Forward Chin-Chuan Cheng Academia Sinica We felt it was time in 2000 to take a concerted action to work on Chinese lexical semantics. The study of word senses in traditional discipline of Xungu appeared to be done piecemeal. The truth-theoretic model of proposition looked illusive at times. Colleagues working on natural language processing demanded a better dictionary for word disambiguation. Few theoretical claims about human manipulation of senses were available. We therefore organized a workshop on lexical semantics to discuss these issues and others at the City University of Hong Kong four years ago. The one-day gathering was fairly informal. But we had fun knowing each of the score of colleagues from Taiwan, Hong Kong, Mainland China, and the United States. We did not call it “First Chinese Lexical Semantics Workshop”. We simply called it a lexical semantics workshop without knowing its consequences. Somehow the gathering in Hong Kong made Professor Yu Shiwen of Beijing University happy. He should be because he had worked on semantics for years. In 2001 he invited more people to the meeting with the title of “Second Workshop on Chinese Lexical Semantics”. The Beijing meeting was enthusiastically followed by the “Third Workshop” in Taipei. Professor Huang Chu-ren was energetic enough to set up the mechanism of abstract submission and evaluation. Some submissions had to be left out because of a large number of excellent papers. The fourth workshop returned to the City University of Hong Kong in 2003. We were not daunted by SARS. Although trip restrictions made us stay home, our papers got exchanged and commented on via the internet. We received hundreds of comments, perhaps more than we would in a face to face conference. It is now 2004. I am pleased to see the workshop fully alive in its fifth year of existence. Our hosts, Drs. Ji Donghong and Lua Kim Teng have kindly accepted our imposition and aptly made arrangements for us to see each other face to face here in Singapore. We are grateful to them for the arrangements. We are also grateful to them for gathering papers in this volume for discussion. I am sure during the workshop we will move away from piecemeal studies of words. We will be a step closer to theoretical generalizations about human cognition of words. 1
Transcript
Forward
Chin-Chuan ChengAcademia Sinica
We felt it was time in 2000 to take a concerted action to work on Chinese lexical semantics. The study of word senses in traditional discipline of Xungu appeared to be done piecemeal. The truth-theoretic model of proposition looked illusive at times. Colleagues working on natural language processing demanded a better dictionary for word disambiguation. Few theoretical claims about human manipulation of senses were available. We therefore organized a workshop on lexical semantics to discuss these issues and others at the City University of Hong Kong four years ago. The one-day gathering was fairly informal. But we had fun knowing each of the score of colleagues from Taiwan, Hong Kong, Mainland China, and the United States. We did not call it “First Chinese Lexical Semantics Workshop”. We simply called it a lexical semantics workshop without knowing its consequences.
Somehow the gathering in Hong Kong made Professor Yu Shiwen of Beijing University happy. He should be because he had worked on semantics for years. In 2001 he invited more people to the meeting with the title of “Second Workshop on Chinese Lexical Semantics”. The Beijing meeting was enthusiastically followed by the “Third Workshop” in Taipei. Professor Huang Chu-ren was energetic enough to set up the mechanism of abstract submission and evaluation. Some submissions had to be left out because of a large number of excellent papers. The fourth workshop returned to the City University of Hong Kong in 2003. We were not daunted by SARS. Although trip restrictions made us stay home, our papers got exchanged and commented on via the internet. We received hundreds of comments, perhaps more than we would in a face to face conference.
It is now 2004. I am pleased to see the workshop fully alive in its fifth year of existence. Our hosts, Drs. Ji Donghong and Lua Kim Teng have kindly accepted our imposition and aptly made arrangements for us to see each other face to face here in Singapore. We are grateful to them for the arrangements. We are also grateful to them for gathering papers in this volume for discussion. I am sure during the workshop we will move away from piecemeal studies of words. We will be a step closer to theoretical generalizations about human cognition of words.
1
Content Table詞語管窺與宏圖.........................................................................................................................1现代汉语中的形式动词.............................................................................................................7The Sinica Sense Management System: Design and Implementation......................................15论汉语释义元语言的特征.......................................................................................................23同源角色框架的表达和计算...................................................................................................39From Frame to Subframe: Collocational Asymmetry in Mandarin Verbs of Conversation......47詞彙素描與多義詞詞義辨識...................................................................................................53隐喻性成语的语义映射...........................................................................................................61漢語情態詞「一定」、「必定」的語義發展與互動...........................................................69略论汉语自然语言的型式 (D-Type Theory ).........................................................................76「別」的語義演變...................................................................................................................84Using WordNet and SUMO to Determine Source Domains of Conceptual Metaphors...........91From Lexical Semantics to Conceptual Metaphors: Mapping Principle Verification with WordNet and SUMO.................................................................................................................99基于现代汉语语法信息词典的词语情感评价研究.............................................................107关于名词多层级分类标注的构想.........................................................................................114基于数据库的汉语语义构词法初探.....................................................................................120论词的语义网络中语义成分的动态性.................................................................................128Feature Representations and Logical Compatibility between Temporal Adverbs and Aspects.................................................................................................................................................133信息处理用现代汉语虚词义类词典研究和工作单设计.....................................................143Multiple-layer Semantic Derivations of Two-part Allegorical Expressions in Taiwanese Southern Min (TSM)...............................................................................................................152基于《知网》的中文信息结构抽取.....................................................................................161词义关系推理.........................................................................................................................169双向考察和验证：并列成分中心语的语义关系和 CCD的名词语义分类体系...............174面向中文信息处理的现代汉语无标记并列结构歧义问题研究.........................................181基于WordNet建立中文同义词词林的上下位关系.............................................................189基于 hownet的无导词义消歧...............................................................................................194蘇軾詩典故用語研究.............................................................................................................201Pan-Chinese Variation on Verbal Synonymy: A Study of Common Reportage Verbs in News Texts........................................................................................................................................213The Usage and Perception of Judgement Termsin the Pan-Chinese Context.........................220漢語詞語與詞素之正負面：三地褒貶詞語初探.................................................................228
汉语词义的微观结构及其切分与描写.................................................................................249汉语的韵律成分界定与句法.................................................................................................254Taxonomy of Fine-grain Semantic Roles for Nominal Modifiers..........................................259閩南語語氣詞 hann、hio、honn的語義比較研究..............................................................264海陸客語副詞"才"與"就"之語義研究..................................................................................268Semantics-related Lexical Access Deficit of Mandarin-Chinese Dyslexia.............................272「少許」、「些許」的複合化過程及語義演變.................................................................276「臉」和「面」在現代華語中的語義區別.........................................................................280义项语汇典例（SVDE ——）的总量控制模型 人机协作对采用汉语注释的语义词汇典例进行计量分析.........................................................................................................................284台灣閩南語中手部動作特指「打」的語義探析.................................................................290大學生對於華語「沒面子」和「丟臉」兩詞彙語義區別的研究.....................................294简直”句式的语义逻辑分析及其生成...................................................................................301从篇章角度看名词性词汇成分的语义关系.........................................................................305《人民日报》语料库专有名词分类的研究.........................................................................311Verbs of Urging in Hakka: A Perspective from Force-Dynamics...........................................316
可以是敬愛老人，也可以是孝順父母，也可以是敬愛愛國詩人。讀者針對一個詞語，廣泛閱讀該詞出現的語言段落，從語境中提煉出‘敬愛’的用法。但是，不是所有的學生都能做總結，有人可能不會偵查或感覺到關鍵所在，因此我們還要加上自動化的輔導來幫助學生提煉出規則來。這樣的輔導需要有詞語的知識，有“詞彙描述”(Kilgarriff and Tugwell 2001) 的功能來幫助學習者，以語境中搭配出現的詞語的統計來告訴讀者‘溺愛’是用於父母對子女的愛護，‘敬愛’的對象可以是同輩、父母、老人等等。初學的人可能不一定能夠在閱讀中了解詞語細微的語義內涵，我們所提出的輔導就是要在最短的時間內幫助學習者得出語言的規則來，加快學習步伐，豐富學習內容。一詞泛讀就是針對一詞來檢查許多出現的語言段落，讓讀者瞭解語意的細緻的意思跟用法，所以我們提出以大量的語言材料來幫助學習者對詞語的學習。
本計畫的題目是“字詞辨識：行為、發展以及神經生理取向的整合性研究”，包括的內容有聲韻表徵與閱讀發展、中文字視覺字性分析相關腦神經機制的時序演進、以鄰項個數效果探討中文字詞辨識歷程--事件誘發電位之系列研究、利用穿顱磁刺激探討中文字詞辨識的腦神經機制、以行為及腦部事件相關電位實驗探討對中文雙字詞的錯誤連結記憶。所用的先進技術包括電位腦電波、腦磁波、功能性磁振造影、穿顱腦磁刺激術等考察詞語辨識的時候腦部的活動，從這些腦功能顯影技術來瞭解語言行為中的大腦運作(Kuo at al . 2003)。六、願景
引用書目黃居仁 . 主編 . 2004.中文的意義與詞義 . 台北： 中央研究院語言學研究所。鄭錦全 . 1998. 從計量理解語言認知 . In Benjamin K. Tsou, Tom B. Y.
Lai, Samuel W. K. Chan, and William S-Y. Wang eds. 漢語計量與計算研究 (Quantitative and Computational Studies on the Chinese Language) 15-30. City University of Hong Kong.鄭錦全 魏培泉 陳昭容 陳克 健 曾 淑娟 . 2002. 語言典藏計畫 . 數位典藏國家型科技計畫中央研究院機構計畫. 鄭錦全 黃居仁 羅鳳珠 蔡美智 . 2004. 兼具教學與研究功能的全球華語文數位教與學資源中心. 國科會數位學習國家型科技計畫整合型計畫 鄭錦全 洪蘭 曾志朗 李佳穎 蔡介立 郭文瑞 阮啟弘 李俊仁. 2004. 字詞辨識：行為、發展以及神經生理取向的整合性研究. 中央研究院主題計畫.
Huang, Chu-Ren. 2003. SINICA BOW: Integrating il ingual WordNet and SUMO ontology. Invited panel talk: Synergy Between Language Resources and Knowledge Resources. The 2003 IEEE International Conference on Natural Language Processing and Knowledge Engineering (NLPKE2003), Special Session on Upper Ontology and Natural Language Processing. Beijing. Oct. 28.
Kilgarriff , Adam and David Tugwell . 2001 WORD SKETCH: Extraction and display of s ignif icant collocations for lexicography. In Proceedings of . Workshop on COLLOCATION: Computational Extraction, Analysis and Exploitat ion 32-38. 39th ACL & 10th EACL, Toulouse, July 2001.
Kuo, W.J. , T. C. Yeh, C. Y. Lee, Y. T. Wu, C. C. Chou, L. T. Ho, D. L. Hung, O. J . Tzeng, and J . C. Hsieh. 2003. Frequency effects of Chinese character processing in the brain: an event-related fMRI study. Neuroimage 18:720-730.
In contemporary Chinese, there is a subclass of verbs called Dummy Verbs. After briefly introducing the lexical meanings of two typical dummy verbs ‘Jiayi’ and ‘Jinxing’, this paper discusses the grammatical attributes of ‘Jiayi’ and ‘Jinxing’ in detail, and further explores their functions as the marks of syntactic constituents and semantic roles.
Keywords : Dummy Verb, Lexical Meaning, Grammatical Attribute, Semantic Role
[9]认为，“这样的句子，翻译成英语时，‘进行’就没有着落”。 现在给出 5种译文：1. We are carrying on reforms on the state-owned enterprises in our country.2. We are making reforms on the state-owned enterprises in our country.3. The reforms on the state-owned enterprises are being carried out in our
country.4. We are reforming the state-owned enterprises in our country. 5. The state-owned enterprises are being reformed in our country.应该说，这 5 种译文基本上都传达了原文的意思。从前面的讨论，可以了解到“我国正在对国营企
The Sinica Sense Management System: Design and Implementation
Chu-Ren Huang, Chun-ling Chen, Cui-Xia Weng, and Keh-jiann ChenAcademia Sinica
1. Background and MotivationIt has been a trend for language engineering to construct a sense-based lexical
knowledgebase as a core foundation. WordNet and Euro WordNet are two well-known examples. There are two important criteria in constructing this knowledgebase: linguistic felicity and data cohesion. Huang et al. (2003) discussed how to achieve linguistic felicity in building a comprehensive inventory of Chinese senses from corpus data. It introduced five criteria as well as operational guidelines for sense distinction. In this paper, we will discuss how to achieve data cohesion for the sense information thus collected through a Sinica Sense Management System (SSMS).
2. Introduction to the Content of the SSMSThe SSMS manages both lexical entries and word senses. This system is designed and
implemented by the Chinese WordNet Team at Academia Sinica. It contains all the basic information that can be merged with the eventual Chinese WordNet. The basic structure of this system is meaning-driven: Each sense of a lemma is identified specifically and given a separate entry. When further differentiation at the meaning facet level is called for, each facet of a sense is also described in a full entry (Ahrens et al., 1998). In addition to sense and meaning facet, this system also includes the following information: POS, example sentences, corresponding English synset(s) from Princeton WordNet, and lexical semantic relation such as synonym/antonym, and hypernym/hyponym. Moreover, the overarching structure of the system is managed by a sense serial number, and inter-entry structure is established by cross-references among synsets and homographs.
In the present stage, the Chinese WordNet Team focuses on analyzing middle-frequent words in Sinica Corpus. The reason to choose middle-frequent words as our target ones is that with only three to five senses of a word, we can investigate senses and meaning facets of each word deeply and accurately, which would avoid the simple situation of one sense in low-frequent words, and the complicate situation in high-frequent words with numerous senses. Up to now, 1000 more lemma have been analyzed, and more than 2000 senses have been distinguished. We also published five technical reports to present these results [4]. In the near future, these fruits will be used as a basis for Natural Language Processing or E-learning application.
3. The Design Principle of SSMSA sense-based lexical knowledgebase with data cohesion must meet three requirements:
unique identification of senses, trackability of sense, and consistent sense definitions. SSMS has four devices to supply these requirements.
3.1 The Unique Serial NumberFirst, each sense or meaning facet is identified by a unique serial number in SSMS. In
Princeton WordNet (Fellbaum 1998), each synset is given a unique offset number. However, the offset number does not have any logical structure to it. Hence, although it guarantees unique identification, it is not very trackable. An alternative is to set up a base ontology and assign senses to an ontological node with a unique ID. However, this is not feasible since we
16
cannot pre-designate all the possible conceptual and semantic relations. And if decision is made to encode only certain higher level nodes, the random assignment issue is unavoidable since more than one lexical sense will be assigned to the same node. In our system, the unique serial number of each sense is composed of three segments: the sequential information of when the lemma was processed, the lemma form, and the sense classification code for each lemma (including the meaning facet level). Take “bao4 zhi3 (newspaper)” for example. “bao4 zhi3” has two senses and two meaning facets being distinguished. The lexical entry of “bao4 zhi3” is as follows.
Example 3-1: The result of sense distinction for “bao4 zhi3 (newspaper)”
Four-level unique serial number is shown as below to express four segments of the unique serial number for one meaning of “bao4 zhi3”.
報紙 “ bao4 zhi3 (newspaper) ” Lemma processing year 03-
Lemma form ID -0018-
The first sense -01-
The first meaning facet -01
The unique serial number for 1st. meaning facet of 1st.sense of “bao4 zhi3” => 0300180101There are four advantages to manage the sense database with unique serial numbers.
First, the sequential number not only gives a unique code to each lemma, it also enables a project manager to track work progress more easily. Second, including the lemma in the serial number helps human users to quickly identify the relevant senses. It also facilitate man-machine interface such as in keyword search for senses. Third, it also provides a logical structure of the sense serial number since each lemma represents a small number of possible senses. Lastly, four digits are reserved to identify senses and meaning facets belong to each lemma. The first two digits are reserved for senses and the last for meaning facets. These four digits also allow the minimal space to identify exact sense in the database. For instance, when stipulating a synonym, we can identify it as word0200, which refers to the second sense of a certain lemma. There is no need to repeat the complete sense serial number. The sense serial number enables unique identification and also contributes to trackability.
3.2 The Cross-reference deviceSecond, SSMS will automatically prompt all possible cross-references. When a lemma is
called up for analysis, all existing records that contain this lemma will be prompted. This includes not only lexical semantic relations such as synonyms and hyponyms, it also includes
17
and sense definition that contain this lemma, as well as any explanatory notes that contain this lemma. This feature allows sense relations to be clearly defined, and inconsistencies to be detected. In addition any anomaly in definition or expression format will also be discovered. This process will also help us to narrow down to a set of control vocabulary for sense definition. This feature contributes to both the trackability of senses and consistency of sense definition.
3.3 The concurrent lexical knowledgebase and coupusThird, SSMS enables parallel concurrent of the lexical knowledgebase and corpus. When
a lemma is chosen in the system, all tagged example of that lemma from Sinica Corpus are retrieved. This allows closer examination of how the senses are used and distributed. It also allows automatic selection of corpus example sentences. In turn, when the sense classification is completed, SSMS allows all the corpus sentences to be sense-tagged and returned to merge with the original corpus. In other words, a sense-tagged corpus is being processed in parallel. This feature allows each lexical sense to be trackable to its actually uses in the corpus. It also allows linguist to examine the data supporting each sense classification.
3.4 Linking to the Sinica BOWFourth, SSMS is also linked to the bilingual wordnet information at Sinica BOW.
Candidate English synset correspondences, including offset number, are shown after a Chinese lemma is chosen. This allows the cross-lingual trackability and consistency.
4. The Implementation of SSMS There are three major phases in this system implementing. In lemma analysis phase,
based on the criteria and operational guidelines proposed in Huang (2003), we distinguish senses and meaning facets for each word. At the same time, Sinica Corpus and WordNet will be referred for POS, examples and English translation. Then through the help of dictionary resources or word mapping by the system, we decide the word relation. The second phase can be divided into two steps. First, we design the schema of the sense management system database for storing the analyzing result of the first phase. Then, as for the data access, we develop the interface to help the Chinese Wordnet Team insert and query from the database. We employ DELPHI tool to design our system interface. Thought the interface, the data in the database also can be exported as Word documents. Last, the third phase of this system implementation is the application phase. Our work project is to build Chinese WordNet web sites for users querying. The development language of these web pages is HTML and ASP. Finally, these web pages in the web sites could be viewed thought web server. By the way of the Internet, people can retrieve data from our sense management database system everywhere at anytime. The flow of the Sinica Sense Management System is displayed in the following chart.
Figure 1: The flow chart of the Sinica Sense Management System.
18
We can represent the overall framework of SSMS diagrammatically in Fig. 2. As the diagram indicates, the Chinese WordNet Team use SSMS to access database and have electric documents as Word report. Moreover, the users in the internet can browse HTML/ASP pages to query database through and web server.
Figure 2: The overall structure of SSMS.
4.1 The Schema of SSMS Database in Class DiagramIn the section, we discuss and design the schema of SSMS Database. The Unified Modeling
Language (UML) [2][7] is a graphical notation that provides the conceptual foundation for assembling a system out of components from the 4+1 views and nine diagrams. Each view is a projection into the organization and structure of the system, focused on a particular aspect of that system.
We employ the class diagram notations in UML to provide a static view of application concepts in terms of classes and their relationships including generalization and association. Therefore, we only introduce the details about class diagrams as follows.
Class diagrams [2][7][6] commonly contain the following features:
1. A class diagram shows a set of classes and their relationships. For example, the class diagram of the Suppliers-and-Parts database as shown in Fig. 3. The terms with italic style in Fig. 2 indicates the concepts about class diagrams.
Figure 3: A class diagram for the Suppliers-and- Parts Database.
2. A class is a description of a set of objects that share the same attributes, operations, relationships, and semantics. A class mainly contains three important parts: its name, attributes, and operations. We explain these terms as follows:
(a) Class name: every class must have a name to distinguish it from other classes. For example, Suppliers or Parts are class names.
(b)Attribute: an attribute represents some property that is shared by all objects of that
19
class. A class may have any number of attributes or no attributes at all. For example, in Fig. 3, the Suppliers have some attributes such as sno, sname, city.
(c) Operation: an operation is the implementation of a service that can be requested from any object of the class to affect behavior. A class may have any number of operations or no operations at all. For example, in Fig. 3, the class of Suppliers has an operation add().
3. There are three kinds of relationships between classes:
(a) Association: an association is a structural relationship that specifies objects of one thing to be connected to objects of another. For example, in Fig. 3, a line drawn between the involved classes (Suppliers and Parts) represents an association named supply.
(b) Aggregation: an aggregation is a ‘whole/part’ relationship, in which one class represents a larger thing (the ‘whole’ class), which consists of smaller things (the ‘parts’ class). Moreover, an aggregation represents a “has-a” relationship, which means that an object of the ‘whole’ class has objects of the ‘part’ class. To represent an aggregation, an empty diamond will be drawn at the ‘whole’ class end of the line linking two classes.
(c) Inheritance: An inheritance relationship can be regarded as a generalization (or specialization), which is a taxonomic relationship between a general (super classes) and a special (subclasses) element, where the special element adds properties to the general one and behaves in a way that is compatible with it. Therefore, it is sometimes called an “is-a-kind-of” relationship. An inheritance relation is represented by means of a large empty arrow pointing from the subclass to the super class. For example, in Fig. 3, Domestic and Foreign suppliers (two subclasses) are a kind of suppliers (the super class).
According to the need of SSMS content and design principle, Fig. 4 is the schema of SSMS database using the concepts of class diagram.
Figure 4: The schema of the Sinica Sense database.
4.2 The Function of SSMS In this section, we will discuss the interface marking for SSMS. The development
language of SSMS interface is DELPHI 7.0. Based on the need of program execution, the function of SSMS is shown in Fig. 5. In SSMS, the programs have many functions and these functions can be represented in windows interface and ASP web pages. Sense management and Sense visualization are two major functions in SSMS. In Sense management function, the Chinese WordNet term can insert, update, and delete data including lexical entries, word
20
sense, meaning facet, POS, example sentences, English synset(s), lexical semantic relation. The Sense visualization is SSMS interface and can be divided into two parts: Sense Query and Word Report. The format of SSMS interface is shown in Fig. 6. The SSMS interface provides a user-friendly interface to operate and maintain. For the Sense query function, the users can enter a serial number or a lexical entry for sense querying in SSMS interface. Another function, the Word report, uses development software Crystal Report9 to produce electric documents shown as Fig. 7.
Figure 5: The class diagram of SSMS function description.
Figure 6: The interface of SSMS.
21
Figure 7: The format of Word report.
6. ConclusionIn sum, SSMS is not only a versatile development tool and management system for
sense-based lexical knowledgebase. It can also serve as the database backend for both Chinese WordNet and any sense-based applications for Chinese language processing.
References[1] Ahrens, K., L. Chang, K. Chen, and C. Huang, 1998, Meaning Representation and
Meaning Instantiation for Chinese Nominals. Computational Linguistics and Chinese Lnaguage Processing, 3, 45-60.
[2] Booch, G., J. Rumbaugh, and I. Jacobson, The Unified Modeling Language User Guide, Addison-Wesley, 1999.
[3] Fellbaum, Christine. Ed. 1998. WordNet: An Electronic Lexical Database. Cambridge, MA: MIT Press.
[4] Huang, Chu-Ren (ed.), 2004, Sense and Sensibility series: Technical Report 03-01~04. CKIP, Taipei.
[5] Huang, Chu-Ren et al., 2003, Sense and Meaning Facet: Criteria and Operational Guidelines for Chinese Sense Distinction]. Presented at the Fourth Chinese Lexical Semantics Workshops. June 23-25 Hong Kong, Hong Kong City University.
[6] Muller, R.J., Database Design for Smarties: Using UML for Data Modeling, Morgan Kaufmann, 1999.
[7] Oestereich, B., Developing Software with UML Object-Oriented Analysis and Design in Practice, Addison-Wesley, 1999.
Representation and Computing of Cognate RoleFrameZhendong Dong Qiang Dong Changling Hao
Research Center of Computer & Language Information Engineering, CAS, Beijing, 100083E-mail: [email protected]
Abstract: Borrowing from the term of “cognate object”, we use “Cognate RoleFrame” to reveal a kind of semantic relations between nouns and verbs like “hatred” and “hate”. By Cognate RoleFrame, we mean that the noun has the same role frame as its correspondent verb. In HowNet we use “CoEvent” as the identifier to describe all the nouns of Cognate RoleFrame. We demonstrate two HowNet-based tools to evaluate our treatment of Cognate RoleFrame.Keyword: HowNet, valency grammar, event role, event role frame, cognate RoleFrame
AbstractThis paper examines the collocational patterns of Mandarin verbs of conversation and proposes
that a finer classification scheme than the flat structure of ‘frames’ is needed to capture the semantic granularity of verb types. The notion of a ‘subframe’ is introduced and utilized to explain the syntactic-semantic interdependencies among different groups of verbs in the conversation frame. The paper aims to provide detailed linguistic motivations for distinguishing subframes within a frame as a semantic anchor for further defining near-synonym sets.
1. IntroductionAs the importance of lexical semantic research grows with the need of representing human
knowledge, various lexically-based information networks have been proposed, such as the comprehensive work of differentiating word senses and sense relations in WordNet (Miller et al. 1990), the ontological hierarchy in SUMO (Das et al 2002, Pease et al 2002, Niles and Pease 2003), and the more linguistically-motivated model of FrameNet (Baker, Fillmore and Cronin 2003). While all providing valuable information regarding certain aspects of word meaning, the first two are constructed in a more intuitive manner. FrameNet, on the other hand, is based on the theory of Frame Semantics (Fillmore and Atkins 1992) and attempts to define meaning within a set of shared knowledge or background information, that is, a frame. However, as pointed out in Liu and Wu (2003), if meaning is anchored in the notion of ‘frame,’ then we need independent motivations for postulating different ‘frames’. What seems to be lacking in the current framework is a cognitive linguistic explanation as to how the individual ‘frames’ are distinguished and interrelated? In other words, what are the ‘semantic relations’ among all the frames? To answer the question, Liu and Wu (2003) proposed an overarching conceptual schema which incorporates all the core frame elements (FEs) and accounts for the interrelationship among various frames in the communication domain. By providing a cognitive schema as a macro-structure, the distinction of frames is then well-motivated. However, there still remains another issue at a micro-level, as indicated by Liu and Wu (2003):
Within each frame, a wide range of verbs are found and one would wonder how these verbs differ from each other. For example, English verbs speak, discuss, quarrel, and gossip, are all found in the Conversation Frame, but obviously, these lemmas encode something different. What are the differences ? There seem to be frame-internal features that also need to be characterized.
In this paper, we will show that within each frame, a more elaborated classification system is needed to account for a variety of verb behaviors. The notion of ‘subframe’ is introduced and utilized to capture the syntactic-semantic interdependencies observed in the corpus data15.
2. Motivation for the Conversation FrameCompared with the other communication frames, the Conversation Frame is unique in that it
profiles the property of reciprocality or two-way communication. Verbs in the Conversation frame encodes reciprocal events where participants are involved as Interlocutors, such as tan談 ‘talk’ ,
15 A preliminary model of the Mandarin VerbNet (http://140.114.75.18/verbnet/webform1.aspx/) has been constructed by researchers from National Chiao Tung University and National Tsing Hua University with the support of a NSC grant.
tanlun談論 ‘converse’, ta-lun討論 ‘discuss’, shangliang商量 ‘discuss’, xietiao協調 ‘negotiate’, xieyi協議 ‘negotiate’, goutong溝通 ‘communicate’, chaojia吵架 ‘quarrel’, zhenglun爭論 ‘argue’, xianliao閒聊‘chat’, and liaotian聊天 ‘chat’, etc16. These verbal events highlight part of the conceptual schema as discussed in detail in Liu and Wu (2003) and represented in (1) below. The core Frame Elements (with bold fonts and grayed areas) help define the frame as a bidirectional communicative activity conducted by both the Speaker and Addressee as Interlocutor 1 and Interlocutor 2 (or Interlocutors), via a certain Medium, on a given Topic.
(1) Conceptual schema for the Conversation frame: Encoding Decoding
- ---- Noise----- -----Noise-----
The Conversation frame proves to be well-motivated in relation to other communication frames, as most of its verbs share the same conceptual schema and realizing similar constructions in coding the core frame elements. There is, however, a fundamental question to be answered, that is, within the Conversation frame, are there semantic subtypes that are also syntactically motivated?
3. Motivation for Distinguishing SubframesAs mentioned above, verbs of conversation involve a set of core Frame Elements:
Interlocutor1, Interlocutor2 (or combined to Interlocutors), Topic and Medium. In most cases, the default Medium is ‘face-to-face’ when not overtly mentioned, as in the following sentence: 他們在 談 / 討論 / 溝通 / 吵 / 聊 人生的意義 ‘They are talking/discussing/exchanging views about/argue about/chat about the meaning of life.’ Intuitively, these different lemmas seem to encode differences in manner, formality or purpose, while sharing the same topic. But what are the grammatical correlates to the lexicalized meaning differences? When looking closely at their collocational patterns, we found that there are asymmetrical distributions in five respects: 1) V+V pattern: some may occur with a preceding verb such as jinxing 進行 ‘proceed’ or dacheng 達成 ‘achieve’; 2) V+NP pattern: the core element ‘Interlocutor2’ may sometimes be coded as the direct object; 3) Metonymic subject: the subject of the event may be inanimate entities taking the role of Interlocutors by the principle of metonym; 4) V+ Complement: some verbs take a postverbal complement or adverbial adjunct denoting effect evaluation, such as chenggong 成功 ‘succeed’ or shibai 失敗 ‘fail’；and 5) In terms of distribution of grammatical functions, they show different frequencies of nominalization. Based on the five criteria, verbs of conversation can be further divided into 5 groups with corresponding sets of near-synonyms. We will address the syntactic-semantic interdependencies revealed by each pattern in the following sections.
16 The lemmas discussed in this paper are high-frequent words of the conversation frame used mainly in Taiwan. Besides the words mentioned, there are some other lemmas , as an anonymous reviewer mentions, such as tung hua通話 ‘interconnect’, shiang tan詳談 ‘descant’ , shen liau 神聊 ‘tittle-tattle ’, sz yu私語, etc. These words can be included as members of one the subframes discussed in the paper.
50
Intl1 Intl 2
UnderstandingSign/SignalMessage
Topic
Medium
Motivation (Internal Cause)
3.1. V+V Pattern: with the preceding verbs jinxing (進行) or dacheng (達成)The use of the light verb ‘進行’ entails a formal register and encodes a process or atelic event. It
tends to occur with an activity verb compatible with the formal register and involving a durative process, as shown in (2): (2) a. 進行 討論/溝通/協調
b. *進行 談/談論/吵架/爭論/聊天/閒聊Below is the distributional tendency found in Sinica Corpus:
(3) Percentage with Jingxing (進行)V1 V2 討論 溝通 Other Verbs (談/談論/吵架/爭論/聊天/閒聊)
進行 4% (3/83) 6% (25/419) 0%
Another verb dacheng (達成) ‘achieve’ is also compatible with some conversation verbs, requiring a formal register but encoding a telic event. The verb dacheng is only found with the nominalized form of such verbs as溝通/協調/協議17, i.e., activity verbs entailing a semantic endpoint with an incremental theme, as shown in (4):(4) a. 達成 溝通/協調/協議
b. *達成 談/ /討論/吵架/爭論/聊天/閒聊
The co-occurrence with the preceding verbs jinxing (進行) or dacheng (達成) serves to distinguish the conversation events in terms of its pragmatic mode (formal vs. informal) and event types (telic vs. atelic).
3.2 V+NP pattern: Interlocutor2 as the Direct ObjectAnother pattern that sets the verbs apart regards the semantic role of the object NP. While most
verbs can only take the Topic as the direct object, some verbs may encode Interlocutor2 as the direct object without adding the associative marker 和/與/跟, as shown in (5):(5) a. [執行秘書]Intl1已協調 [相關單位]Intl2
b. 由[他] Intl1負責溝通 [校方] Intl2
This suggests that with the verbs協調/溝通, the co-participant, Interlocutor2, may be viewed as the undergoer or the affected target of the event. Among the sentences of協調/溝通 followed by an object NP, an average of 23 percent may take an Interlocutor 2 as the direct object in Sinica Corpus:(6) Percentage with Interlocutor2 as DODO V 溝通/協調 Other Verbs (談/談論/吵架/爭論/聊天/閒聊/討論)
Interlocutor2 as DO 23%(28/123) 0%
3.3 Use of Inanimate Subject Interlocutors in the conversation events are by default human participants. However, some
verbs may take inanimate subjects (place or institute names) as Interlocutors via metonymic extensions from institute/building to human organization:(7) a. 台北和北京 談/討論/溝通/協議 了很久
b. 台北和北京 *聊天/*閒聊 了很久The application of metonymy tends to be associated with verbs that comply with the formal
17 In Sinica Corpus, only ‘達成協議’ is found, but in the Kimo website, we found examples: 1)具備這三項條件就可以達成溝通; 2) 兩邊的陣營終於達成協調.
51
register requiring also a formal, non-personal topic (e.g., public affairs). Sinica corpus shows that verbs with marked manners tend not to be used with metonymic subjects:
(8) Percentage with Inanimate Subject: VSubj. type
談 討論 溝通 Other Verbs(吵架/爭論/聊天/閒聊)
Inanimate Subject 2%(4/191)
6%(5/83)
25%(21/85)
0%
3.4 Postverbal Complement with Effect EvaluationAmong the conversation verbs, only the ‘negotiate’ verbs (e.g., 協調, 溝通) may collocate
with effect evaluation complement such as 成功 ‘successfully’ and失敗‘failingly’, as shown below with examples and percentage rate from the Sinica Corpus:(9) a. 國防部和兩廳院已初步協調[成功]
b. 在協調[失敗]後，水公司終於昨天宣布放棄其中一口(10) Percentage of result evaluation complement: Comp. V 協調 Other Verbs (談/談論/吵架/爭論/聊天/閒聊/討論)
成功/失敗 12.5% (2/16) 0%
The co-occurrence with effect evaluation complement indicates that the semantics of the ‘negotiate’ events encode some kind of an ‘effect’ or ‘result’ that is being sought by the negotiation process. This also implies that the two-way communication in the event of協調/溝通 is a solution-seeking process which is semantically bounded and may be evaluated as to whether the solution or purpose has been achieved.
This pattern also correlates with the use of dacheng 達成 ‘achieve’, as mentioned above, in signaling an evaluation of the attainment of the desired result.
3.5 Frequency of NominalizationSome groups of verbs tend to be nominalized more frequently than the others. Comparing the
high-frequency verbs and their distributions over grammatical functions, we see clear skewing in nominal uses:(11) Distribution of Predicate vs. Nominal UsesFunc. V 談 討論 溝通 吵架 聊天Predicate 97%
(680/701)52%
(83/161)55%
(415/1013)76%
(123/162)94%
(134/142)
Nominalized 3%(21/701)
48%(78/161)
45%(598/1013)
24%(39/162)
6%(8/142)
Nominalization serves to change verbs to event nominals that may be referred to as a quantifiable entity. Nominalization is also highly correlated with the formal register of written texts.
4. Subframes as an Anchor for Near-synonymsThe asymmetrical distributions of the conversation verbs over different collocational
associations clearly suggest that verbs can be further divided into subtypes. Although sharing the same conceptual frame, subclasses of verbs show distinct patterns of syntactic-semantic interdependencies that may serve as the basis to further define near-synonym sets. These subtypes may be viewed as anchored in different subframes. Below is a summary of the collocational patterns associated with the 5 subframes within the conversation frame:
52
(12) Collocational Patterns associated with the Conversation Subframes CPSubtypes 進行+V 達成+V Intl2 as DO Inanimate
SubjectComplement成功/失敗 [+Nom]
1. Converse: 談/談論 No No No Yes No Low
2. Discuss: 討論 Yes No No Yes No High
3. Negotiate協調/溝通 Yes Yes Yes Yes Yes High
4. Quarrel: 吵架/爭論 No No No Yes No Mid-High
5. Chat: 聊天/閒聊 No No No No No Low
Based on the distributional variations of collocational patterns, we can group all the other conversation lemmas into the five subframes: 1) ‘Converse’ subframe: 談, 談論, 交談, 講, 談話, 講話, 說話, 會話,交流,會談,對話,對談2) ‘Discuss’ subframe: 討論, 商量, 商討, 商談, 會商3) ‘Negotiate’ subframe: 協調, 協議, 溝通, 協商, 協議, 交涉, 接洽, 洽商4) ‘Quarrel’ subframe: 吵架, 爭論, 爭吵, 爭執, 爭辯, 辯論5) ‘Chat’ subframe: 閒聊, 聊天, 閒談, 閒聊, 談天.
5. Concluding Remarks and Theoretical ImplicationsWith the proposal of subframes within the theoretical construct of Frame Semantics, verb
meanings may be defined with finer distinctions that are syntactically motivated. However, further fine-grained semantic distinctions are still needed to differentiate near-synonyms within each subframe, such as交談 vs.交流 or 討論 vs.商量, etc. It is exactly at the subframe level that we may anchor all the near-synonym sets as closely related. In sum, to fully represent the meaning relations among verbs, we’d like to propose the following classificational scheme for representing verb meanings:
Domain ->Frame -> Subframe -> Near-synonym Set -> Lemma
The five-layered structures allow verbs to be represented in a frame-based semantic hierarchy with detailed lexical information to further disambiguate near-synonyms.
ReferencesBaker, Collin F., Charles J. Fillmore and Beau Cronin (2003) The Structure of the Framenet Database.
International Journal of Lexicography 16(3).281-296. Das, Subrata, Kurt Shuster, and Curt Wu. 2002. Ontologies for Agent-Based Information Retrieval and Sequence
Mining. Proceedings of the Workshop on Ontologies in Agent Systems (OAS02), held at the 1st International Joint Conference on Autonomous Agents and Multi-Agent Systems,.Italy, July 15-19.
Fillmore, Charles J., and Atkins, Beryl T. 1992. Toward a Frame-Based Lexicon: The Semantics of RISK and Its Neighbors. Frames, Fields, and Contrasts, ed. by Adrienne Lehrer and Eva Feder Kittay. 75-102. Hillsdale. New Jersy: Lawrence.
Kennedy, C. and Levin, B. 2002. “Telicity Corresponds to Degree of Change” Handout to Speech at Georgetown University.
Liu, Mei-Chun and Yiching Wu. 2003. Beyond Frame Semantics: Insight from Mandarin Verbs of Communication. Paper presented at the 4th Chinese Lexical Semantics Workshop. City University of Hong Kong, Hong Kong. June 22-July 11. (http://icl.cityu.edu.hk/conference/4CLSW/BIG5/home.htm）
Miller, A., R. Beckwidth, C. Fellbaum, D. Gross, K.J. Miller. 1990. Introduction to WordNet: An on-line Lexical Database. International Journal of Lexicography. 3.235-244.
Niles, Ian and Adam Pease. Linking Lexicons and Ontologies: Mapping WordNet to the Suggested Upper Merged Ontology. Proceedings of the 2003 International Conference on Information and Knowledge Engineering (IKE’03), Las Vegas, Nevada, June 23-26, 2003.
Pease, A., Niles, I., and Li, J. 2002. The Suggested Upper Merged Ontology: A Large Ontology for the Semantic Web and its Applications. Working Notes of the AAAI-2002 Workshop on Ontologies and the Semantic Web, Edmonton, Canada, July 28-August 1, 2002.
Website Addresses:FrameNet: http://www.icsi.berkeley.edu/~framenet/HowNet (知網): http://www.keenage.com/html/c_index.html/Sinica Corpus (平衡語料庫): http://www.sinica.edu.tw/SinicaCorpus/SUMO Ontology: http://ontology.teknowledge.com/WordNet: http://www.cogsci.princeton.edu/~wn/The Academia Sinica Bilingual Ontology WordNet (BOW): http://bow.sinica.edu.tw/
的意義，如華語動詞「送」，可以「送」信，「送」禮，「送」客，也可以「送」命。根據 Li & Huang (1999)的統計，《同義詞詞林》收錄的 52716目詞當中多義詞約 7800個，佔 14.8%。對於使用者而言，多義詞理解上沒什麼困難，例如華語人士可以毫不費力地分辨下列五句「送」的詞義：(1a)遞送，(1b)贈送，(1c)陪送，(1d)斷送，(1e)推送。
(1)a. 把手伸出來，他還會抓住你的手往嘴裡送。 b. 花也可以送。c. 最後她哭了，也不許別人送，一個人回家去了。 d. 你把孟瑞的命送了。
e. 貝珍把東尼的小包往瑟勒思娣面前一送。反觀自然語言處理，多義詞的詞義辨識，也就是所謂的詞義消歧(Word Sense Disambiguation)，卻是發展資訊擷取自動化極重要的環節，電腦遇到多義詞，必須有辦法根據上下文自動排除歧義，才能正確分析和解讀語料。早期詞義消歧仰賴專人撰寫規則，耗費時日，能處理的詞項極為有限。八０年代以降，詞典成為詞義消歧主要的依據，電腦根據詞典中的詞義解釋自動取得辨識詞義的訊息。
詞典解歧法落實了詞義自動抽取和判讀的理念，但如 Li & Huang (1999)所述，一旦上下文異於詞典的定義，消歧效果便大打折扣，Lam et al. (1997)利用《現代漢語詞典》的詞義解釋及《同義詞詞林》的義類代碼處理多義實詞，詞義辨識率僅 45.5%。鑑於詞典無法預測可能出現的語境，也無法提供所有的組合特徵，利用電腦自動取得大量組合知識的語料庫解歧法於焉產生，如李涓子(1999)在詞典的基礎上結合語料庫，多義詞辨識率提高為 52.13%；Chen et al. (1999)同樣藉由語料庫解決自動翻譯中的歧義和多義問題，效能也見提昇。儘管一詞多義相當普遍，但在實際語句中起作用的大多只有一個。認知語言學家
Choueka et al. (1983)實驗證明，人類僅憑上下文中一個或少數幾個成分就能分辨多義詞的詞義。因此除了詞典和大型語料庫之外，晚近的消歧策略進一步將搭配成份的語法訊息納入考量。如此一來，詞性相同的多義詞可以經由比較搭配成份的語法功能和語意類型，達成詞義辨識，如王惠(2002)辨別華語名詞「辦公室」的兩種意義，Chuang (2003)區隔英語動詞 set和華語動詞「擺」的多重意義。Ahrens et al. (2003)並利用MARVS模組屬性動詞語意表達模式，提出多義詞詞義(sense)與義面(meaning facet)的辨識標準。
(24) a. 吳德貴率先五指一抓一捏就往口裡送。b.每天派人給和尚送飯。c. 成品將一波波從趕工的生產線上送到市場。
(25) a. 有一農婦採集殞石後送莫斯科研究。 b. 當事人送由警備總部依法嚴辦。(26) 左手向前一送，藏在衣袖中的匕首已刺了出去。(27) a. 爸爸親自送我們到車站。 b.* 我們被爸爸親自送到
車站。c. 雙雙昏倒在地，被救護車送到醫院。
(28) a. 又要養家活口，又要送子女到國外唸書。b. 偷偷送棍子到學校給老師的，也是家長。
60
(29) a. 由男方送兩頭活豬代表雙雙對對給女方。b. 西裝多是由朋友送的。 c. 生日那天，收到三位女同事送的花。
(30) 這兩名受傷病患是由陸軍總部送來的。主要參考書目中國社科院語言所詞典編輯室編 2001，《現代漢語詞典》繁體字版，香港：商務印書錧。王惠 2002，基於組合特徵的漢語名詞詞義消歧，CLCLP 7.2, 77-88.呂叔湘 1980，《現代漢語八百詞》，北京：商務印書館。李涓子 1999，漢語詞義排歧方法研究，博士學位論文，清華大學圖書館。袁暉主編 2001，《現代漢語多義詞詞典》，太原：書海出版社，二版。黃居仁 2003，〈語言的網路，知識的架構：由意義誘發的語言科技〉，台灣師範大學主辦第三屆語言與科技研討會邀請講席，2003年 12月 20日。梅家駒、竺一鳴、高蘊琦等 1983，《同義詞詞林》，上海：上海辭海出版社。Ahrens, K., C.-R. Huang, and Y-H Chuang. 2003. Sense and meaning facets in verbal semantics: A
MARVS perspective. Language and Linguistics 4.3: 469-484.Chen, H.-H., G.-W. Bian, and W.-C. Lin. 1999. Resolving translation ambiguity and target polysemy in
cross-language information retrieval. CLCLP 4.2, 21-38.Choueka, Y., and S. Lusignan. 1983. A connectionist scheme for modeling word sense disambiguation.
Cognition and Brain Theory 6.1, 89-120.Chuang, Y.-H. 2003. Sense Distinction of Verbs in English and Mandarin Chinese: An Analysis of the
Verbs “Set” and “Bai3”. MA thesis. National Taiwan University.Lam, S.-S., K.-F. Wong, and V. Lum. 1997. LSD-C –A linguistics-based word sense disambiguation
algorithm for Chinese. Computer Processing of Oriental Languages 10.4, 409-422.Levin, B., and M. Hovεv. 2001. What alternates in the dative alternation? CSSP. Li, J., and C. Huang. 1999. A model for word sense disambiguation. CLCLP 4.2, 1-20.
源21” (Vehicle)之间的互动，隐喻本身指包含两个部分的整体（胡壮麟，2004）。因此，隐喻涉及到两个领域，用其中一个领域 A 来说明另一领域 B 。两个领域中，被说明的领域（即 B）称为目标领域(target domain)，另一个（即A）称为源领域(source domain)。两个领域有不同的语义网络，语义结构特征具有相异性。但是，本体和喻源之间之所以能够互动，是因为它们之间有“共同点”(ground)或者“相似性”。相似，使隐喻成为可能；相异，使隐喻更加鲜明，使认知更深入。3.2 隐喻的语义映射
隐喻意义是两个语义领域之间的语义映射。隐喻是以喻源和本体之间的相似性作为意义转移的基础的。以“A 是 B ”结构出现的映射是将有关喻源域 B 的适当的知识结构的一部分映射到目标域结构 A 上。为了以喻源域的词语来理解目标域，人们必须具有对喻源域的适当的知识。
语言中的隐喻是一种以词语或句子为焦点，以语境为框架的语用现象。孤立的词严格意义上讲不能称为隐喻。只有在具体的语境中才能判断一个词是否用作隐喻（束定芳，1998）。但是，成语则不然了。成语是人们长期以来习用的定型词组或短语，脱离了字面的意思，获得了统一的概念或者比喻义。因此，隐喻性成语从隐喻程度性上来说，成语应该算作死寂隐喻(dead and buried)或者是非活跃隐喻(inactive metaphor) 22。隐喻涉及到语义的转移，这一转移带有方向性。一般情况下，喻源的有关特征被转
T=T1 + T2+ … + Tk，gT是某个 Tk的属性 gT。由此可定义 gT1 + gT2+ … + gTk, 使有 gT=gT1 + gT2+ … + gTk，若 T=T1 G T2 G … G Tk，gT是某个 Tk的属性 gT。由此可定义 gT1 G gT2 G … G gTk, 使有 gT=gT1 G gT2 G … G gTk.只要适当地定义关系的相应结合运算即可。等等。（从略）。
�（限于篇幅，略。本节与以下各节内容概从略）3.5 代换与概括等（尝试关于“具体”、“抽象”、“隐喻”等转换/变换的数学表述）4.（由此导出的关于信息元、情境等对象的信息量的序）（主要目的是讨论显式的给出型式的有限近似的问题）5. 结束语.从初步讨论看，自然语言语义学需要考虑建立它的数据/对象的类型理论。型式是信息内容与结构的抽象，是展开待开发的理论的核心概念。信息量的序关系简单合用。参考文献[1]Barwise,J.,J.Perry. Situations & Attitude, MIT Press, 1983; Re-issued by CSLI Publications, 1999[2]Devlin,K., Logic and Information, Cambridge University Press, 1991[3]Jon Barwise, Recent Developments in Situation Semantics, in Language and Artificial Intelligence, M. Nagao (ed.) 1987[4]Shalom Lappin, The Handbook of Contemporary Semantic Theory, Blackwell Publishers Ltd 1996, 1997[5] Cooper, Robin, The Role of Situations in Generalized Quantifiers, in S. Lappin[4][6]Hideki Tsuiki, A Domain-Theoretic Theory of Lax Generic Functions, Theor. Compt. Scien.v.294 (2003) 307
pp.1-36[8]Cooper, Robin, Three Lectures on Situation Theoretic Grammar, in Natural Language Processing, EAIA’90,2nd Advanced School in Artificial Intelligence(Guarda, Portugal, October 1990)Proceedings(102-140)[9]Jackendoff, Ray, Semantic Structures, Cambridge: The MIT Press,1990[10]卡纳普，世界的逻辑构造， 陈启伟译，上海，上海译文出版社，1999 （原著：Carnap, Rudolf, Der Logische Aufbau der Welt）[11]Montaque, Universal Grammar, in Formal Philosophy : selected papers of Richard Montague, Originally published in Theoria 36:373-98(1970)[12]Jacobson, Pauline, The Syntax/Semantics Interface in Categorical Grammar, in S. Lappin[4]
Abstract In previous work by Ahrens et al. (2003) and Chung et al. (2003a, 2003b), the extraction of conceptual metaphors was carried out based on prototypicality. Based on this view, the most prototypical mappings of source to target domains in the corpora help determining the Mapping Principle of the conceptual metaphors (Ahrens 2002). However, this approach is limited in the aspect that the identification of source domains must be manually determined. Therefore, the purpose of this paper is to suggest a way to reduce the manual work required for the determination of the source domain as well as to operationalize the steps to define different source domains. This paper incorporates the WordNet lexical representation and SUMO ontology in the identification of the related nodes to a group of linguistics expressions. The lexicons in the WordNet and SUMO are linked through works such as Niles and Pease (2003).
1 Introduction
The Conceptual Mapping (CM) Model (Ahrens 2002) proposed that a specific Mapping Principle can be generated through observing the source and target domain mappings in conceptual metaphors. For instance, the list of words in Table 1 below were suggested by Ahrens, Chung & Huang (2003) to be related to the source domain of PERSON. They extracted a total of 2000 instances from the Academic Sinica Balanced Corpus of Mandarin Chinese. From these instances, they extracted metaphorical expressions through observing the source-target domain mappings. For instance, chengzhang ‘growth’ below shows a mapping between the concrete source domain of PERSON and the abstract domain of ECONOMY. The analyses were carried out manually.
Table 1: ECONOMY IS A PERSON 成長 (growth) 衰退 (dysfunction) 成長期 (growth period) 病狀 (symptoms) 命脈 (lifeblood) 衰頹(weakness and degeneration) 成長 (grow) 衰退 (to become dysfunctional) 復甦 (regain consciousness) 惡化 (deteriorate) 恢復 (recover)
Ahrens, Chung and Huang (2003) then counted the number of instances of which these expressions occurred. Based on the most frequent (i.e., the most prototypical) mappings, the Mapping Principles (MP) were suggested. For instance, the MP for ECONOMY IS A PERSON in Table 1 is ‘economy is person because people have a life cycle and economy has growth cycle.’ In this paper, we further extend the work of Ahrens et al. (2003) but minimize the manual work required when determining the source domains. We suggest that the identification of source domains can be solidified through examining a) the WordNet senses and explanations
93
and b) the SUMO nodes and their definitions. By doing so, we propose a way to operationalize the steps in defining source domains. This extension of the CM Model is aided by studies linking the lexicon in WordNet and SUMO such as in Niles and Pease (2003). After the links were established, the ontological nodes in SUMO were provided hyperlinks to their semantic definitions by WordNet. Also facilitating this analysis is the work of Huang (2004) in which all the nodes in SUMO are available in both Chinese and English through the Academia Sinica Bilingual Ontological WordNet (BOW; available at http://bow.sinica.edu.tw).
2 Methodology
Our methodology involves two major steps, i.e., extracting metaphorical instances from the corpus and examining the WordNet explanation and SUMO nodes. For the corpus search, a total of 1062 instances of shiye事業 ‘career’ were extracted from the Academia Balance Corpus of Mandarin Chinese. All metaphorical instances were manually marked. A metaphorical mapping occurred when there was a mapping from the source (such as PERSON) to the target domain (CAREER). For instance, in the phrase shiye youhuan yishi 事業憂患意識 ‘ the worried consciousness of career,’ the abstract target domain CAREER is mapped onto the consciousness of PERSON. What makes this paper different is that we did not categorize all the expressions according to source domains at this stage. We used WordNet and SUMO instead to determine the categorization of the metaphorical expressions found.
3 Results
For the A total of 59 types (84 tokens) of metaphorical expressions were found. These instances are listed in Table 2 on the next page. Upon keying in the target word such as yishi ‘consciousness’ (highlighted in the Table 2) into Sinica BOW, there were a list of senses suggested by the WordNet lexical knowledge database. The most concrete sense was pulled out from the list to further search for their related SUMO nodes (which were provided after the explanation of each WordNet sense).
Table 2 : Metaphorical expressions related to shiye ‘career’Conceptual Metaphors Tokens Conceptual
Metaphors Tokens Conceptual Metaphors Tokens
新創 1 紮實 1 溶進...之中 1
創造 5 起(步) 1 收起來 1
開創 1 (走)向 2 投身...中 1
共創 1 第一(步) 1 基礎 5
再創 2 闖 2 追求 4
挑戰 1 關 5 風險 1
策略 2 關卡 1 供輸 1
趨勢 3 過程 1 賭 1
(幕後)功臣 1 前途 2 投向 1
異(軍)突起 1 前程 1 投下 1
改革 1 競爭 1 衝刺 1
兵符 1 抗爭 1 角色 1
憂患意識 1 打拚 1 大舞台 2
搖身一變 1 掙 1 階梯 1
創傷 1 拼 1 登上...位子 1
94
躍進 1 軌道 1 投入 1
放手 1 轟轟烈烈 2 玩掉 1
生命力 1 走(上) 1 包袱 1
壯大 1 火車頭 1 開發 1
成長 1 退出 2
Table 3 shows the example of yishi ‘consciousness’ as well as other examples taken from Table 2 in which their related WordNet senses and SUMO nodes are listed.
Table 3: Defining Source Domains through WordNet and SUMOExpressions WordNet
Senses Explanations SUMO Nodes
憂患(意識) 1: consciousness an alert cognitive state in which you are aware of yourself and your situation Awake ( 清醒 )
挑戰 4: challenge a call to engage in a contest or fight Requesting ( 請求 )
策略 6: ambush the act of concealing yourself and lying in wait to attack by surprise
ViolentContest( 暴力性的競爭 )
基礎 1: foundation lowest supporting part of a structure Region( 區域 ) 軌道 2: track a pair of parallel rails providing a
runway for wheelsTransportationDevice( 運輸工具 )
For the expressions in Table 3, we first searched for their WordNet senses and examined their WordNet explanations. The shaded words in Table 3 above provided the cue to which source domains these expressions belong. For instance, the concept of cognitive state might refer to a PERSON; contest and attack to WAR whereas rails and wheels to VEHICLE. However, the sorting of source domains using the WordNet alone was not enough. For instance, the WordNet explanation regarding jichu ‘foundation’基礎 in Table 2 yielded the idea structure. An analysis using the WordNet alone cannot identify whether it is referring to the structure of an artifact or any other object. Hence, the second step involved examining the related SUMO nodes. All WordNet senses are linked to the related SUMO nodes, which are listed on the rightmost column of Table 2. Through examining the SUMO nodes, we decide whether a group of expressions belongs to the same concept, i.e, the concept of the source domain onto which the target term CAREER was mapped. In order to display more specifically which ontological nodes correspond to one another, this paper shows the mechanisms through using the instances from CAREER IS A PERSON
4 Defining the Source Domain of PERSON
In this section, we unfold how the source domain of PERSON was obtained. The expressions in (1) below were found related to the source domain of PERSON.
(1) 憂患(意識) ‘ the worried consciousness’創傷 ‘wound’放手 ‘let go’生命力 ‘the force to live’成長 ‘grow’起(步) ‘start a step’(走)向 ‘walk towards’第一(步) ‘first step’
For these expressions, the search in WordNet and SUMO produced the explanations in Table 4:
Although no obvious pattern was found among the nodes in (2), when their definitions were searched for in SUMO, these nodes showed linkage to one another, as in (3). The linking concept, i.e., ‘Organism,’ is shaded in all the definitions in (3).
(3) (a) Awake: Attribute applies to Organisms that are neither Unconscious nor Asleep.(b) EmotionalState: The Class of Attributes that denote emotional states of Organisms.
(c) Organism: Generally, a living individual, including all Plants and Animals. (d) Growth: The Process of biological development in which an Organism or part of
an Organism changes its form or its size.
(e) Walking : Any BodyMotion which is accomplished by means of the legs of an Organism on
land for the purpose of moving from one point to another.
From investigating the SUMO definitions, all the expressions in Table 4 were found related to ‘Organism’ in one way or another. Among these five SUMO nodes in (3), 3(a) and (b) are related to the upper node of ‘State of mind’ or a psychological process. 3(c) and (d) are linked to the ‘Organism’ and the ‘Organism Process’ or biological process. ‘Walking’ in which (3e) is related ‘Body motion’ in SUMO. What can be concluded from this discussion is that there were overlaps between all the expressions in Table 4 which were identified through the CM Model. The linking concept between them is ‘Organism.’ However, the concept of ‘Organism’ is too broad because it comprises all living things, including all plants and animals. For the conceptual metaphors, we are looking for source domains that are in contact with human conceptualization, i.e., more concrete concepts which human can easily recall when describing the abstract idea. Furthermore, we are not only looking for any organisms which have the abilities to grow and to walk, but one with a state of mind and emotional state. Based on these attributes, we rule out the possibility of PLANT and ANIMAL, but suggested that these expressions are related
In order to verify our findings regarding CAREER IS A PERSON, this paper also carried out the same procedures with the target domain of CULTURE. A total of 2000 instances was extracted from the Academia Balance Corpus of Mandarin Chinese. From these 2000 instances, 335 metaphorical expressions were extracted manually. From these 335 instances, 88 were found related to the source domain of PERSON. These instances constituted the following related WordNet senses and SUMO nodes (See Table 5).
Table 5: CULTURE IS A PERSON and its related WordNet senses and SUMO NodesExpressions WordNet
Senses Explanations SUMO Nodes
生命 3: life the organic phenomenon that distinguishes living organisms from nonliving ones Living ( 活的 )
思想 2: thought the organized beliefs of a period or group or individual Proposition ( 命題 )
成長 6: growth
the process of an individual organism growing organically; a purely biological unfolding of events involved in an organism changing gradually from a simple to a more complex level
Growth ( 生長 )
反省 3: introspectiveness thoughtfulness about your own situation and feelings
TraitAttribute( 人格特質 )
生命力 1: animation the property of being able to survive and grow Living ( 活的 )
氣質 1: temperamentexcessive emotionalism or irritability and excitability (especially when displayed openly)
PsychologicalAttribute( 心理屬性 )
命脈 1: lifeblood the blood considered as the seat of vitality Blood ( 血液 )
心靈 2: soul the immaterial part of a person; the actuating cause of an individual life Human ( 人類 )
(走) 向（v) 1: foot informal or colloquial synonyms of 'walk' " Walking ( 行走 )
決定 3: decide reach, make, or come to a decision about something
IntentionalPsychologicalProcess( 意向性心理歷程 )
行為 4: behavior manner of acting or conducting oneself BodyMotion ( 身體的移動 )
滅絕 1: die_out become extinct Death ( 死亡 )
From Table 5, the SUMO nodes in the rightmost columns were found related to HUMAN (or PERSON). These nodes have the following definitions in SUMO, all of which are related to Organism in one way or another:
(4) (a) Living ( 活的 ) : This Attribute applies to Organisms that are alive.(b) Proposition ( 命題 ) :Propositions are Abstract entities that express a complete
thought or a set of such thoughts.(c) Growth ( 生長 ) : The Process of biological development in which an Organism or
part of an Organism changes its form or its size. (d) TraitAttribute ( 人格特質 ) : Attributes that indicate the behavior/personality traits
of an Organism.(e) PsychologicalAttribute ( 心理屬性 ) : Attributes that characterize the mental or
behavioral life of an Organism.(f) Blood ( 血液 ) : A fluid present in Animals that transports Nutrients to and waste
products away from various BodyParts.
(g) Human( 人類 ) : Modern man, the only remaining species of the Homo genus.(h) Walking( 行走 ) : Any BodyMotion which is accomplished by means of the legs of
an Organism on land for the purpose of moving from one point to another.
(i) IntentionalPsychologicalProcess( 意向性心理歷程 ) : An IntentionalProcess that can be realized entirely within the mind or brain of an Organism.
(j) BodyMotion( 身體的移動 ) : Any Motion where the agent is an Organism and the patient is a BodyPart.
(k) Death( 死亡 ) : The Process of dying.
The target domain of CULTURE produced more instances than the CAREER metaphor. Therefore, a comparison can be made. From the information in Table 5, the first decision that the source domain could be PERSON was observed through the SUMO nodes, all of which can be related to HUMAN. The further investigation of their SUMO definitions in (4) found that these nodes are related to ‘Organism’ by possessing ‘thoughts,’ ‘biological development,’ ‘animals,’ ‘body parts,’ ‘modern man,’ ‘Homo genus,’ ‘body motion,’ ‘mind or brain’ and ‘the process of dying.’ In other words, the analysis showed that there is a connectivity from the beginning linguistic expressions (metaphorical instances) to the WordNet explanations and SUMO nodes (Table 5) and lastly to the explanation of the SUMO nodes (example (4) above). The connectivity can be drawn by first defining the linguistic expressions within a source domain, as shown in Figure 1:
Figure 1: Representations of the Metaphorical Instances According to Source and Target Domians and their Relationship with WordNet and SUMO
The overlapping area between CULTURE and CAREER in Figure 1(a) constitutes lexical items that are more lexicalized as they can apply to more target domains than the other items. In other words, lexical items within several overlapping source domains (such as chengzhang 成長 ‘growth’ and zouxiang (走)向 ‘walk towards’) tend to be lexicalized faster than the other lexical items. All the lexical items in 1(a) can be mapped onto the ontological concepts of SUMO in 1(b).
6 Conclusion
The results from the WordNet lexical representation and SUMO searches prove two major points: First, the manual analysis made using the CM Model in previous studies can be made more automatized. Second, there is conceptual connectivity between the expressions found within a source domain. This connectivity can be established through observing the overlapped concepts in all the expressions within the same source domain. The approach proposed by this paper takes a linguistic perspective in which instances are first analyzed based on a linguistic model of metaphors. In order to computerized the process of extracting source domain from the instances, a large amount of data has to be collected. The project has covered numerous source domains with corpora instances of metaphorical examples. When these instances are fully coded with the linguistic information, a computerized program can be designed for the extraction purposes. From the perspective of the linguists, this analysis provides the evidence to pinpoint at which level of knowledge conceptual metaphors occur. From the perspective of anthropology and language processing, this study provides the linguistic evidence to understand the mental representation of human when expressing metaphors.
AcknowledgementsWe would like to thank the CLSW-5 reviewers for their comments on this paper.
References Ahrens, Kathleen. 2002. “When Love is Not Digested: Underlying Reasons for Source to Target Domain
Pairings in the Contemporary Theory of Metaphor.” In Yuchau E. Hsiao (ed.). Proceedings of the First Cognitive Linguistics Conference. Cheng-Chi University, 273-302.
Ahrens, Kathleen, Chung Siaw-Fong, Huang Chu-Ren. 2003. “Conceptual Metaphors: Ontology-based Representation and Corpora Driven Mapping Principles.” In the Proceedings of the ACL Workshop on the Lexicon and Figurative Language. pp. 35-41.
Chung, Siaw Fong, Kathleen Ahrens and Huang Chu-Ren. 2003a. “ECONOMY IS A PERSON: A Chinese-English Corpora and Ontological-based Comparison Using the Conceptual Mapping Model.” Proceedings of the 15th ROCLING Conference, National Tsing-Hwa University, Taiwan.
Chung, Siaw Fong, Kathleen Ahrens and Huang Chu-Ren. 2003b. “ECONOMY IS A TRANSPORTATION_DEVICE: Contrastive Representation of Source Domain Knowledge in English and Chinese.” Proceedings of the special session of UONLP, 2003 International Conference on Natural Language Processing and Knowledge Engineering (NLP-KE), Beijing.
Huang, Chu-Ren, Ru-Ying Chang, Shiang-Bin Lee. 2004 “Sinica BOW (Bilingual Ontological Wordnet): Integration of Bilingual WordNet and SUMO.” To appear in the Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC2004). Lisbon. Portugal. 26-28 May, 2004.
Lakoff, George and Mark Johnson. 1980. Metaphors We Live By. Chicago: The University of Chicago Press. Niles, I. and Pease, A. 2003 “Linking Lexicons and Ontologies: Mapping WordNet to the Suggested Upper
Merged Ontology.” In Proceedings of the 2003 International Conference on Information and Knowledge Engineering,, Las Vegas, Nevada.
100
From Lexical Semantics to Conceptual Metaphors: Mapping Principle Verification with WordNet and SUMO
Abstract The goal of this paper is to further develop methods for verifying Mapping Principles between source and target domain pairings of conceptual metaphors. Previous work (Ahrens, Chung & Huang, 2003) integrated the Conceptual Mapping Model with an ontology-based knowledge representation (i.e. SUMO) in order to demonstrate that conceptual metaphor analysis could be restricted and eventually, automated. However, that study relied on an operational definition that required a large number of instances of source-target domain pairings in order to determine the most frequent mapping. In this study, we further operationalize the frequency-based definition and examine the possibility of using information derived from WordNet and SUMO to verify Mapping Principles for source-target domain pairings that do not occur frequently enough (if at all) in a corpus.
1 Introduction
Lexical resources are central to any natural language processing system and the plethora of multi-lingual resources such as EuroWordNet (EWN) and Academia Sinica’s Bilingual Ontological WordNet (BOW) are based firmly in the lexicon. Recent work in figurative language processing, however, has pointed out shortcomings of these resources for processing of conceptual metaphor. Veale (2003), for example, points out that determining the aptness is dependent upon parameters that are, as yet, hidden. Lonneker (2003) decries the lack of relationship between literal and figurative synsets in WordNet, while Alonge and Castelli (2003) suggest that the EWN Top Ontology needs to be extended with more concepts in order to deal with figurative language, since at this point in time it only notes general semantic distinctions. Ahrens (2002), taking an intuition-based approach, proposed that the linguistic expressions that are used metaphorically can be analyzed in terms of the entities, qualities and functions that can map between a source and a target domain. When these conventionalized metaphorical expressions have been analyzed, an underlying reason for these mappings can then be postulated. This study relied on native speaker intuition regarding what is mapped conceptually from the source to the target domain. For example, in the four examples from the metaphor LOVE IS PLANT, given below, the Mapping Principle (MP) of “Love is understood as plant because plants involve physical growth and love involves emotional growth” was extracted based on the fact that all the examples in some way had to do with growth.
1. (a) liang ren de ai miao zuijin cai gang mengya two people MOD love seedling lately just recently sprout ‘Their love just begins to sprout lately.’
(b) wo dui tade ai-yi jianjian zizhang I for his love gradually grow ‘My love for him has grown gradually.’
101
(c) aiqing xuyao xinqin de quanqai love need industriously water
‘Love needs to be watered industriously.’
Ahrens, Chung & Huang (2003), followed up on this study and presented a corpus-based approach to determining the systematicity between source and target domain pairings (i.e. Mapping Principles (MPs)). They propose that each source-target domain pairing will have a prototypical instance of mapping as indicated by a lexical item that is frequently mapped, as compared with other mappings. In addition they use the Suggested-Upper-Merged-Ontology (SUMO) to delimit the source domain knowledge. SUMO has the advantages of being able to be used to infer knowledge through automatic reasoning as well as to constrain the falsifiability of the MP. For example, when they looked at the metaphor ECONOMY IS A BUILDING, they find 102 instances (in their randomly extracted corpora of 2000 instances from Sinica Corpus) of this source-target domain pairing. Of these instances, 39 are mapped with the lexical item jianshe ‘construction’, and the suggested Mapping Principle is therefore: Economy is a building because buildings involve a (physical) structure and economy involves an (abstract) structure (p. 38).” However, clear numerical criteria for determining the minimal number of tokens necessary, as well as the percentage of tokens necessary to reach consensus for a mapping principle was not established. In addition, if enough conventional metaphor examples cannot be found in the corpora, there is currently no way to verify the intuition-based mapping principle. Thus, the determination of mapping principles by means of frequency needs to be researched further. In this paper, then, we will create guidelines for determining when MPs can be established on the basis of frequency, as well as extend Ahrens et al.’s (2003) scope and examine the possibility of using information derived from WordNet and SUMO to verify MPs for source-target domain pairings that do not occur frequently enough in a corpus.
2 Using Frequency to Verify Mapping Principles
In this section, we attempt to determine frequency criteria for verifying MPs (Table 1). Two thousand examples were extracted for ECONOMY metaphors from the Academia Sinica Balanced Corpus (http://www.sinica.edu.tw/SinicaCorpus/), and then manually sorted into metaphorical and literal examples. In addition, we also look at data on STOCK MARKET extracted from on-line newspapers in Taiwan (Chung, Ahrens & Sung 2003).
Table 1: Total Number of Metaphorical Instances Extracted from 2000 SamplesMetaphor Total # of
Tokens Most Frequent Percentage MPs matchECONOMY IS A PERSON 121 88 72% YesECONOMY IS BUILDING 102 39 38% YesSTOCK MARKET IS OCEAN WATER 58 53 91.00% YesECONOMY IS COMPETITION 40 14 35% YesSTOCK MARKET IS WAR 25 2 8% Not clearSTOCK MARKET IS A BALL 24 17 71.00% YesECONOMY IS WAR 23 5 21% Not clearSTOCK MARKET IS A PERSON 19 2 10% Not clear
STOCK MARKET IS EARTHQUAKE 17 16 94.12% YesSTOCK MARKET IS DISASTER 16 11 68.75% YesECONOMY AS JOURNEY 15 5 33% YesECONOMY IS AN AEROPLANE 10 8 80% Yes
102
The second column in Table 1 lists the total number of metaphorical examples found for each source domain. Next, examples with the same lexical item were counted. For example, in ECONOMY IS A PERSON, ‘chengzhang’ growth is found 88 times or in 72% of the data. In ECONOMY IS A BUILDING, ‘jianshe’ construction occurs 39 times (or in 38% of the data). In both these cases (and in all the shaded cases in the table) the most frequent lexical item is related to what was originally postulated as a Mapping Principle on the basis of linguistic intuition. However, in the three non-shaded cases, the relationship between the frequency of the lexical items and the Mapping Principle is not straightforward. For example, in STOCK MARKET IS WAR, no one lexical item receives priority in terms of frequency. In ECONOMY IS WAR, there are five instances of ‘gua-shuai’ to take position of the commander in the army, but this does not clearly relate to the MP of “Economy is war because war involves a violent contest for territorial gain and the economy involves a vigorous contest for financial gain.” Therefore, we propose that the following two conditions be met for a frequency-based account to be valid: First, at least 10 tokens must found and, in addition, at least 30% of these tokens must be of one particular lexical item.
3 Using WordNet and SUMO to Verify Mapping Principles
In the cases where there are fewer than 10 instances found in the corpora, or in cases where the percentage of a single lexical item is less than 3o%, we need to find another method to verify MPs. Our goal is aided by studies linking the lexicon in WordNet and SUMO (http://ontology.teknowledge.com, Niles & Pease, 2003). After the links are established, the ontological nodes in SUMO provide hyperlinks to their semantic definitions by WordNet. Also facilitating this analysis is the work of Huang (2004) in which all the nodes in SUMO are available in both Chinese and English through the Academia Sinica Bilingual Ontological WordNet (http://bow.sinica.edu.tw). We examine two instances here: LOVE IS A PLANT (Table 2) and LOVE IS FOOD (Table 3).
Table 2: LOVE IS PLANT Definitions from WordNet and SUMOItems WordNet
Senses Explanation SUMO Category
萌芽 2: sprout grow sprouts, of a plant Growth ( 生長 )
苗 1: seedling young plant or tree grown from a seed FloweringPlant ( 開花植物 )
長 1: grow come to have, of physical features and attributes Growth ( 生長 )
灌溉 1: water pour water on Wetting( 濕潤 ) 開花 1: bloom produce or yield flowers Growth ( 生長 )
Ahrens (2002) proposes the following MP for LOVE IS A PLANT: “Love is understood as plant because plants involve physical growth and love involves emotional growth.” Since corpora searches do not come up with any instances of this metaphor, it is difficult to ascertain the validity of this principle. We therefore propose looking at the 1) WordNet sense, 2) the WordNet definition, and the 3) SUMO node for the WordNet sense (Niles & Pease 2003) for the intuition-based examples in order to see if there are any semantic overlaps within, or between, these three types of information. Table 2 shows that the word “Growth” appears three times in the SUMO category, out of the five examples. “Grow” also appears three times in the sense and definition columns from WordNet. Thus, the combination of WordNet information and the SUMO representation agrees with the MP originally given. In another example that has less than ten corpora examples, LOVE IS FOOD (Table 3), both the WordNet information and the SUMO information again matches up with the Mapping Principle suggested in Ahrens (2002), that “Love is understood as food because food
has different tastes as love involves different feelings.” Table 3 shows that taste is mentioned five times in the WordNet sense and definition, and three out of four times in the SUMO category. Thus, determining the number of overlapping lexical items in WordNet definitions and SUMO categories to verify Mapping Principles seems to hold promise for instances where there are not enough exemplars to make a judgment based on frequency alone.
Table 3: LOVE IS FOOD: Definitions from WordNet and SUMOItems WordNet
Senses Explanation SUMO Nodes
滋味 1: taste distinguishing a taste by means of the taste buds Tasting ( 嚐 )
苦 1: bitter one of the four basic taste sensations; sharp and disagreeable; like the taste of quinine
TasteAttribute ( 味覺屬性 )
味道 1: taste distinguishing a taste by means of the taste buds Tasting ( 嚐 )
吃 1: eating the act of consuming food Eating( 吃 )
4 Testing the Verification Principle
In Section 2, we pointed out that there were three instances in the STOCK MARKET and ECONOMY metaphors that did not meet the 30% criteria. We look at these three metaphors below in order to determine if WN and SUMO can be used to verify the MPs.
Table 4: ECONOMY IS WAR: Definitions from WordNet and SUMOItems WordNet
Senses Explanation SUMO Nodes侵略 4:
invasion the act of invading; the act of an army that invades for conquest or plunder
ViolentContest ( 暴力性的競爭 ) 戰 1: war the waging of armed conflict against an
enemy War ( 戰爭 ) 老兵 1: veteran a serviceman who has seen considerable
active serviceSocialRole ( 社會角色 ) 攻防戰 x策略 6:
ambush the act of concealing yourself and lying in wait to attack by surprise
ViolentContest ( 暴力性的競爭 ) 大權 x千瘡百孔 全力(衝刺) x
掛帥 x大權…操在…的手上 x
犧牲 1: sacrifice kill or destroy Killing( 殺害 )
犧牲品 1: sacrifice
personnel that are sacrificed (e.g., surrendered or lost in order to gain an objective) Human( 人類 )
Table 4 shows that there are three instances of the concept of ‘Invasion’ found in the WN definitions, but they are all in the same definition. An alternate hypothesis is that ‘ViolentContest’ is the critical issue since it occurs in the SUMO nodes of two different words. (The blank cells indicate that this word was not found in BOW and so does not have a
listed sense or associated SUMO node.) In addition, Ahrens et al. (2003) noted that ECONOMY IS WAR is a subset of the ECONOMY IS CONTEST metaphor, with the MP of ‘Economy is war because war involves a violent contest for territorial gain and the economy involves a vigorous contest for financial gain.’ Moreover, the SUMO node of WAR is linked to ViolentContest. This example demonstrates that not only do we need to have an expansion of the SinicaBow to link to more items in WordNet, we also need to expand our notion of semantic space to include related Sumo nodes. In sum, our current analysis suggests that the previous MP was correct. Table 5 gives instances of STOCK MARKET IS WAR. Although different lexical items are mapped as compared with ECONOMY IS WAR, the mapping of the concept of ‘contest’ is the same.
105
Table 5: STOCK MARKET IS WAR: Definitions from WordNet and SUMOItems WordNet
Senses Explanation SUMO Nodes
戰 1: war the waging of armed conflict against an enemy War ( 戰爭 )
策略 6: ambush the act of concealing yourself and lying in wait to attack by surprise
ViolentContest ( 暴力性的競爭 ) 焦土 x
地雷 1: land_mine
an explosive mine hidden underground; explodes when stepped on or driven over Weapon ( 武器 )
攻勢 x
關卡 1: checkpoint
a place (as at a frontier) where travellers are stopped for inspection and clearance LandArea ( 陸地 )
大關 x防衛 2: defend be on the defensive; act against an attack Contest ( 競爭 ) 功臣 x全軍覆沒 x殺空 x撤出 x殺出 x
上(攻) 1: attack take the initiative and go on the offensive: "The Serbs attacked the village at night" Contest ( 競爭 )
挑戰 4: challenge a call to engage in a contest or fight Requesting ( 請求 )
撤離 1: evacuation
the act of evacuating; leaving a place in an orderly fashion; esp. for protection Motion ( 移動 )
Thus, the proposed MP is the same; ‘Stock market is war because war involves a violent contest for territorial gain and the stock market involves a vigorous contest for financial gain.’ Note that this does not violate the Mapping Principle Constraint proposed in Ahrens (2002) since two different target domains are recruiting the same source domain for the same reason. The Mapping Principle Constraint says that ‘A target domain will select only source domains that involve unique mapping principles.’ This constraint applies to the source domains that are selected by a single target domain.
The last metaphor in our study that does not meet the frequency requirement of 30% is STOCK MARKET IS A PERSON. Although no one lexeme predominates, variants of the word ‘life’ occur several times. It is interesting to note that although in the previous examples a case could be made for looking solely at the SUMO nodes (cf. Tables 2-5 above), in this instance, only the words used in the WN definition are found to coalesce around a concept. In this instance, we suggest that the MP is: Stock market is a person because people have varying degrees of physical activity (liveliness) and the stock market has varying degrees of financial activity. It may also be beneficial to look at whether analyzing the relationship between the words given in the definitions (i.e. if some words are near-synonyms) would aid in determining the underlying reason for the source-target domain pairing. However, another possibility is that this source domain is too broad or not clearly delimited. In fact, one step that needs to be taken in this regard is to incorporate the steps that Chung, Ahrens and Huang (2004) take to delimit source domains before analyzing MPs.
Table 6: STOCK MARKET IS A PERSON: Definitions from WordNet and SUMOItems WordNet
Senses Explanation SUMO Nodes
活力 4: pep liveliness and energy BiologicalAttribute ( 生 物屬性 ) 創傷 2: wound any break in the skin or an organ caused by
violence or surgical incision Injuring( 傷害 )
瀕 (死) 的 2: die pass from physical life and lose all bodily attributes and functions necessary to sustain life
Death( 死亡 )
敏感的 x疲軟 x
激勵 12: spur give heart or courage to
IntentionalPsychological Process ( 意向性心理歷程 ) 輸血 x
起死回生 1: resurrect cause to become alive again OrganismProcess ( 生命 歷程 )
變 (臉) 1: face the front of the head from the forehead to the chin and ear to ear BodyPart ( 軀體部件 )
應聲倒地, 應聲而倒 x
復甦 1: come_to
return to consciousness: "The patient came to quickly"
BiologicalProcess ( 生物 歷程 ) 失血 x
5 Conclusion
We propose the following steps to determine Mapping Principles: First, if the number of tokens of a conceptual metaphor is greater than 10, and the highest number of tokens mapped is greater than 30% of all tokens, then this lexical item is the basis for the MP. Second, if either the first or second criteria are not met, determine the sense of each item in WordNet (extracted through SinicaBow) and the associated SUMO category for that particular sense. Next calculate the number of times that a particular lexical item is used in the WN sense, explanation, or linked SUMO category and determine if that lexical item is related to what intuitively seems to be mapped between the source and target domains in the example sentences. Further research will focus on examining whether it is necessary to extend the parameters of the second step to include synonyms, thus expanding the semantic space by which a MP can be determined. In addition, the biggest hurdle this type of account faces is that the WN sense must currently be stipulated manually, although we hope that the implementation of semantic tagging can overcome this problem in the future. We hope that the line of research discussed herein will stimulate more research in how a computational approach can help set parameters for determining metaphorical sense and point the way to the application of creating a systematic relationship between literal and figurative synsets in WordNet.
AcknowledgementsWe would like to thank the CLSW-5 reviewers for their comments on this paper and NSC grant #NSC92-2411-H-002-076-ME for supporting the research discussed herein. Any remaining errors are the sole responsibility of the authors.
References Ahrens, Kathleen. 2002. “When Love is Not Digested: Underlying Reasons for Source to Target Domain
Pairings in the Contemporary Theory of Metaphor.” In Yuchau E. Hsiao (ed.). Proceedings of the First Cognitive Linguistics Conference. Cheng-Chi University, 273-302.
Ahrens, Kathleen, Chung Siaw-Fong, Huang Chu-Ren. 2003. “Conceptual Metaphors: Ontology-based Representation and Corpora Driven Mapping Principles.” In Proceedings of the ACL Workshop on the Lexicon & Figurative Language. pp. 35-41.
Alonge, Antonietta & Margherita Castelli. 2003. “Encoding Information on Metaphoric Expressions in WordNet-like Resources.” In Proceedings of the ACL Workshop on the Lexicon and Figurative Language. pp. 10-17.
Chung, Siaw-Fong, Kathleen Ahrens and Chu-Ren Huang. 2004. “Using WordNet and SUMO to Determine Source Domains of Conceptual Metaphors.” In Proceedings of the 5th Chinese Lexical Semantics Workshop (CLSW-5), Singapore. June 14-15.
Chung, Siaw Fong, Kathleen Ahrens and Sung Yahui. 2003. “STOCK MARKETS AS OCEAN WATER: A Corpus-based, Comparative Study in Mandarin Chinese, English and Spanish.” In the Proceedings of the 17th Pacific Asia Conference on Language, Information and Computation (PACLIC), Singapore. pp. 124-133.
Lonneker, Birte. 2003. “Is There a Way to Represent Metaphors in WordNets? Insights from the Hamburg Metaphor Database.” In the Proceedings of the ACL Workshop on the Lexicon and Figurative Language. pp. 18-26.
Huang, Chu-Ren, Chang, Ru-Yng, Lee, Shiang-Bin. (2004) Sinica BOW (Bilingual Ontological Wordnet): Integration of Bilingual WordNet and SUMO". To be presented at the 4th International Conference on Language Resources and Evaluation (LREC2004). Lisbon. Portugal. May 26-28.
Niles, I. and Pease, A. 2003 “Linking Lexicons and Ontologies: Mapping WordNet to the Suggested Upper Merged Ontology.” In Proceedings of the 2003 International Conference on Information and Knowledge Engineering, Las Vegas, Nevada.
Veale, Tony. 2003. “Systematicity and the Lexicon in Creative Metaphor.” In Proceedings of the ACL Workshop on the Lexicon & Figurative Language. pp. 27-34.
摘要本文将情感评价属性特征纳入现代汉语语法信息词典的词语属性描述体系，基于人民日报基本标注语料库，探讨以定性和定量相结合的方式对汉语词语的情感标注进行研究。根据真实文本实例的统计、归纳，对词典中词语的情感倾向加以描述，然后在词典中形式化。词语的情感评价属性的计算处理对文本过滤、信息抽取、网页评价等有重要的参考价值。关键词 词语情感评价 语义韵律 搭配规律Research on Word Emotional Evaluation Based on The Grammatical Knowledge-base of Contemporary Chinese
Wang Zhimin Zhu Xuefeng Yu ShiwenInstitute of Computational Linguistics Peking University 100871 China
This paper introduces the attributes of emotional evaluation into The Grammatical Knowledge-base of Contemporary Chinese. The word emotions tagging is studied by the approach of using both qualitative and quantitative methodology. Based on the statistics result in People’s Daily tagging corpus, the word emotional trends are described and formulized in our Knowledge-base. Moreover, there are some potential applications of emotions tagging in such related fields as text filtering, information retrieve and webpages’ evaluation. Keywords : evaluation of word emotions semantic prosody collocation regulation
语言是动态的，静态词库里的词语在实际的语言应用中会受到其搭配词语的影响，一个本来没有多少褒贬意义的词语在进入句子框架后可能会表现出强烈的情感倾向。这方面的研究，国外已经开始得到重视。他们称之为语义韵律。Partington（1996）将语义韵律定义为“超越单个词语界限的色彩意义的扩展”(Semantic prosody refers to the spreading of connotational coloring beyond single word boundaries) ，语义韵律由音韵学上的“韵律”概念转化而来，音韵学上的韵律现象是用来概括语音研究中切分成分在语流中具有超切分特征。音韵学中的同化、异化、连读都属于韵律的研究范畴。受这种现象的启发，J.Sinclair把韵律加以推广，认为这种现象同样存在语言的词汇层面。J.Sinclair(1987)利用大规模的语料库资源对于词语的韵律进行了搭配意义方面的研究,指出观察词语意义的习惯搭配是揭示语义韵律的最佳途径。综合国内外的研究成果，词语的情感评价应该从两个层面来考虑。一方面应该是静态词汇
博士提出了很好的建议，在此向他们表示衷心的感谢。 参考文献Parington ,Alan.1996. Patterns and Meanings :Using Corpus for English Language Research and Teaching.pp:68Palmer,F. ,ed. 1968. Selected Papers of J.R.Firth. Landon:Longman. 1952-1959 Sinclair, J.M.1987.Looking Up.London and Glasgow;William Collins.Sinclair, J.M.1991.Corpus.Concordance,Collocation. Oxford :Oxford University Press.戴维.克里斯特尔[英]，沈家煊译，2002，现代语言学词典 商务印书馆 第 4版。俞士汶等，2003，现代汉语语法信息词典详解 清华大学出版社 （第 2版）pp:40-41。应英等，2002，汉语情感意义的机器标注研究初探 中文信息学报 （第 2期）。符淮青，1985，现代汉语词汇 北京大学出版社 pp:28。刘叔新，1990，汉语描述词汇学 商务印书馆 pp:11。北京大学中文系现代汉语教研室，1997，现代汉语 商务印书馆。吕叔湘，1979，汉语语法分析问题，商务印书馆。吕叔湘，1980，现代汉语八百词，商务印书馆。
2.2名词归类实践我们以国家语委、北京大学计算语言学研究所、山西大学计算机应用研究所编制的三份比较有权威性、代表性的词表为蓝本，使用的检索工具是山西大学开发的“中文语料研究软件系统（1.0版本）”，与之配套的语料库已经过分词、词性标注和人工校对，可信度比较高。碰到数据稀疏时，再利用其他语料库，直至用 google 进行网上在线检索，平均搜索语料总字数达2000万字左右。这样，得到按义项和功能分布设立的约 4万个名词项。目前，我们初步提出的名词多级分类标准和词表，已交由山西大学计算机应用研究所算法化后编制成程序输入计算机，正等待实际结果的反馈，以做后续的修订完善。
Abstract: This paper at first presents the general idea of the research on Chinese semantic word-formation, then introcuduces the relevant aspects regarding the realization of a Chinese semantic word-formation database with 52366 disyllabic compound words, the distribution of characters from semantic point of view, the types of relationships between character senses and word senses, and the rules of semantic word-formation, based on the database. The characteristics of semantic word-formation rules are further discussed at last.Key words: Chinese semantic word-formation, lexical semantics, database, Chinese information processing
双音合成词语义构词的具体规则，并将这些规则进一步归纳为四个大的类型。下列规则中“A、B、C、D、E、F、G、H、I、J、K、L”为语义类大类的标记，具体规则中“AB”表示双音合成词中前一个字位的语义类为 A 类、后一个字位的语义类为 B 类，其他类推。6．1同类规则构成双音合成词的两个字位属于同一个语义类，所构成的词的语义类与其基本相同。AA
的词义 100%为 A 类，BB的词义 88.89%为 B 类，CC的词义 83.82%为 C 类，DD的词义 86. 83%为 D 类，EE的词义 84.98%为 E 类，FF的词义 68.03%为 F 类，GG的词义 84.5%为 G 类，HH的词义 88.41%为 H 类，II的词义 74.12%为 I 类，JJ的词义 74.72%为 J 类，KK的词义82.08%为 K 类，LL的词义 60%为 L 类。可见，除了 II、JJ、LL三类外，其他类构成的词义与其同类的都在 80%以上。属于同类构成的双音合成词共有 17565 个，占33.54%。6．2 后向型规则构成双音合成词的两个字位属于不同的语义类，所构成的词的语义类与后一个字位的语义
那么，如何才能让计算机识别在什么情况下三个处所名词构成 a 类关系，什么条件下构成 b 类关系，什么条件下构成 c 类关系？本文作者认为可以尝试使用特征核查的方法来处理对其作出形式化的描写。例如可以首先对每个词进行语义特征分析：中国＝［+国家、+东方的、+发展中的…］日本＝［+国家、+东方的、+发达的…］瑞士＝［+国家、+西方的、+发达的…］
Feature Representations and Logical Compatibility between Temporal Adverbs and Aspects
Shih-Min Li, Su-Chu Lin, Keh-Jiann ChenCKIP, Institute of Information Science, Academia Sinica, Taipei{shihmin, jess} @hp.iis.sinica.edu.tw ; [email protected]
Abstract In this paper, we propose clear-cut definitions to distinct temporal adverbs and provide descriptive features for each class of temporal adverbs. By adopting the corpus-based approach and measuring time points in temporal axis, the temporal adverbs listed in Lu & Ma 1999 are revised and reclassified into four main classes namely, time, frequency, duration, and time manner. The descriptive features suffice to discriminate temporal relations and to predict logical compatibility between temporal adverbs and aspects.
1 IntroductionThere are about 130 temporal adverbs in Mandarin Chinese. Lu & Ma classify the temporal adverbs into two groups, speaking-time related adverbs (abbr: ST-related adverbs, 定時時間副詞) and reference-time related adverbs (abbr: RT-related adverbs, 不定時時間副詞). The ST-related adverbs consist of 27 temporal adverbs, which are subdivided into three subclasses. In the class of RT-related adverbs, 104 temporal adverbs are listed and subdivided into 18 subclasses. Lu and Ma’s subdivision of temporal adverbs based upon aspects of situations. However, the subdivision is vague and the definition is ambiguous. For example, cengjing 曾經, ceng 曾, yeyi 業已 and yejing 業經 are grouped into two different subclasses of ST-related adverbs. The former two cengjing and ceng are grouped into the same subclass, which express the actions or situations have been existed or happened before speaking time. The later two yeyi and yejing are grouped into the same subclass, which indicate the actions or situations have been completed or occurred. In fact, it is difficult to differentiate the actions or situations that have been ‘happened’ from those ‘completed’, especially when the situation type is achievement with SHORTLY-PRECEDE(t1,t2) or NEARLY-EQUAL (t1,t2).75 Moreover, temporal adverbs may not have the same syntactic behaviour even though they are classified into the same subclass. For instance, the ST-related adverbs cong 從, conglai 從來, zhijin 至今, xianglai 向來, sulai 素來, lilai 歷來, su 素, and yixiang 一向 are grouped into the same subclass. When co-occurring with aspect markers le 了, guo 過 and zhe 著, cong, conglai and zhijin are incompatible with le and zhe; however, xianglai, sulai, lilai, su and yixiang are incompatible with le and guo. The cause of the difference in the compatibility of temporal adverbs with aspects will be also under discussion.
In this paper, we propose clear-cut definitions and provide descriptive features for each
75 The terms SHORTLY-PRECEDE(t1,t2) and NEARLY-EQUAL(t1,t2) are mentioned by Yang & Bateman (2002: 1130-31). The predicate SHORTLY-PRECEDE(t1,t2) indicates that the time point t1 only shortly precedes the time point t2. The predicated NEARLY-EQUAL(t1,t2) indicates that the time point t1 lies close to the time point t2. Applying the terms to the explanation of achievements, t1 can be treated as Yang & Bateman’s ti, the event initial time, and as Smith’s (1991) I, initial points, and t2 can be regarded as Yang & Bateman’s tf, the event finishing time, and as Smith’s F, final points. Yang & Bateman’s expression of the predicates used to express temporal relations are altered from Allen 1984.
subclass of temporal adverbs. The descriptive features help to define temporal relations and to predict the compatibilities between temporal adverbs and aspect markers.
2 Literature Review and MethodologyTo make a clear-cut differentiation, we use the Academia Sinica Balanced Corpus (Sinica Corpus) and adopt the corpus-based approach to analyze Mandarin Chinese temporal adverbs. Time points in temporal axis will be used to define the temporal relations of the temporal adverbs in Lu & Ma 1999.
Smith (1991) discusses aspectual systems in language. She illustrates each situation type and viewpoint type with temporal schema. Below are the temporal schemata of Mandarin Chinese aspectual markers le, guo and zhe, which are represented by symbols I, F, F+1, …. and /.76
(1) Temporal schema fore the –le Perfective (Smith 1991: 348)77
Klein (1994) points out five temporal features and notes TT, TU and TSit. TT, topic time, is the time span to which the speaker’s claim on the occasion is confined. TU is time of utterance, which is the time at which the utterance is made. TSit is time of situation, which presents the time at which event occurs.
In addition to Smith’s and Klein’s temporal terminology of time points and time interval, the event modules in the framework MARVS (Module-Attribute Representation of Verbal Semantics) proposed by Huang et al (2000) can be also applied to analyze temporal relations of temporal adverbs. Event modules are the basic building blocks of the event contour. Five event modules stand alone or in combination, including Boundary, Punctuality, Process, State and Stage. The event module Boundary is defined as an event module that can be identified with a temporal point and must be regarded as a whole (including complete Event), which is adopted in this paper to define the notion of Boundary Point.
Yang & Bateman further discuss the semantic temporal relations of aspect system and propose principled semantic conditions for aspect combination. In their opinion, Chinese aspect system is actually composed of both aspect morphemes (-le, -zhe, -guo4, etc.) and aspect adverbials. Moreover, they propose that the Chinese aspect system has basically seventeen simple primary aspect forms. These simple primary aspect forms belong to the three subsystems of perfective, imperfective or future-existing according to the semantic properties in individual cases. Some simple primary aspect forms can combine to form an aspect of secondary type if their temporal attributes are in harmony. The temporal relation of
76 I and F indicate initial and final points. F+1 indicates a stage distinct from the final stage. The dots indicate internal stages and the slashes indicate the interval process.77 RVC is the abbreviation of Resultative Verb Complements.
137
the combination is represented graphically by time point ti, tf, tr and ts78.
In this paper, we adopt the terms proposed in the research above to help us clarify the temporal relation of each subclass of the temporal adverbs listed in Lu & Ma. We use the notations of ST, RT, ET, BP, Start and End to define temporal relations. Each respectively denotes the speaking time, the reference time, the event time, the boundary point, the start point of the event, and the end point of the event. For instances, the temporal features for le, guo and zhe are defined as follows in our system which are compatible with the notions of (Smith 1991).
le: BP≦ST, which means the prominent boundary point of the referred event precedes the speaking time.
guo: End<ST, which means the end point of the referred event precedes the speaking time.
zhe: ET=RT, which means the referred event time overlaps with the speaking time.
3 Temporal Relations and Compatibilities between Temporal Adverbs and Aspects
Temporal adverbs in Lu and Ma are divided into two classes, ST-related adverbs and RT-related adverbs, according to the time point of the situation referred to. The ST-related adverbs establish constraints between event time (ET) and speaking time (ST), but the RT-related adverbs establish constraints between ET and reference time (RT). For example, the ST-related adverb cengjing only denotes events or situations happened or existed before ST. Precisely speaking, the end points of events are prior to ST. For example, the sentence wo shangci cengjing jieshao guo 我上次曾經介紹過 ‘I have introduced it last time.’ is grammatical; however, *wo xiaci/mingtien cengjing jieshao guo 我 *下次/明天 曾經介紹過 ‘*I have introduced it next time/tomorrow.’ is ungrammatical. The RT-related adverb yijing 已經 is used to refer to events happened or existed before either a certain specific time, an event or ST. For instance, the sentences shang libaiwu wo yijing chuguo le 上禮拜五我已經出國了 ‘I have gone abroad last Friday.’ and xia libaiwu wo yijing chuguo le 下禮拜五我已經出國了 ‘I have been abroad next Friday.’ are both grammatical even though xia libaiwu is a future specific time.
Lu & Ma divide ST-related adverbs into three subclasses and RT-related adverbs into eighteen subclasses. Three subclasses of ST-related adverbs are subdivided into nine sub-subclasses. Eighteen subclasses of RT-related adverbs are subdivided into thirty sub-subclasses. By looking up Sinica Corpus, we find some RT-related adverbs in Lu & Ma’s classification behave like ST-related adverbs. Some temporal adverbs denote the order of events, which establish constraints only between ET. Furthermore, some temporal adverbs refer to habitual situations, which are not concerned with temporal relations. Therefore, our classification of temporal adverbs is somewhat different and revised from that of Lu & Ma’s. In this paper, temporal adverbs are divided into four classes according to temporal relations. More detailed sub-classification will be introduced in the sections below.
78 As the second footnote mentions, ti is the event initial time and tf, is the event finishing time. The symbols tr and ts is the reference time and the speaking time respectively.
138
3.1 Temporal Relations and Compatibilities between ST-related Adverbs and Aspects
In our classification, ST-related adverbs are classified into eleven subclasses. Each subclass is characterized and represented by its temporal features and is assigned one semantic type of Time, Duration, or Frequency. Table 1 shows the temporal features of each subclass.
Table 1 Temporal Relation and Semantic Role of Each Subclass of ST-related Adverbs
For instance, cengjing and ceng are grouped into the same subclass of ST-related adverbs. Since they denote the actions or situations that have been existed or happened before ST, the end points of the situations will precede ST. Thus the temporal relation of cengjing and ceng is defined as End<ST.
The temporal features clearly point out the relations among these ST-related adverbs. Furthermore, these temporal features can predict and verify correctness of co-occurrences of temporal adverbs and aspectual markers, le, guo and zhe, by unifying the temporal features of them. The temporal features of le, guo and zhe are BP≦ST, End<ST, and ET=RT respectively. If the corpus data shows the sentence is grammatical after combination, its combination of temporal relation will be in harmony; otherwise, there may be something wrong with the temporal relations or some other essential factors bringing out the incompatibility. Vice versa, we assume that by detecting the harmony of the combination of temporal relations between temporal adverbs and aspects, we can predicate whether the temporal adverb is compatible with le, guo and zhe or not. Here we take example by cengjing and ceng. The corpus data in Sinica Corpus shows the combination of cengjing with le, guo
139
and zhe as below:
(4) ceng jing co-occurs with le, guo and zhe
a. 我們的通信曾經給了她很大的快樂 women de tongxin cengjing gei ta henda de kuaile
She had been very happy while we lettered each other.
b. 我們曾經提過 women cengjing ti guo
we had mentioned before
c. 曾經提著小包袱 cengjing ti zhe xiao baofu
ever carried little backpack
The combination of cengjing with le, guo and zhe is grammatical so that we conclude the temporal relation of cengjing is correct and assign the semantic role of time, not frequency nor duration, to the subclass of cengjing and ceng. Vice versa, from the observation of the temporal relations of le, guo, zhe and cengjing, the temporal relation of cengjing is compatible with that of le, guo and zhe respectively. Thus before looking the corpus data up in Sinica Corpus, we predict the co-occurrence of cengjing with le, guo and zhe respectively is grammatical. The grammatical sentences 4a, 4b and 4c prove our hypothesis and prediction correct.
Temporal adverbs classified into the same subclass with the same temporal relation may not have the same syntactic behaviours. For instance, cong, conglai, zhijin, xianglai, sulai, lilai, su, and yixiang are grouped into the same subclass and assigned the semantic role of duration. The temporal relation of these temporal adverbs is defined as Start<ST. We find cong, conglai and zhijin are incompatible with le and zhe; however, xianglai, sulai, lilai, su and yixiang are incompatible with le and guo. Although this subclass of ST-related adverbs has the same temporal relation, their compatibility with aspectual markers is somewhat different. The Sinica Corpus sentences below reveal that cong, conglai and zhijin are usually in negative sentences. The feature [+NEG] brings about the different syntactic behaviours.
(5) 好像這件事從沒發生過hoaxing zhejian shicong mei fasheng guo
It seems this thing haven’t happened before.
(6) 從來不會逃避conglai buhui taobi
never evade
(7) 至今尚未出現在他生命中zhijin shangwei chuxian zai ta shengming zhong
haven’t appeared in his life
Temporal adverbs usually co-occur with aspects. The compatibility of temporal adverbs with aspect markers can testify the correctness of temporal relations. Therefore, not only
140
temporal feature representations of temporal adverbs and aspects are discussed but also the co-relations between temporal adverbs and aspects are our concerns. Below is the reversion of Table 1, including the compatibility of temporal adverbs with aspects. The symbol asterisk stands for ungrammaticality.
Semantic Role
Temporal Adverbs Temporal Relation
Compatibility with le, guo and zhe
time cengjing 曾經, ceng 曾 End<ST le guo zhe
time yeyi業已, yejing業經 End≦ST le guo zhe
time zhongyu終於, bijing畢竟, daodi到底
End≦ST le guo zhe
time xianxing先行 BP<ST le guo *zhe
time zaori 早日, jizao及早, chenzao趁早
ST<Start *le *guo *zhe
time zhongjiang終將, zhongjiu終久, zhonggui終歸, zonggu總歸, bijiang必將, chizao遲早, zaowan早晚
Table 2 Temporal Relation and Compatibility between ST-related Adverbs and Aspects
3.2 Temporal Relations and Compatibilities between RT-related Adverbs and Aspects
By adopting the corpus-based approach and applying the analysis of time points in temporal axis, Table 3 is the sub-classification of RT-related adverbs.
Semantic Temporal Adverbs Temporal Compatibility
141
Role Relation with le, guo and zhe
time yijing已經, yi已, zaoyi 早已, zaojiu 早就, dou都
BP<RT le guo zhe
time gang剛, ganggang剛剛, cai才 BP≦RT le guo zhe
time xian先, yuxian預先, shixian事先 BP≦RT le guo zhe
time jijiang即將, jiangyao將要, jiuyao就要, kuai快, xingjiang行將
Table 3 Temporal Relation and Compatibility between RT-related Adverbs and Aspects
The RT-related adverb yijing represents actions or situations have been finished, happened or existed before a certain specific time, another action or situation. The BP of the situation denoted by yijing must precede RT; therefore, its temporal relation is BP<RT. Sinica Corpus data point yijing is compatible with aspect markers.
(8) 一定已經打破了紀錄yiding yijing dapo le jilu
have certainly broken the record
(9) 已經吃過飯yijing chiguo fan
have had a meal
142
(10) 蛇已經咬著他了she yijing yao zhe ta le
A snake has bitten him.
We find the combination of temporal relations of some classes and that of aspect markers are contradictory to corpus data. The temporal adverbs classified in the subclasses with like, dunshi and turan(jian) present situations or events happen immediately after ST or another situations. The temporal relation of them is defined as RT≦Start. The corpus data show that these RT-related adverbs are incompatible with guo; however, their temporal relation RT≦Start is compatible with guo. Sentence 11 points out the relation of like between RT and ST:
(11) 他一聽到消息，立刻回了電話ta yi tingdao xiaoxi like hui le dianhua
He immediately called back when he got the news.
In sentence 11, RT is ta yi tingdao xiaoxi and ST is the present. The temporal relation of like in sentence 11 is defined as RT≦Start<ST, which is compatible with aspect markers le, guo and zhe. However, the corpus data show that guo is not compatible with like. The temporal relation RT≦Start focuses on start points of events while the aspect marker guo focuses on end points of events. One event can be viewed as whole or as different parts. When it is viewed as whole, the temporal relation is defined by BP. When it is viewed as different parts, its temporal relation is defined by either the start point or end point. Since Start is the focus of the temporal relation RT≦Start, the aspect marker guo pointing out End is incompatible with it. Thus the RT-related adverbs divided into the same subclasses with like, dunshi and turan(jian) are actually incompatible with guo even though their temporal relations seems compatible.
3.3 Temporal Relations and Compatibilities between Successive-related Adverbs and Aspects
In Lu & Ma’s classification, some temporal adverbs are neither ST-related adverbs nor RT-related. These temporal adverbs denote the order, sequence or successive of two situations. Thus, we classify these temporal adverbs into another class, Successive-related adverbs. Table 4 lists the Successive-related adverbs and their feature representations and co-relations with aspects.
Table 4 Temporal Relation and Compatibility between Successive-related Adverbs and Aspects
Since these temporal adverbs refer to the successive of events, their temporal relations have
143
no relation with ST and RT. They only concerns with the sequence of the event time.
3.4 Feature Representations and Compatibilities between other Temporal Adverbs and Aspects
Some temporal adverbs mark habitual situations or the manner of situations. Habitual situations represent that situations go on without focusing on any time points in time axis. Temporal adverbs denoting the manner of situations have no concern with temporal relations. Consequentially, these temporal adverbs are classified into another subclass in table 5.
Semantic Role
Temporal Adverbs Features Compatibility with le, guo and zhe
time manner jian漸, jianjian漸漸, jianci漸次, zhujian逐漸, rijian日漸, rijian日見, zhubu逐步
slowly
stage change
*experience
le *guo zhe
Time manner suishi隨時 preparatory *le *guo zhe
Time manner anshi按時, anqi按期 preparatory
regular
*le *guo *zhe
Table 5 Feature Representation and Compatibility between other Temporal Adverbs and Aspects
4 ConclusionFrom section 3.1 to 3.4, we classify the temporal adverbs listed in Lu & Ma into four main classes. The division of the subclasses in each main class is somewhat different from Lu & Ma’s although the majority is similar. Lu & Ma’s two-classes classification of temporal adverbs is vague so we adopt more detailed four-classes classification. The application of corpus-based approach helps us to testify the correctness of temporal relations. Besides, the co-relations of temporal relations between temporal adverbs and aspect markers le, guo and zhe can help to predict the co-occurrence of temporal adverbs with aspect markers. The feature representations of temporal relations are helpful to assign semantic roles to temporal adverbs in Sinica Treebank.
In future work, the subclasses of temporal adverbs may be subdivided into sub-subclasses. Temporal relations or other linguistic factors can be further applied to explain why certain classes of temporal adverbs can co-occur grammatically while certain classes cannot.
144
References Allen, J. F. 1984. Towards a General Theory of Action and Time. Artificial Intelligence. 1984, 23: 123-154.Huang, Chu-Ren, Kathleen Athens, Li-Li Chang, Keh-Jiann Chen, Mei-Chun Liu, and Mei-Chih Tsai. 2000. The
Module-Attribute Representation of Verbal Semantics: >From Semantics to Argument Structure. International Journal of Computational Linguistics and Chinese Language Processing, 5(1): 19-46. Also appeared in Proceedings of the Symposium on Selected NSC Projects in General Linguistics from 1998-2000, 119-46. 2001.
Klein, Wolfgang. 1994. Time in Language. London: Routledge.
Lu, Jianming & Zhen Ma (陸儉明&馬真). 1999. 〈關於時間副詞〉《現代漢語虛詞散論》, 98-127.Smith, Carlota S. 1991. The Parameter of Aspect. Dordrecht: Kluwer Academic Publishers.Yang, G-W & Bateman John A. 2002. The Chinese Aspect System and its Semantic Interpretation. Proceedings
of 2002 COLING II, 1128-34. Taipei: Academia Sinica.
Study on the Machine Tractable Thesaurus Dictionary of Contemporary Chinese Functional Words for Intormation Processing and Design Intormation Terms for Dictionary Entries
Chen QunxiuThe State Key Laboratory of Intelligent Technology and System
Department of Computer Science and Technology, Tsinghua University, Beijing 100084, ChinaE-mail: [email protected]
ABSTRACT: At present, worldwide scholars from different countries attach great importance to the knowledge resource construction of language information processing. Knowledge includes the aspects of lexicology, syntax, semantics, pragmatics and even general knowledge. The knowledge of semantics is the core, of which the most basic and most important one is the semantic knowledge of vocabulary. At last 10 years, scholars of Tsinghua University and Cooperators made systematic and complete studies of lexical semantic knowledge of main categories in Chinese content words (including verb, adjective, noun, etc), and developed and studied “The Machine Tractable Dictionary of Contemporary Chinese Predicate Verbs”, “The Machine Tractable Dictionary of Contemporary Chinese Predicate Adjectives”, “The System of Relation of Slots Centering on Nouns for Contemporary Chinese” and “The Thesaurus Dictionary of Contemporary Chinese for Information Processing”. But the study of lexical semantic knowledge (in particular for information processing) of Chinese functional words is up to now empty point. First of all, this paper discuses the value of functional words in Chinese Information Processing. Chinese in one of languages whose typology is isolate language and root word based analytical
＊本论文有关研究承 863 项目（项目号 2001AA114210）的资助。language. Vast majority of Chinese words can not themselves clearly express grammatical sense of them. Chinese syntactical methods mainly depend on functional words and words order. In addition it is important that Chinese functional words express meaning of sentences and structures of texts, e.g. voice, tense & acpect, mood, modality, positive and negative, relation of sentences, degree, range, draw into object, etc. Therefore, the study of lexical semantics of Chinese functional words is great importance to Chinese Information Processing. Secondly, this paper introduces the research of the Machine Tractable Thesaurus Dictionary of Contemporary Chinese Functional Words, in other words, it is one of expressing typology of studying category of Chinese functional words at Chinese semantic level in the 21th century. Namely it is The Modality Representation System of Chinese Sentences for Information Processing. Then, this paper shows the result of initial research. The Chinese functional words are divided into 10 kinds of main-level categories, e.g. Voice Category, Tense & Aspect Category, Mood Category, Modality Category, Positive and Negative Category, Relation Category of Sentences, Degree Category, Range Category, Object Category, Other Category, etc. Each of main-level category is divided into a number of middle-level categories or minor-level category once and again according to expressing different meaning. After that, this paper presents some ideas of designing information terms for dictionary entries. This paper finally proposes the conception of constructing a base platform of Contemporary Chinese Semantic Knowledge Bank.
Keywords: The Machine Tractable Thesaurus Dictionary of Contemporary Chinese Functional Words, The Modality Representation System of Chinese Sentences for Information Processing, Voice Category, Tense & Aspect Category, Mood Category, Modality Category, Positive and Negative Category, Relation Category of Sentences, Degree Category, Range Category, Object Category, Other Category, Dictionary Entries, Information Terms
1. Introduction Two-part allegorical expressions Xie-hou-yu, colloquial expressions whose second parts bear the main semantic weight, exhibit rich semantic and phonological operations. Generally, there are three kinds of two-part allegorical expressions: the first involving multiple-layer semantic derivations, the second only one layer of semantic derivation, and the third, both semantic and phonological derivations. This paper aims to deal with the first kind, which exhibits multiple semantic operations among metaphor, metonymy, and knowledge frame (cf. Lakoff 1993, Kovecses and Radden 1998, Fillmore 1977, Taylor 1995, among others). How the intended meanings of this type in TSM are obtained is the main concern of the study.
2. Cognitive mechanismsThis section will introduce the cognitive mechanisms before we probe into the analysis in Section 3. In Section 2.1, the notion of frame proposed by Fillmore (1977) and Taylor (1995) will be reviewed as the ground of the following cognitive mechanisms. Then, the concepts of metonymy and metaphor, as proposed by Kovecses and Radden (1998) and Lakoff (1993), will be provided in Section 2.2 and Section 2.3, respectively. And finally, Section 2.4 will discuss Ruiz de Mendoza’s (2003) interaction types between metaphor and metonymy in analyzing idioms.
2.1 FrameThe concept of frame is first pointed out by Charles Fillmore in the mid 1970s. By frame, we refer to the shared “knowledge network linking to multiple domains associated with a given linguistic form” (Taylor 1995). For example, the knowledge structure of [TREE] involves the domains of color (e.g. green, brown, white, etc.), parts (e.g. the root, the trunk, the leaves, etc.), shape (e.g. straight, curve, etc.), function (e.g. for building a house, for setting a fire, for cooking, etc.), height, and chemical components (e.g. chlorophyll, glucose, cellulose, etc.). The multiple domains are associated with the linguistic form [TREE]. In this study, Fillmore’s terminology of “frame” is adopted to refer to the shared knowledge network of human experience. Take the example of [BUY] frame in Fillmore’s (1977) to illustrate. A scenario can be structured in which a person acquires the ownership of some goods by giving another person who possesses the goods by agreement. The needed background involves the knowledge of BUYER, SELLER, GOODS, and MONEY. That is, the owner of the goods (the SELLER) has the control of the goods and the buyer cannot take the goods without paying an amount of money to the seller by the agreement; otherwise, he will commit a crime. Besides, in this society, money is considered as a valuable thing that people can exchange some goods by it. The following diagram summarizes Fillmore’s [BUY] frame.
The [BUY] frame not only shows the domains involved in a frame but also brings up the notion of perspectivization. In John bought a car yesterday, we describe the commercial event from the buyer’s perspective. That is, only the buyer and the goods are realized in the main structure of the sentence. The buyer is highlighted as the subject and the goods, as the object. 2.2 MetonymyAs mentioned in the previous section, perspectivization highlights one domain from the various ones in a frame to indicate the intended meaning. Such a process gradually leads to metonymy. Metonymy involves highlighting of an entity in a domain.
Traditional view of metonymy treats it as a matter of language in literary language where it is a transfer of the meanings of words which have reference and the stand-for relationship of the two words should be contiguously related (Stern 1931, Ullmann 1951, 1962, and Waldron 1967, cited in Kovecses and Radden 1998). However, as more studies are devoted to this field, metonymy is not what was thought before. In cognitive linguistic view, metonymy is conceptual in nature, involving not only literary language but also our daily life expressions (Kovecses and Radden 1998). It is a vehicle or access that enables us to process from one conceptual entity to another in the same domain, or idealized cognitive models (ICM) (following Kovecses and Radden 1998). For example, in The fishburger asks for another cup of coke cola, we process from the food to the person who ordered the food in the fishburger domain. There are two general configurations: (1) whole ICM and its parts; (2) parts of an ICM. Under each configuration, there are some more specific metonymic types. These are shown in the following (Kovecses and Radden 1998: 49-58).
(1) Whole ICM and its part(s) (2) Parts of an ICMA. Thing-and-part ICM A. Action ICMB. Scale ICM B. Perception ICM C. Constitution ICM C. Causation ICMD. Complex event ICM D. Production ICME. Category-and-member ICM E. Control ICMF. Category-and-property ICM F. Possession ICM
G. Containment ICM Meaning extensions of metonymy enable us to process from one conceptual entity to another in the same domain. However, meaning extensions can involve meaning transfers across two domains. Section 2.3 will introduce the mechanism behind the conceptual mapping of a concept from a concrete domain to an abstract one—metaphor. 2.3 Metaphor Metaphor is traditionally defined as “a novel or poetic linguistic expression where one or more words for a concept are used outside of their normal conventional meaning to express a ‘similar’ concept” (discussion on classical theory in Lakoff 1993: 202). This view treats metaphor as a “language,” or a “figure of speech” where it stems from the similarity of two concepts. However, as a great number of metaphors being discovered to have generalizations of our thought, contemporary theory on metaphors holds that metaphor is conceptual in nature. Metaphor is a conceptual mechanism through which we comprehend abstract concepts in terms of more concrete concepts. That is,
156
metaphor is “a cross-domain mapping in the conceptual system” (Lakoff 1993). We map the concrete entities in the source domain to abstract entities in the target domain. For example, the metaphor LOVE IS A JOURNEY is a mapping from the source domain (a journey) to the target domain (love) (Lakoff’s 1993). Defined by Lakoff (1993), metaphors follow the Invariance Principle: “metaphorical mappings which preserve the cognitive topology (that is, image-schema structure) of the source domain, in a way consistent with the inherent structure of the target domain.” That is, in the example LOVE IS A JOURNEY, the source domain entities is exactly mapped to the corresponding entities in the target domain, not others. Hence, lovers correspond to travelers; relationship corresponds to vehicles; goals correspond to destinations; and difficulties correspond to impediments on the road. In addition to conceptual metaphors, novel metaphors also obey the Invariance Principle. Novel metaphors include three types: image metaphors, generic-level metaphors, and extensions of conventional metaphors (Lakoff 1993). Unlike conceptual metaphors whose mappings are correspondence between entities in the two domains, namely, source domain and target domain, image metaphors map between characteristics in two conventional mental images (Lakoff 1993). The metaphoric mappings of image metaphors work in the same way as those of conceptual metaphors. Mapping the structure of the source concept to the structure of target concept just like conceptual metaphors, image metaphors are “one-shot” metaphors, which map one source image onto another target image. The following example from Andre Breton’s line adopted by Lakoff (1993: 229) My wife… whose waist is an hourglass can well illustrate the mappings of image metaphor. After reading the line, we immediately have the image that the woman’s waist is very thin since we automatically map the image of an hourglass onto the image of a woman’s owing to their similar shapes. This operation is conceptual. The words of the line do not tell us the thinness of the woman’s waist nor do they tell us which part of the hourglass is mapped to the woman’s waist. It is our conventional knowledge about the special part of the hourglass and our mental images about an hourglass and the woman’s waist that enhance us the connection. Besides characteristics of an image, the evaluation of one image can also map or percolate to another image as in the example quoted by Lakoff My horse with a mane made of short rainbow (Lakoff 1993: 230). The curved bands of the rainbow mapped onto the curly hair of the horse. Such an image mapping also maps the evaluation of the rainbow onto the horse. Hence, we maintain that the horse is beautiful by our evaluation on the rainbow. Provided with the knowledge of metonymy and metaphor, we can proceed to the question how metaphor and metonymy interacts. Ruiz de Mendoza (2003) provides us three interaction types.
2.4 The interaction of metaphor and metonymy Ruiz de Mendoza (2003) provides us with three types of interaction between metaphor and metonymy by the distinction between part-for-whole and whole-for-part metonymies. The first type is source-in-target metonymy within the metaphoric source where the output of metonymy serves as the source of a metaphor as in example 1. The second type is also source-in-target metonymy, but different from the first type, this type exhibits a different mapping where the output of a metaphor provides the source for the metonymic mapping as in example 2. The third type is a variation of the second type where a metonymic mapping highlights the specific corresponding part of a metaphor as illustrated in the example 3.
Eg 1: He got up on his hind legs to defend his views.
157
In this example, the metonymy serves as the source of a metaphor where the animal’s behavior maps to human behavior through a metaphor. First, we image a scene in which an animal rears up in an attacking position. The hind legs stand for the overall scene. Such is a part for whole metonymy where the source is within the target. Then, the scene metaphorically maps to a scene in which a person energetically stands up to argue in public. Finally, the energetic action of a person standing up to argue in public is metaphorically mapped to a behavior where a person defends his opinions with vigor whether he stands up or not.
Eg 2: He kept his eyes peeled for pickpockets.
Peeling his eyes metaphorically refers to “open one’s eyes widely.” The target of a metaphor is the source of a metonymy where the action of opening one’s eyes stands for the overall situation where one keeps his eyes open to be alert to possible dangers.
Eg 3: She could read my minds.
158
In this example, the mind is like a text which could be read. There is a system of metaphorical mappings, including the reader, the action of reading, and the text. The function of the metonymy is to point out which part of the metaphorical mapping will be interpreted. In this example, the mind is the part we focus to interpret. After provided with conceptual tools for analysis, we will advance on the analysis by using these conceptual mechanisms in the following section.
III. Analysis Ruiz de Mendoza’s (2003) notion of interaction types in analyzing idioms is modified to account for the data, among which four interaction types are found:
Eg 4: Bi-hun tang — pah-pah khang (米粉筒—百百孔).
Example 4 illustrates the first type, where the source of the metaphor contains target-in-source metonymy. In the first stage, the knowledge frame of [THE TUBE FOR MAKING RICE NOODLE] is activated, including several domains such as domain of its shape, domain of its weight, domain of its function, domain of the steps for making rice noodles, and so on. Then, subsequently, the domain of shape is perspectivized. Specifically, its characteristic shape—having hundreds of holes on the bottom—is perspectivized to stand for the whole object, activating DEFINGING PROPERTY FOR CATEGORY metonym where the target is within the source. Then, the characteristic of the tube, serving as the source of a metaphor, is metaphorically mapped to the characteristic of the result of an event. Hence, hundreds of holes in the tube represent full of holes/problems of the result. Through this process, the intended meaning ‘full of holes/problems’ is derived. The metaphor involved here is a type of Lakoff’s (1993) novel metaphor—image metaphor—where the image of the tube for making rice noodles is mapped to the image of the result of an event. The Invariance Principle operates also in the metaphor. Holes of the tube correspond to holes of the result. The meaning derivation of the first type is shown above.
Eg 5: Kau-bu pang-jio—nng-kha (狗母放尿—軟腳)
In the second type (eg 5), target-in-source metonymy is embedded in the metaphoric source and the target of the metaphor serves as the source of a metonymy which itself is the source of the next source-in-target
159
metonymy. In the first stage, the domain of physical posture is perspectivized from the knowledge frame of the event [A BITCH URINATES], which contains domain of physical posture, domain of the urinating steps, domain of the urinating places, and so on. The characteristic posture of a bitch’s urination—bending hind legs—is perspectivized to stand for the whole events, activating DEFINGING PROPERTY FOR CATEGORY metonym in which the target is embedded in the source. Then, the posture of a bitch’s bending legs is metaphorically mapped to the posture of a frightened person’s bending legs. Such a mapping illustrates an image metaphor in which the image of a bitch’s posture is mapped to that of a person’s posture. The Invariance Principle operates here, where a bitch’s bending legs corresponds to a person’s bending legs. Then, the knowledge frame provides us with the knowledge that someone will unconsciously bend his legs when he is over-frightened. Such an action is out of fear. Bending legs stands for fear, activating PHYSIOLOGICAL EFFECT FOR EMOTION metonym, which is a part-for-part relationship. Then, the knowledge that being afraid of everything is one of the main characteristics of being a coward lays the ground for the next meaning derivation. The defining feature—fear is employed to stand for the category—cowardice, activating DEFINING PROPERTY FOR CATEGORY metonym. Hence, through these metonymic and metaphorical mappings, the intended meaning—cowardice is obtained. The derogative meaning of the two-part allegorical saying is preserved in the mapping. Both bitches and urination are evaluated as negative in our culture. Such negative evaluation of the surface form is percolated to the intended meaning of the two-part allegorical saying. Hence, speakers could easily obtain a derogative meaning in the two-part allegorical expressions.
HowNet-Based Extraction of Chinese Message StructuresQiang Dong Changling Hao Zhendong Dong
Research Center of Computer & Language Information Engineering, CAS, Beijing, 100083E-mail: [email protected]
Abstract: The Chinese message structure is composed of several Chinese fragments which may be characters words or phrases. Every message structure carries certain information. We have developed a HowNet-based extractor that can extract Chinese message structures from a real text and serves as an interactive tool for building large-scale bank of Chinese message structures. The system utilizes the HowNet Knowledge System as its basic resources. It is an integrated system of rule-based analyzer, statistics based on the examples and the analogy given by HowNet-based concept similarity calculator.Keyword: Chinese message structure; Knowledge Database Mark-up Language (KDML); parsing; chunk;
算语言学联合学术会议论文集），孙茂松、陈群秀主编，清华大学出版社［2］ 颜国伟、谭慧敏，基于《知网》的常识知识标注，中文计算语言学期刊，第 4卷第 2期，1999［3］ 汉语计量与计算研究，邹家彦，香港城市大学语言资讯科学研究中心，1998［4］ Jorn Veenstra, Memory-based text chunking, In Nikos Fakotakis Machine Learning in human language
technology, workshop at ACAL99, 1999［5］ Jian Sun, Ming Zhou, Jianfeng Gao, “A Class-based Language Model Approach to Chinese Named Entity Identification”, Computational Linguistics and Chinese Language Processing,Vol.8, No.2, August 2003, pp.1-28［6］ Hua-Ping ZHANG, Qun LIU, Hong-Kui YU, Xue-Qi CHENG, Shuo BAI, “Chinese Named Entity Recognition Using Role Model”, Computational Linguistics and Chinese Language Processing, Vol.8, No.2, August 2003, pp.29-60网络资源：http://www.keenage.com
1 引言词与词在语义上存在一定的意义关系,词汇分解是说明词与词之间意义关系的一种办法。Dowty 使用词汇分解方法，把生成语义学家的研究成果在 Montague 语义学框架里做出了描写。词汇的语义通过深层结构对句子意义的表达，或使用一些抽象算子（如CAUSE, BECOME 等）和逻辑语言的基本词表示为逻辑公式，进而发现词汇语义的关系（同义，蕴涵等）。例如，给出及物动词 kill、不及物动词 die和形容词 dead 的语义表示如下：a. kill’ = lxly[CAUSE(BECOME(alive’(y)))(x)]b. die’ = ly[BECOME(alive’(y))]其中 alive’是基本逻辑词，加撇号“’”的词汇表示与该词汇对应的逻辑公式语言中的语义项。那么容易得出：Zhang San killed Li Si.蕴涵 Li Si died.
意义公设也是一种描写词汇之间意义关系的办法，比如下位关系的描述。X 是 Y 的下位词，当且仅当 X和 Y 有以下意义公设：�x[X’(x)Y’(x)]。作为示例，有a. �x[猫’(x) 哺乳动物’(x)]b. �x[哺乳动物’(x) 动物’(x)]。
在句子的语义解释中，语义关系的作用是决定性的。词汇分解方法给词汇作了意义表达，意义公设方法给词汇之间的某些意义关系作了描述。然而，短语、句子之间的意义关系是如何系统地得到呢？例如，Zhang San killed Li Si.蕴涵 Li Si died.能不能依照某种机制推出。词汇语义关系可以按一定的规则在词汇形成短语或句子的过程中传播，因而找出传播规则就能够方便地得到短语或句子之间的语义关系，这个思路很自然地与逻辑证明的推理方法联系起来。
Hownet, Wordnet 都是词汇语义关系的例子，所以这里做的研究也可看作是对词汇的上下位关系的组合传播方式的理论探讨，进一步的研究可以考虑对其他的词汇语义关系的传播作研究, 或研究任何词汇语义关系向上下位关系的转换。参考文献 Dowty, D. Word Meaning and Montague Grammar. 1979. Dordrecht, Reidel.Montague, R. Formal Philosophy. Selected Papers of Richard Montague, edited and with an introduction by R.
Thomason. New Haven, Conn. 1974. Yale University Press.Moortgat, M. Multimodal linguistic inference. 1995, Bulletin of the IGPL 3(2,3). Special issue on Deduction
2.4 并列成分中心语语义既不相似也不相关也不相对语言中存在少数的并列结构，其并列成分中心语的语义既不相似也不相关也不相对。例如：（5） a 草案确定了反恐怖活动战略计划的执行[机构和办法]。
b 领导干部受到[人民和法律]的监督。对此我们还无法进行有效的描述和解释。似乎是，只要两个语言成分符合句法约束就可形成并列，而不需要“额外”的语义约束。储泽祥等（2002）从“语用需要、经济原则”的角度描述两个名词的非常规联合，但没有涉及句法语义的解释。3 从并列成分中心语语义相似看 CCD 的名词语义分类体系把名词性并列结构并列成分中心语语义相似看作是一种客观存在的语言特性，那么并列结构为我们提供了一个很好的视角来审视 CCD的语义分类体系，这种审视对其他的语义分类体系也很具参考价值。3.1 “人们”、“社会团体”语义类名词和“人”语义类名词可形成并列——考虑移动并入CCD在“群体（group）”语义类下设有“人们（people）”一小类（记作[＋人们]），表示“任何一群人（any group of human beings）”，而 25 个基本语义类中又设有“人（person）”一类。现代汉语中，[＋人们]和[＋人]名词经常形成并列，例如：（6） a 日益被[各层领导和社会公众]所认识。 （领导[＋人]，公众[＋人们]）
b 雅俗共赏，极受[专家和人民群众]喜爱。 （专家[＋人]，群众[＋人们]） [＋人们]、[＋人]名词能自由形成并列，而可以不论是否是“群体”，即数的多少，这是由于汉语没有数的形态变化而造成的。b中并列结构若翻译成英语必得是“experts and common people”。“人们”、“人”这两个语义类在汉语中是相近的，应该合并在一起。事实上，董振东先生的HowNet、北京大学的《语义词典》（王惠等，2003）都是将“公众”、“群众”这样的词置于“人”语义类下。因此，在 CCD中可以将“人们”小类从“群体”类中移出并入“人”语义类。
b [旅客和航空公司]都受到损失， （旅客[＋人]，公司[＋社会团体]）[＋人]名词和[＋社会团体]名词都具有“施事”功能，很多动词对它们具有相同的选择限制（selectional restriction）。例如，“反映”的主体可以是“求职者”，也可以是“单位”，因此可形成 a的并列；“受到损失”的主体可以是“旅客”，也可以是“公司”，因此可形成 b 的并列。除了人所特有的一些生理动作[＋社会团体]名词不能胜任，例如不能说“**单位吃”，“**公司跑”，[＋社会团体]名词可以充当大多数动词的施动者，如“单位赠送锦旗”，“公司转让债权”，“航空公司请求延期”等等。虽然表面上[＋社会团体]名词不具有生命，但它由具有生命的人所组成，并且由其中的代表法人来实施某种行为，因此句法功能上[＋社会团体]名词和[＋人]名词有很多相似之处86。《语义词典》将“社会团体”置于“人”语义类下作为一个次类87，这是比较合适的。由此，在 CCD中可以将“社会团体”小类从“群体”类中移出并入“人”语义类。同为“群体”语义类的[＋人们]名词和[＋社会团体]名词可形成并列。例如：（8） a 要求[各国政府和全人类]采取紧急行动， （政府[＋社会团体]，人类[＋人们]）
b [地方政府和人民群众]积极支持部队， （政府[＋社会团体]，群众[＋人们]）因此，从并列结构形成的角度考虑，“人们”、“社会团体”可适当并入“人”语义类。3.2 “社会团体”语义类名词和“行政区”语义类名词可形成并列——考虑移动靠近“群体”语义类下的“社会团体”名词和“处所位置”语义类下的“行政区（district）”名词（记作[＋行政区]）可以形成并列。[＋社会团体]名词既可以指称共享某些社会关系的人们，也可以表示这些人们所在的处所位置，例如“银行”，在（9）a中表示“银行的领导者或人们”，而在（9）b中表示“处所位置”之义。反之，[＋行政区]名词既可以表示占有一定空间的处所位置，又可以指称相关的社会团体，例如“北京”，在（10）a 中表示“处所位置”之义，而在（10）b中表示“北京的领导者或人们”88。表现在并列结构上，[＋社会团体]名词和[＋行政区]名词可以自由形成并列，如（11）中的例子。
（9） a 这 13家银行也作出了积极反应。 b 他走进了那家银行。（10）a 1921年 9月 26日生于北京。
b 北京按照国际奥委会的要求，如期将申办报告文本送交国际奥委会审阅。（11）[俄罗斯和北约]建立战略伙伴关系。（俄罗斯[＋行政区]，北约[＋社会团体]） 由此可见，“社会团体”和“行政区”这两个语义类在汉语中是相近的，语义类设置
形成并列，但[＋人]名词却很少能够和[＋行政区]名词形成并列。可见，词语之间的并列关系是不可传递的。86 从WordNet 对“社会团体”的解释中，可以更清楚地看到“社会团体”和“人”的密切联系：“people sharing some social relation”，“社会团体 ”其实就是一群人。87《语义词典》将此语义类记作“团体（group）”。88 从《语义词典》的一例填写错误中我们也可以窥见“社会团体”和“行政区”两个语义类的密切联系：某同学将“俄罗斯”的语义类填写成“机构”（相当于 CCD中的“社会团体”），而将“俄国”的语义类填写成“处所”，可见人在判断这两个语义类时有时也是模糊不清的。
b 可持续发展的[定义和内容]。 （定义[＋通信]，内容[＋认知]）c 被[习惯势力和陈旧观念]所束缚。 （势力[＋属性]，观念[＋认知]）d 有着[悠久的文明和丰富的文献传统]。 （文明[＋群体]，传统[＋认知]）e 一国两制事业的[可行性和辉煌前景]。 （可行性[＋属性]，前景[＋状态]）
人们对抽象事物的认识其实并没有那么清晰的分类意识。假如问一个人“环境”、“传统”、“可行性”的语义类分别是什么，他或许会回答“环境”的语义类是“认知”，“传统”的语义类是“通信”，“可行性”的语义类是“状态”。各家语义分类体系对抽象词语的归类也存在诸多的不一致性。《语义词典》将“环境”和“传统”笼统地归入“抽象事物”，将“可行性”归入“属性”。HowNet 将“环境”和“可行性”归入“属性”，将“传统”归入“规矩”（相当于 CCD中的“认知”类）。而对具体事物（实体）就是另一番光景了，没有人会认为“桌子”是“食品”，或者“狗”是“人工物”。表现在语言上，具体事物（实体）的语义类之间不能随意并列，偶尔并列也依赖于一定语境的支撑。现代汉语并列结构的形成启示我们，对路径长度相同的两对节点，“具体事物”类下的两个节点语义距离较大，而“抽象事物”类下的两个节点语义距离较小。由此可知，一方面在并列结构的自动识别过程中，对抽象名词的语义相似性要求可适当放宽，而对具体名词的语义相似性要求需适当加严；另一方面在语义类设置过程中，对抽象类名词的分类宜粗而不宜过细。4 从并列成分中心语语义相对看 CCD 的名词语义分类体系4.1 高度抽象的词语容易形成并列——宜从一个新的角度进行语义归类上文 2.3谈到，表示相对意义的词语经常形成并列。例如：（13） a [社会心理、人们情感]变化是值得抒写的。（心理[＋认知]，情感[＋情感]）
b 赢得了宝贵的[时间和空间]。 （时间[＋时间]，空间[＋处所位置]）需要注意的是，（13）中各并列成分多是概括的、抽象的词语，它们不指称具体的概念，一旦将其中一个抽象概念换作具体概念，并列就不能成立。这种抽象概念往往就是某个语义类的“标签”，其本身的意义和用法与语义类内部具体词语差别很大。例如“时间”的语义类是[＋时间]，但和具体的时间词“今天”、“明年”用法语义差别很大，可以说“今天学习”，但不能说“**时间学习”。“时间”能和别的语义类的词语形成并列，如“时间和精力”，“时间和空间”，但具体的时间词却只能跟时间词自己并列，不能跟别的语义类的词语形成并列，“**今天和精力”，“**明年和空间”这样的并列在语言中是不存在的。又例如“情感”的语义类是[＋情感]，但它和具体表示情感的词语（例如“温情”、“恐慌”）在用法语义上差别很大，“情感”和“温情”、“恐慌”在形成并列结构时鲜有共同点。像语义类标签的这些高度抽象的词语可看作是广义的属性（attribute），语义类之下的具体实例 （instances）可看作是属性值（attribute values），我们有VALUE(时间) = 今天|明年，VALUE(情感)=温情|恐慌。我们怀疑，类似语义类标签的这些高度抽象的词语，是否应该从一个新的角度进行语义归类。
183
5 结语本文一方面基于 CCD的语义分类体系，考察了现代汉语名词性并列结构并列成分中心语的语义关系，包含有四种：语义相似、语义相关、语义相对、语义既不相似也不相关也不相对。语义相似是并列成分之间最主要的语义关系，但并不是所有的并列成分都呈现出语义相似的特性。另一方面，透过并列成分之间的语义关系其中主要是语义相似的关系，我们重新审视了 CCD的语义分类体系，对词语之间的语义关系进行了深入的思考，为语义类的设置提供了一些有价值的参照座标。本文对并列成分中心语的语义关系和 CCD的名词语义分类体系进行了双向考察和验证，这其实也就是形式和意义的相互验证。参考文献Agarwal, R. and Boggess, L. 1992. A simple but useful approach to conjunct identification. In Proceedings of
30th Annual Meeting of Association for Computational Linguistics.Fillmore, C. J. 1982. Frame Semantics. Linguistics in the Morning Calm. Seoul: Hanshin Publishing
Corporation.中译文，詹卫东译，框架语义学.《语言学论从》第 27 辑.Kurohashi, S. and Nagao, M. 1994. A syntactic analysis method of long Japanese sentences based on the
detection of conjunctive structures. Computational Linguistics, 20(4).Miller, A. 1999. Nouns in WordNet. In Fellbaum, C., (ed.), Wordnet: An Electronic Lexical Database.
Cambridge: MIT Press.Okumura, A. and Muraki, K. 1994. Symmetric pattern matching analysis for English coordinate structures. In
Proceedings of the 4th Conference on Applied Natural Language Processing.Resnik, P. 1993. Selection and Information: A Class-Based Approach to Lexical Relationships. Ph.D.
摘 要歧义问题是语言学中的一个难点，也是自然语言理解和机器翻译中难以解决的问题。无标记并列结构——n+n、v+v、a+a是几种例化后可能产生歧义的结构，本文在潜在歧义理论的基础上，重点分析每种结构中两个词的语法、语义、语用特性，对汉语中的这几种结构做了比较全面的研究，目的在于找到这些结构中产生歧义的原因和消解歧义的策略。希望本文能为自然语言计算机处理中歧义的消解提供一定的参考作用。Ambiguity Study of Unmarked Coordination in Contemporary Chinese For Chinese Information
Processing
Yang Quan Feng Zhiwei
Abstract
Ambiguity is the difficult problem in linguistic. It is one of the keys in nature language understanding and machine translation. v+v, n+n and a+a are some structures which may have ambiguity after instantiation. According to Potentional Ambiguity Theory, this thesis analyzes these phrases by integrating syntactic, semantic and pragmatical processing. It makes a comprehensive study of unmarked coordination in contemporary Chinese, aiming at finding out the reason why ambiguity exists and searching out some disambiguity rules. We hope it can help the disambguity of computer processing nature language.
参考文献：［1］朱德熙，1980，汉语句法中的歧义现象，现代汉语语法研究，商务印书馆。［2］吕叔湘，1984年（5），歧义类例，中国语文。［3］冯志伟，1989年（2），中文科技术语的结构描述和潜在歧义，中文信息学报。［4］冯志伟，1995年（4），论歧义结构的潜在性，中文信息学报。［5］冯志伟，1996，自然语言的计算机处理，上海外语教育出版社。［6］詹卫东，常宝宝，俞士汶，1999年（3），汉语短语结构定界歧义类型分析及分布统计，中文信息学报。［7］詹卫东，2000，面向中文信息处理的现代汉语短语结构规则研究，清华大学出版社出版。［8］尤庆学，2002，现代汉语歧义结构研究，武汉大学博士学位论文。［9］Hirst, G. 1986.Semantic Interpretation and the Resolution of Ambiguity. Cambridge University Press, Cambridge. ［10］James Allen, 1995.Natural Language Understanding, The Benjamin/ Cummings Publishing Company, Inc. ［11］Donald Hindle, Mats Rooth,1993. Structural Ambiguity and Lexical Relations.
其中，ID号是概念节点的唯一编码，Tag 是该节点的名称，通常取同义词集合中熟悉度最大的一个词，大括号内是该节点的语义定义，通过同义词集合以及该节点和其他节点的关系来描述。我们认为，这种结构最大的优点是描述简洁，便于计算机处理。参考文献[1] 黄昌宁, 李涓子.2000.词义排歧的一种语言模型.北京：语言文字应用，2000，（8）[2] 董振东.1998.语义关系的表达和知识系统的建造.北京：语言文字应用，1993，（3）[3] 陈群秀.1998.一个在线义类网络.北京：语言文字应用，1998，（2）[4] HSIN-HSI CHEN, CHI-CHING LIN, AND WEN-CHENG LIN. National Taiwan University. Building a Chinese-English Wordnet for Translingual Applications. ACM Transactions on Asian Language Information Processing, Vol. 1, No. 2, June 2002.
[5] 刘群、李素建， 基于《知网》的词汇语义相似度计算， 第三届汉语词汇语义研讨会（台北，2002）[6] 张晶、姚建民、赵铁军、李生.2001.基于wordnet和hownet建设双语语义词典，高技术通讯（2001.12）[7] VOSSEN, P. 1998. EuroWordnet: Building a multilingual database with Wordnets for European languages. ELRA Newsl. 3, 1, 7-10.[8] HSIN-HSI CHEN, CHI-CHING LIN, AND WEN-CHENG LIN. 2000.Construction of a Chinese-English Wordnet and its application to CLIR. In Proceedings of the Fifth International Workshop on Information Retrieval with Asian Languages (Hong Kong, Sept.-Oct. 2000). ACM Press, New York, NY, 189-196.
摘要无导词义消歧避免了人工词义标注的巨大工作量，可以适应大规模的多义词消歧工作，具有广阔的应用前景。这篇文章提出了一种无导词义消歧的方法，该方法以 hownet的词库为词典，采用二阶 context构造上下文向量，使用 k-means 算法进行聚类，最后通过计算相似度来进行词义的排歧.实验是在抽取术语的基础上进行的，在 13 个汉语高频多义词的测试中取得了平均准确率 81.33％的较好的效果。An Unsupervised Approach To Word Sense Disambiguation Based On Hownet
ChenHaoDepartment of Computer ScienceCentral China Normal University 430079Wuhan, China
AbstractAn unsupervised WSD(word sense disambiguation) can avoid big labor cost and it is possible to adjust to deal with large-scale ,so WSD has extensive applications in many fields. This paper presents an unsupervised approach which is used as dictionary based on hownet , constructing context vector by means of second-order context, clustering by k-means and disambiguates by calculating the similarity. Our experiments are based on the extraction of term and average accuracy is 81.33% for 13 ambiguous words in open test by this method.
References:[1] 鲁松,白硕,黄雄,张健 基于向量空间模型的有导词义消歧 计算机研究与发展 Vol.38,No.6June 2001 662－667（LuSong, BaiShuo, HuangXiong,andZhangJian Supervised Word Sense Disambiguation Based On Vector Space Model Journal Of Computer Research & Development Vol.38,No.6June 2001 662－667 ）[2] 鲁松, 白硕,黄雄 基于向量空间模型中义项词语的无导词义消歧 软件学报Vol.13, No.6 1082－1089(LU Song, BAI Shuo, HUANG Xiong An Unsuptervised Approach to Word Sense Disambiguation Based on Sense-Words in Vector Space Model Journal of Software Vol.13, No.6 1082－1089)[3] Li, Juan-zi. The research on Chinese word sense disambiguation [Ph.D. Thesis]. Beijing: Tsinghua University, 1999 (in Chinese).[4] 董振东，董强（2000），“知网”，http://keenage.com (Dong Zheng-Dong,DongQiang(2000),”hownet” http://keenage.com)[5] 行小帅,潘进,焦李成 基于免疫规划的Ｋ-means聚类算法 计算机学报 2003 vol.26 No.5 605－610(Xing Xiao-Shuai, PanJin,Jiao Li-Cheng ＡNovel K-means Clustering Based on the Immune Programming Algorithm Chinese Journal Of Computers 2003 vol.26 No.5 605－610)[6] Patrick Pantel & Dekang Lin. 2001. A Statistical Corpus-Based Term Extractor. Canadian Conference.on.AI2001.p.36-46[7] 李飞,薛彬,黄亚楼.初始中心优化的 K-Means聚类算法.计算机科学,2002,29(7):94～96(LiFei, XueBing, HuangYaLou A Novel K-means Clustering Based On Initial Central Superior Computer Science 2002,29(7):94～96)[8]Miller, G.A., Charles, W. Contextual Correlates of Semantic Similarity. Language and Cognitive Processes, 1991,6(1):1~28.
Abstract This paper discusses the semantic and syntactic analysis of four very common transitive verbs in newspaper reportage: shuo1 (說/say), biao3shi4 (表示/express), zhi3chu1 (指出/point out) and cheng1 (稱 /claim). These four verbs have high frequencies in various Chinese-speaking communities, including Hong Kong, Taiwan and Beijing. On the basis of authentic linguistic data from the LIVAC (Linguistic Variation in Chinese Speech Communities) synchronous corpus, we compare the constructions and usages of these four verbs in the three regions. Our analysis unveiled the subtle difference in usage among these four near-synonyms, and provided useful data on verb argument structures for benchmarking the competence of humans in the verb acquisition process. This contrastive analysis will also aid the construction of Chinese semantic networks and thesauri which are applicable in many Chinese-speaking communities.
1 IntroductionStudies on verbal synonymy have seldom addressed the variations between the standard written form of a language and local dialects. In this study, we analyze four closely-related common verbs used in news reportage: shuo1 (說 /say), biao3shi4 (表示 /express), zhi3chu1 (指出 /point out) and cheng1 (稱 /claim). According to the LIVAC98 synchronous corpus, these four verbs are amongst the most frequently used transitive verbs in Hong Kong news texts in the past few years. Similar dominance99 is also observed in Taiwan news and magazine texts (CKIP, 1993). Interestingly, all of them can correspond to the very general verb in reported speech, waa6 (話/say), in colloquial Cantonese.
In HowNet (Dong, 1999), these four verbs are arranged into two sememes: biao3shi4 and zhi3chu1 belong to the sememe “express|表示”, and shuo1 and cheng1 belong to “speak|說”. These two sememes have hypernym-hyponym relation, such that “express|表示” is the parent of “speak|說”. These four verbs have very similar syntactic and semantic valencies that are hard to be recognized solely by intuition. Therefore corpus data is needed to reveal some important but opaque distributional differences among them.
Both syntactic and semantic information are crucial to the study of synonymy. It is assumed that “the syntactic realization of arguments – their syntactic type and grammatical function – is predictable to a large extent from the meaning of a word” (Liu cf. Levin and Rappoport, 2003). Roland and Jurafsky (1998) compared the subcategorization frequencies between different types of corpora 100 and classified the subcategorization frames into sixteen categories. Arppe and Järvikivi (2002) uncovered the preferences of the usage of near-synonymous verbs in Finnish, and realized that syntactic and semantic analysis can be used to enhance and enrich lexical description.
Similar approaches are adopted in the study of synonyms in Chinese. Chief et al. (1998) differentiated two Chinese adjectives through corpus-aided studies. Syntactic difference between near-synonyms is used to indicate the existence of subtle semantic difference. Similarly, Tsai et al. (1998)
98 Linguistic Variation in Chinese Speech Communities. http://www.livac.org/99 biao3shi4, shuo1, zhi3chu1 and cheng1 ranked 11, 16, 23 and 298 respectively among 19,907 verbs.100 Corpora analyzed: psychological sentence production data, written text, and telephone conversation data.
220
used the distributional differences in the syntactic patterns of near-synonymous Chinese predicates to deduce the relevant components of verb meanings. Four different types of syntactic information were also proposed for the guideline of similar research: syntactic function, argument structure, aspectual type and sentential type. Recently in 2003, Liu distinguished between the Chinese synonymous verbs tao3lun4 (討論 /discuss) and shang1liang2 (商量 /discuss), and suggested that a huge difference in frequency of use readily indicates that the verb with lower frequency might be semantically more restricted and marked, and hence more restricted in use. On cross-language verbal synonymy, Gao and Cheng (2003) investigated “verbs of contact by impact” in English and their equivalents in Mandarin Chinese, and showed the dissimilarity of semantic components between English and Mandarin Chinese.
However, there is not yet any significant study of verbal synonymy between formal literate and literary dialect, or between standard and local dialect. Chinese dialects do not readily enjoy mutual intelligibility. Although Modern Standard Chinese (MSC) is regarded as the official language of China, dialect speakers do write in a dialectal style which is different from MSC in terms of lexicon and syntax. This research thus aims at studying verbal synonymy in the Pan-Chinese context with the corpus linguistic approach, which has been popular and useful in similar studies.
Apart from corpus studies, other experiments might help in discovering the mapping between dialect and formal literate. For example, the “subjective acceptability rating” (Arppe and Järvikivi, 2002) is used for testing the relation between corpus-based results and the linguistic intuitions of native speakers. Together with the syntactic and semantic characteristics of the verbs deduced from corpus data, we can test human subjects for their performance and competence on the verbs, so as to better study the process of verb acquisition, as part of our future work.
In Section 2, we will present the methodology, with which we analyzed the data for frequency distribution, semantic features and syntactic properties. The results will be presented in Section 3 and then discussed in Section 4. Section 5 will be the conclusion and the preview of future work.
2 MethodologyA subset of the LIVAC101 corpus (Tsou et al., 2000) was used in this study. The data consist of Chinese news texts from three Chinese-speaking communities: Hong Kong (HK), Taiwan (TW) and Beijing (BJ), collected from 1995 to 2002. The analysis on frequency distribution was based on the data of all seven years, and the sample sentences containing shuo1, biao3shi4, zhi3chu1 and cheng1 were extracted from the data of 1997-1998.
2.1 Frequency Distribution
From the subset of the corpus, frequencies of the four verbs in every year are calculated, as well as the total frequency of words. As the sentences containing the target verbs were sampled from 1997 to 1998, figures of that year were analyzed in detail.
2.2 Semantic Analysis
Within the data collected from 1997 to 1998, 50 sample sentences for each verb from each place were randomly selected for investigation. All sample sentences have the target verbs as the main verbs with objects corresponding to “message”102. The constituents of the sentences were annotated with one of the eleven abstract thematic roles, which are generally organized into “major” and “minor” roles:
Major roles: AGENT, PATIENT, EFFECTOR, THEMEMinor roles: GOAL, TIME, LOCATION, INSTRUMENT, SOURCE, FREQUENCY, QUANTITY
101 LIVAC consists of newspaper texts from Hong Kong, Beijing, Taiwan, Singapore, Shanghai and Macau, covering local news, international news, financial news, sports news, and entertainment news, collected over the same period of time.102 For example, in the sentence 他說實話(he says the truth), 實話(truth) is not the “message” communicated by 說(say), and therefore this sentence will not be sampled.
221
Major roles are essential to the formation of a sentence, and minor roles are optional. Features of the major roles and the distribution of the minor roles should give hints to the preference and constraint of a verb respectively. Three major features of AGENT were identified:
Name: Phrases with human proper name or surname as head Pronominal: Simple pronouns (e.g. 你/ 我/ 他/ 他們) excluding 誰/ 雙方/ 各人/ 後者/ etc. Inanimate: Inanimate subject excluding name of places or organizations. E.g. 它/ 報導/ 報告/ 消息/
文章/ 評論/ 會議Others103: E.g. 美國/ 首相/ 時報/ 美聯社
2.3 Syntactic Analysis
Our syntactic analysis was focused on the association pattern and the collocation behavior. With reference to the subcategorization frames classified by Roland and Jurafsky (1998), as well as the authentic corpus data, three frames were roughly identified:
Quotation104: Direct quotations with quotation marks. E.g. 江澤民 表示 「今後來香港的機會愈來愈多，但具體什麼時候去，現在就說不好」
Sentence: Complete sentences. E.g. 他們 說 該建築計劃會導致當地水浸Reflexive: Speaker being the subject of the clause. E.g. 董 稱 樂做任何有需要的事
3 Results3.1 Frequency Distribution
Among the transitive verbs in HK in the year 1997-1998, our target verbs shuo1, biao3shi4, zhi3chu1 and cheng1 ranked 2, 3, 7 and 8 respectively. The outstanding ranking of shuo1 can be explained mainly by its diversity in English translation105 (Ding, 2001), and hence the generality of its meaning. However, the explanations on the frequency distribution of the other three verbs need more in-depth syntactic and semantic analysis.
Table 1: Relative frequencies of the verbs from 1995 to 2002
First of all, we will look into the total relative frequencies106 of the four verbs in the three places from 1995 to 2002, which are shown in Table 1. It can be seen that shuo1 is similarly frequent in the three regions, and the figures of biao3shi4 and zhi3chu1 are comparatively low in TW, and that of cheng1 is high only in HK.
Figures were then broken down by year in order to observe the trends and tendencies of the relative distribution of each verb across the various places. Throughout the years in HK, the steady relative frequencies separated the four verbs into two groups, with zhi3chu1 and cheng1 at the bottom and shuo1 and biao3shi4 on the top. For BJ, relative frequencies of shuo1 and biao3shi4 were high and increasing, but that of cheng1 was far less than that of zhi3chu1. In HK and BJ, the verb with highest relative frequency alternated between shuo1 and biao3shi4. The figures of TW were different from those of the others, with shuo1 dominating the other verbs, while decreasing in relative
103 This category will not be shown in the results.104 Partial and embedded quotations are not counted.105 Ding (2001) translated shuo1 into over ten English verbs including “mention”, “state” and “describe”, and classified the translation into four means, namely direct, extended, alternative and elliptical translation.106 Relative frequency is equal to the frequency of the verb divided by the total number of words.
222
frequency.The trends of the relative frequency of the four near-synonyms within each place were then
investigated. In the seven years, the relative frequencies and the tendencies of shuo1 were similar among the three regions. Within the same period of time, the trend for biao3shi4 was diverging. Its relative frequency remained highest and increased over years in BJ data, but decreased and remained lowest in TW. Similar relative distributions were also seen in the data of zhi3chu1. For cheng1, large discrepancy was found, in which the figures of HK were often five times higher than those of TW and BJ.
In summary, from the distributions of the relative frequency, shuo1 exhibited greatest similarity among the three regions and cheng1 the least, and the trend of the four verbs differs in the three places.
3.2 Semantic and Syntactic Analysis
Totally 600 sample sentences were annotated. The figures are shown in Table 2 to Table 4107.
3.2.1 Summary of the Results
Table 2 shows the summary of the frequency distributions of the thematic roles among the four verbs, disregarding the sources of the corpus data. From the table, the general preference and restriction of the verbs were deduced.
Reflexive 12.00% 56.00% 2.00% 30.00% 100%Table 2: Summary of thematic role distributions
Interestingly, in every region, the minor roles FREQUENCY and LOCATION were absent from the monosyllabic verbs shuo1 and cheng1, and all the four near-synonyms could take proper names as AGENT and full sentences as PATIENT. Pronominal and inanimate AGENTs were not preferred by cheng1 and biao3shi4 respectively. Similarly, zhi3chu1 and cheng1 did not take reflexives and quotations as PATIENT respectively.
3.2.2 Cross-regional Comparison
The frequency distributions of the thematic roles for the four verbs were compared across the three regions, and the resulting percentages are shown in Table 3. The distributional differences showed the regional variation on the verbs’ usage.
For shuo1, constraints and preferences on both AGENT and PATIENT generalized from HK texts differ from those deduced from TW and BJ data, but the usages of shuo1 are similar in the latter regions. In HK, inanimate AGENTs are rarely used, and quotations and reflexives are mostly preferred.
For biao3shi4, greatest restriction is found in TW as FREQUENCY and LOCATION are always absent. In contrast, all the minor roles are found in both BJ and HK, and therefore the restriction of biao3shi4 is relatively flexible in these two regions. For the major roles, pronominal AGENTs and reflexive PATIENTs dominate in BJ. On the other hand, inanimate subjects and proper names are mostly used for AGENT in TW. The preference on PATIENT is similar in HK and TW.
In general, no significant similarity on the usages of zhi3chu1 is found among various regions,
107 As EFFECTOR, THEME, INSTRUMENT, SOURCE and QUANTITY were not seen in the sample sentences, their frequencies are not shown in the tables.
223
i.e. different complexities with similar extent are observed. Most inanimate AGENTs, and quotations and reflexives for PATIENT are found in BJ. For the minor roles, FREQUENCY and TIME are often seen in TW texts, whereas GOAL and LOCATION are only found in HK and BJ respectively.
For cheng1, strong preferences are seen in HK and TW as proper names and inanimate subjects are often taken as AGENT in these two places respectively. The fact that GOAL, FREQUENCY and LOCATION do not collocate with cheng1 in any region reflects the verb’s strong restriction. In short, the constraint on cheng1 is similarly high in the three places.
3.2.3 Cross-word Comparison
Table 4 shows the comparison of the four near-synonyms within each region. In HK, zhi3chu1 welcomes all kinds of AGENT. Pronominal and inanimate AGENTs often
collocate with shuo1 and cheng1 respectively, but seldom with biao3shi4. For the minor roles, GOAL comes with shuo1 most of the time, and the rest often come with biao3shi4. Discrepancy in the preference on PATIENT is observed in HK, such that shuo1, zhi3chu1 and biao3shi4 frequently collocate with quotations, sentences and reflexives correspondingly.
In TW, only TIME is frequently seen among all the four minor roles. For AGENT, pronominal and inanimate subjects are preferred by shuo1 and cheng1 respectively. For PATIENT, quotations often come with shuo1, and reflexives collocate only with biao3shi4 and cheng1.
From the BJ texts, surprisingly, the restriction on minor roles is the same for both the frequent shuo1 and the infrequent cheng1. On the other hand, biao3shi4 exhibited the least constraint as all the minor roles can collocate with it. For the major roles, biao3shi4 and cheng1 do not prefer inanimate and pronominal AGENTs respectively. For PATIENT, shuo1 and biao3shi4 prefer quotations and reflexives respectively.
4 DiscussionsDue to the nature of news reportage, the focus is more often on the content of the message than the speaker or other relevant information. The message content is best stored in direct quotations, and it will be highlighted when pronoun is serving as AGENT. The strong collocation tendency with quotations and the dominance of pronominal AGENTs can therefore accounts for the top ranking of shuo1 in the news texts.
The verb biao3shi4 is widely used in HK and BJ, sometimes even more than shuo1, as illustrated in Section 3.1. From the results of the semantic and syntactic analyses, biao3shi4 exhibited no constraint on minor roles in HK and BJ, and it only collocated with some of them in TW. The general high frequency of biao3shi4 is thus explained by its generally flexible usage. That this verb was used least in TW is probably because its usage is comparatively restricted in this region.
According to Section 3.1, biao3shi4 has higher frequency than zhi3chu1. As shown in Table 2, with respect to the capacity to collocate with minor roles, the usage of zhi3chu1 is as flexible as that of biao3shi4, although the former took on minor roles less often than the latter. This observation is therefore coherent with the claim of Liu, that semantically more restricted and marked verbs will have lower frequency than its more flexible counterpart.
As demonstrated in Section 3.2.2, significant regional variation was found on zhi3chu1 as there was inconsistency on its preference and restriction across the three regions. As a result, zhi3chu1 is worth further inspection of temporal and regional variation in linguistic convergence.
From Section 3.2.1, it was shown that the constraint of cheng1 is comparable to that of shuo1, with respect to the distributions of minor roles. These two verbs differ mainly on their preference on AGENT and PATIENT, where direct quotations and pronominal subjects do not frequently collocate with cheng1. However, the strong preference of the latter on inanimate AGENTs distinguished itself from the other three near-synonyms. In addition, with its monosyllabic characteristic, it is often seen in the headlines or subheadlines of the news texts. These characteristics keep cheng1 at a relative high position within all the reportage verbs. In addition, the restricted usage of these two monosyllabic verbs implied the relationship between a Chinese verb’s restriction and its word length.
In general, the restrictions of the four near-synonyms are identified. It was found that the constraints of shuo1 and cheng1 are similar with respect to the possession of the minor roles, and so
224
are those of biao3shi4 and zhi3chu1. These groupings are coherent to the definition of HowNet, as mentioned in Section 1. However, this analysis has gone further to unveil the difference in preference within a synonym group, and most importantly in the Pan-Chinese context.
5 Conclusions and Future WorkIn conclusion, this research distinguishes four near-synonyms by disclosing their underlying preferences and restrictions through semantic and syntactic analysis. The regional variation of near-synonyms across three Chinese-speaking communities is unveiled. It was found that, across the three regions, the same verb can have distinctive usages.
Temporal variation on linguistic convergence, which is the study of the chronological change in the usage of a verb, will be the next step. The synonym pair biao3shi4 and zhi3chu1, given their apparent difference in preferences and restrictions, will be suitable for such an investigation.
In addition, all the semantic and syntactic knowledge denoted by the annotations will provide useful data on verb argument structures for benchmarking the competence of humans in the verb acquisition process. The results and the methodology adopted in this study will help modify Chinese lexico-semantic resources and capture the linguistic norms in different Chinese-speaking communities, and thus contribute to the construction of Chinese semantic networks and Pan-Chinese thesaurus.
AcknowledgementsThis work is supported by Competitive Earmark Research Grants (CERG) of the Research Grants Council of Hong Kong under grant Nos.CityU1233/01H and CityU1317/03H.
References Arppe, A. and Järvikivi, J. (2002) Verbal Synonymy in Practice: Combining Corpus-Based and Psycholinguistic
Evidence. In Proceedings of the Workshop on Quantitative Investigations in Linguistics (QITL-02), Osnabrück, Germany. http://www.cogsci.uni-osnabrueck.de/~qitl/
Chief, L.C., Huang, C.R., Chen, K.J., Tsai, M.C. and Chang, L. (1998) What Can Near Synonyms Tell Us. International Journal of Computational Linguistics and Chinese Language Processing, 5(1): 47-60.
Chinese Knowledge Information Processing Group (CKIP). (1993) The Most Frequent Verbs in Journal Chinese and Their Classification. Taiwan: Institute of Information Science, Academia Sinica.
Ding, X. (2001) English Translation of the Chinese verb “Say” or “Speak”. Journal of Shangqiu Teachers College, 17(5).
Dong, Z. (1999) HowNet. http://www.keenage.com.Gao, H and Cheng, C.C. (2003) Verbs of Contact by Impact in English and Their Equivalents in Mandarin
Chinese. Language and Linguistics, 4(3): 485-508.Liu. M.C. (2003) From Collocation to Event Information: The Case of Mandarin Verbs of Discussion. Language
and Linguistics, 4(3): 563-585.Roland, D. and Jurafsky, D. (1998) How Verb Subcategorization Frequencies Are Affected By Corpus Choice.
In Proceedings of the 17th International Conference on Computational Linguistics and the 36 th Annual Meeting of the Association for Computational Linguistics (COLING-ACL 1998), Montreal, Canada, pages 1117-1121.
Tsai, M.C., Huang, C.R., Chen, K.J. and Ahrens, K. (1998) Towards a Representation of Verbal Semantics – An Approach Based on Near-Synonyms. In Proceedings of the 10th Conference on Computational Linguistics and Speech Processing (ROCLING X), pages 34-48.
Tsou, B.K., Tsoi, W.F., Lai, T.B.Y., Hu, J. and Chan, S.W.K. (2000) LIVAC, A Chinese Synchronous Corpus, and Some Applications. In Proceedings of the International Conference on Chinese Language Computing (ICCLC), Chicago, pages 233-238.
225
shuo1 (100%) biao3shi4 (100%) zhi3chu1 (100%) cheng1 (100%)HK TW BJ HK TW BJ HK TW BJ HK TW
Reflexive 27.78% 44.44% 0.00% 27.78% 0.00% 54.55% 0.00% 45.45% 4.76% 66.67% 4.76%Table 4: Cross-word frequency distributions of thematic roles in HK, BJ and TW
226
The Usage and Perception of Judgement Termsin the Pan-Chinese Context
Oi Yee Kwong and Benjamin K. TsouLanguage Information Sciences Research Centre
City University of Hong KongTat Chee Avenue, Kowloon, Hong Kong
{rlolivia, rlbtsou}@cityu.edu.hk
Abstract This paper reports on a synchronous corpus-based study of the everyday usage of a set of Chinese judgement terms. An earlier study on Hong Kong data found that these terms were more polysemous than their English counterparts within the legal domain where they originate; and are even more fuzzily used in general news reportage. The current study further compares their usage in general texts from other Chinese-speaking communities (Beijing, Taiwan, and Singapore) to explore the regional differences in lexicalisation and perception of the relevant legal concepts. Corpus data revealed the distinctiveness of the Singapore data, and that the contrasting frequency distributions of the terms and senses could be a result of the varied focus in reportage or the use of alternative expressions for the same concepts in individual communities. The analysis will contribute to the construction and enrichment of Pan-Chinese lexico-semantic resources, which will be useful to many natural language processing applications such as machine translation.
1 IntroductionIn this paper, we discuss a synchronous corpus-based study of the everyday usage of a set of semantically related legal terms in news reports from various Chinese-speaking communities. These terms were shown to be close in meaning and easily confusable (Tsou & Kwong, 2003). Although the confusion might be insignificant to the perception of the general readers, the subtlety therein could bear significant conceptual difference in the stricter legal domain, which is far less tolerable of impreciseness and ambiguity. Tsou and Kwong (2003) have discussed that while the legal system in Hong Kong has operated through English solely for more than 150 years, the difference in cross-lingual lexicalisation of legal concepts between English and Chinese has become a substantial linguistic hurdle in the implementation of legal bilingualism since the 90’s. It was observed, from a corpus of bilingual court judgments and a general corpus of news articles from Hong Kong, that the set of Chinese legal terms under investigation was considerably polysemous. They also identified sets of near-synonyms among the observed senses of the various Chinese legal terms, and attempted to arrange them in terms of their semantic relatedness into a verb hierarchy and a noun hierarchy, in a similar manner as in WordNet (Miller et al., 1990).
The current work aims at expanding on the above study, to further explore and analyse the different usages and finely grained senses of the same set of Chinese legal terms, but among news texts from various Chinese-speaking communities including Hong Kong, Beijing, Taiwan, and Singapore. From the usages of these terms as manifested in the corpus texts from different communities, we look for any local differences in the lexicalisation and thus the perception of the corresponding legal concepts, which might be a result of the differences in social structure or legal system.
Efforts have been made by researchers in lexical semantics on the study of semantic relations among Chinese lexical items, with a view toward organising the lexical items into semantic networks.
227
Gao (2001), for example, proposed a quantitative measure for the closeness and differentiation of near-synonyms among verbs denoting physical actions from a range of lexical semantic features. Cheng (2001) discussed the differentiation of related words from their individual focus and orientation. Nevertheless, work on lexical semantics and corpus-based lexicography often only drew reference from one particular corpus. Huang et al. (2000) worked on verbal semantics and near-synonyms, but of Mandarin Chinese as used in Taiwan only. Tongyici Cilin ( 梅等 , 1984) is based exclusively on Chinese as used in post-1949 Mainland. However, linguistic variation is significant and especially salient for Chinese language used in different communities (Tsou et al., 2004). Our corpus-based, Pan-Chinese approach, initiated with a set of domain-specific lexical items, thus has an additional advantage for its indigenousness, portability and versatility. Such a Pan-Chinese lexical resource, when done in large scale, would contribute to natural language processing applications like machine translation and would make a rich reference for legal and paralegal professionals, and more importantly, the resource would capture the linguistic norms from more than one Chinese-speaking community.
In Section 2, we first briefly review the polysemy of legal terms and the complexity manifested in the translation of legal terms from English to Chinese. Then in Section 3, we present the approach and the corpus analysis done in the current study. Results are discussed in Section 4, and we will conclude with future directions in Section 5.
2 Polysemy of Legal TermsTsou and Kwong (2003) started with a set of Chinese legal terms, all with senses related to “judgement”108 or “the action of judging” in the legal context, to study how the preciseness of legal concepts lexicalised in English is captured in their Chinese translations as shown in bilingual court judgments, and how the preciseness of the latter is in turn preserved in a general corpus. The set of “judgement” terms includes “裁定” (hold, convicted), “裁決” (determine, verdict), “判決” (judgement, conviction), “裁斷” (find, finding), and “裁判” (Magistracy)109. They observed that in Hong Kong, despite the implementation of legal bilingualism for several years, legal concepts are not as precisely lexicalised in Chinese as in English. The fuzziness is carried over from legal contexts as in court judgments to informal contexts as in news reports. For instance, “裁決” has been identified as the translation equivalent for “decision”, “verdict”, and “award” in a bilingual corpus of court judgments. Similarly, the word “decision” has been rendered as “裁定”, “決定”, “裁決”, and “判決”, among many possibilities. Such a complex correspondence (as shown in Fig. 1) between English and Chinese legal terms can be explained by the fact that English is much more maturely used in the Common Law system in Hong Kong. Many legal concepts are thus lexicalised and can be precisely expressed in English, whereas this preciseness is greatly weakened when terms are translated to Chinese.
108 An explicit distinction between the use of “judgment” and “judgement” is drawn here, as the inclusion or omission of the “e” is not arbitrary. “Judgment” refers specifically to the concluding writing upon a court trial, while “judgement” refers to the action of judging in general.109 The English terms are the more common translations of the corresponding Chinese terms as observed from Hong Kong court judgments, included here for reference only and are not necessarily the absolute or correct translations per se.
determine
decision
hold
verdict
award
find
裁定判決裁決裁斷
228
Fig. 1 Example of Multiple Renditions between English and Chinese Legal Terms
Despite the complexity of multiple renditions, the morphemic structure of the individual Chinese terms might nevertheless indicate a core sense of the term, and thus suggest the focus of the relevant concept. For example, sharing an identical morpheme “裁” (to judge), the words “裁定”, “裁決” and “裁斷” could be differentiated by their second morphemes, which focus on “conclusion”, “decision”, and “inference” respectively. This distinction is similar to Cheng’s (2001) discussion of word families where similar and related words could be differentiated by their individual focus and orientation, or meaning facet.
In the current study, we are interested to see if the same kind of polysemy appears in the usage of the same legal terms among different Chinese-speaking communities. It is hypothesised that we may not find exactly the same usage of the same terms among the various communities, as the different social structures and legal systems among them might lead to different perception of the corresponding legal concepts and the same concepts may not be equally salient for people in different communities. We will probe, from authentic corpus data, the perception of the various legal concepts by different communities, and see how the salience of “judgement” is reflected in the language used in different places; and if they do not use the terms in the same way, what words alternative to those investigated here are used to express similar concepts.
3 A Synchronous Corpus-Based Study3.1 MaterialsIn this study, we further analyse three terms (called “target words” hereafter) which Tsou and Kwong (2003) studied, namely “裁定” (hold, convicted), “裁決” (determine, verdict), and “判決” (judgement, conviction). We leave out “裁斷” (find, finding), as it was only found in the bilingual court judgment corpus but not at all in the general corpus (LIVAC, as introduced below), and “裁判” (Magistracy), as it mostly referred to the sense of “umpire” or “adjudication in a contest” when used in the general corpus.
Sentential contexts for the target words were extracted from a subset of the LIVAC corpus (Tsou et al., 2000). LIVAC (http://www.livac.org) is a synchronous corpus developed by the Language Information Sciences Research Centre of the City University of Hong Kong. The corpus contains mostly newspaper articles collected synchronously and regularly from six Chinese speech communities. The subset we use in the current study consists of texts from Hong Kong (HK), Beijing (BJ), Taiwan (TW) and Singapore (SG), covering local news, international news, sports news, entertainment news, and financial news, collected over the same period of time (for two years, 1997-98 and 2002-03). Each sub-corpus, i.e. texts from each of the four places, contains about 5M Chinese characters, which yields about 3M words upon segmentation.
3.2 The AnalysisFor each target word, 30 samples of their sentential contexts (where there is sufficient data) were randomly selected from each sub-corpus, and assigned a sense from the sense inventory as in Tsou and Kwong (2003) where appropriate. New senses were recorded when found. Upon sense tagging, the samples from the various sub-corpora were further analysed with respect to the sense distribution for each word in each community, and the similarity and difference of such distributions across the various communities. Again, the assignment of senses takes into account the collocation patterns of the different senses and subcategorisation patterns of the verbal usages where applicable. In addition, bi-syllabic words containing the
229
morphemes “裁”, “定”, “決” or “判” were retrieved from the sub-corpora. The retrieved words and their relative frequency were studied, to disclose any perceptual difference of the related legal concepts and any alternative expressions of such concepts by the different communities.
4 Results and Discussion4.1 Relative Frequency DistributionThe frequency of the target words from the various sub-corpora is shown in Table 1. The relatively low frequency of all the target words in BJ data is most notable. The small numbers readily indicate that court news does not enjoy as much attention in Beijing newspapers as in other places.
Word HK BJ TW SG
裁定 122 38 80 19
裁決 142 32 54 139
判決 160 66 210 341
Table 1 Frequency of the Target Words in the Sub-corpora
Just comparing the absolute frequencies, “判決” ranked highest across the board. Its relative frequency is especially high for TW and SG. In the most dramatic case of SG data, there are 341 occurrences of “判決” but only 19 occurrences of “裁定”. This is very different from, for example, HK data where the relative frequency of “判決” and “裁定” differs less than 10%. To a certain extent, this difference in relative frequency suggests the variation of focus in the reportage of the two communities, assuming their use of the words is not arbitrary. According to many legal dictionaries (e.g. 《法學詞典》編輯委員會, 1985; 劉清景, 2001), “判決” and “裁定” refer to different aspects of the ruling of a court. In particular, “判決” is often associated with the final determination on the main issue in a trial, whereas “裁定” usually refers to other factual disputes during the trial.
4.2 Sense DistributionSample sentences of the target words from BJ, TW and SG data were examined and each occurrence of the words was assigned a sense with reference to the sense set defined in Tsou and Kwong (2003). The sense distributions were compared to those reported for Hong Kong data in the same study. The results are tabulated in Tables 2 to 4 for “裁定”, “裁決” and “判決” respectively (the number in brackets next to each place refers to the number of samples checked and all figures reported are percentages). The second and third columns refer to data gathered from a bilingual corpus of Hong Kong court judgments and a subset of the current HK data from LIVAC respectively as reported in the earlier study. In this subsection we focus on the sense distribution with respect to individual words, and in the next we will further explore the regional difference observed.
4.2.1裁定“裁定” was earlier distinguished into four verb senses and two noun senses. No new sense was found from the data in the current study. The dominance of its verb usages is observed in all places except BJ. Over 70% of the samples for BJ were assigned sense 6 (as in “作出終審裁定”, “阻礙判決、裁定的執行”, “維持原判的裁定”, etc.). This is an enormous contrast not only with other places but also with data from the legal domain. There is also a dramatic contrast from SG data where no nominal usages
230
were found at all for “裁定”. This may be a consequence of the small number of SG samples but is more likely a genuine difference in the usage of the word, as we will further discuss below. Another interesting observation is that in the legal corpus, “裁定” is seldom used to state the order given by the court (sense 3) and the BJ data were more or less in line with this. However, over 25% of the samples from HK, TW and SG fall under sense 3. This thus raises interesting questions regarding the saliency of the concepts for individual places.
Sense and Example Legal (30) HK (30) BJ (30) TW (30) SG (19)1. [v.] the court decides on the outcome of a case
e.g. 法庭裁定…罪名成立。 43.33 43.33 3.33 16.67 31.60
2. [v.] the court resolves an issue in a casee.g. 法官裁定所提出的要求沒有得到滿意答覆。 36.67 20.00 6.67 30.00 26.30
3. [v.] the court gives an ordere.g. 法官裁定港府要即時釋放他們。 3.33 26.67 6.67 26.67 42.10
4. [v.] to judge on some issue to resolve disputee.g. 法庭需要裁定臨立會的合法性。 0.00 6.67 0.00 3.33 0.00
5. [n.] the resolution of an issue of disputee.g. …裁定回覆不能令人滿意。本席認為上述裁定正確無誤。 10.00 0.00 6.67 6.67 0.00
6. [n.] the decision on the outcome of a casee.g. 我認為暫委法官的裁定是正確的。 6.67 3.33 76.67 16.67 0.00
Table 2 Sense distribution of “裁定”
4.2.2裁決Sense tagging is notorious for its difficulty as the meaning in the new occurrence of a word is not always so clear-cut that a pre-defined sense could be unambiguously assigned to it. In this regard the tagging for “裁決” remained most difficult and confusing. The difficulty may be largely attributed to its relatively general meaning. For example, according to the hierarchies suggested in Tsou and Kwong (2003), the verb sense 裁決/1 and the noun sense 裁決/5 are the top nodes110 in their respective hierarchy. Having the most general sense amongst others, it means that the word can be used in a relatively wide variety of contexts. When it comes to general news reports, they are not necessarily the correct and legitimate contexts. This is evident in two respects from Table 3.
Sense and Example Legal (30) HK (30) BJ (25) TW (30) SG (30)
1. [v.] the court makes a decision based on evidencee.g. 若有法律觀點分歧，最終交由法庭裁決。 16.67 6.67 0.00 23.33† 33.33
2. [v.] the court decides on the outcome/sentence/etc.e.g. 法官裁決…把謀殺罪減為誤殺罪… 0.00 3.33 0.00 23.33 20.00
3. [n.] the court’s decision on the outcome of a casee.g. 陪審團達至誤殺的裁決。 30.00 6.67 20.00 13.33 0.00
4. [n.] the court’s decision on monetary compensatione.g. …拒絕執行公約裁決… 23.33 0.00 0.00 0.00 3.33
5. [n.] the court’s decision on a case and orderse.g. …會就法院上月底裁定港府要釋放十名越南人的裁決上訴。 16.67 53.33 40.00† 30.00 23.33
6. [n.] the resolution of an issuee.g. …關於證據接納性的裁決… 13.33 20.00 40.00† 0.00 0.00
*8. [v.] the court decides on the outcome of a case {裁定/1, 判決/1} 3.33 10.00
110 For the verb hierarchy, {裁決/1} subsumes {裁定/1, 判決/1}, {裁定/2, 判決/2}, {裁定/3, 裁決/2, 判決/3}, while {裁定/2, 判決/2} further subsumes {裁定/4}. For the noun hierarchy, {裁決/5, 判決/4} subsumes {裁定/6, 裁決/3}, {裁定/5, 裁決/6}, {裁決/4}, while {裁定/6, 裁決/3} further subsumes {判決/5}. Senses within curly brackets belong to the same synonym set. (Tsou & Kwong, 2003)
231
e.g. 法官路易斯迪蘇沙裁決他的罪名成立。*9. [v.] the court resolves an issue in a case {裁定/2, 判決/2}
e.g. …裁決搭客沒有起訴保險公司的權利。 6.67
*10. [v.] to judge on some issue to resolve dispute {裁定/4}
e.g. 三司必須裁決的只是關係到公眾利益的法律問題。 3.33
*11. [v.] the action of judging by referee in sports eventse.g. …但二壘審竟判定外野手是接殺，然後訴請裁決… 6.67
Table 3 Sense distribution of “裁決”
First, the use of “裁決” is almost abused in the SG data. Senses 8 to 10 in Table 3 were unexpectedly found from SG. Their low relative frequency and their absence from other regions suggest that these senses might more appropriately and specifically be replaced by the relevant senses of “裁定” and “判決” instead, as the examples for these senses in Table 3 do sound slightly unnatural. Second, the loose restriction on “裁決” is also reflected from the BJ and TW data (marked with † in Table 3) as it is often used to refer to the decisions (or the action of making decisions) from non-judiciary units (e.g. “完全由行政機關裁決” – TW, “世貿組織曾做出裁決” – BJ, etc.). Apart from that, an additional sense for “裁決” relating to referee’s judgement (sense 11) was found from the data; while sense 4 (decision on monetary compensation) is so domain-specific and technical that it is rarely found outside the legal documents.
4.2.3判決As seen from Table 4, all regions show a similar sense distribution for “判決”, where senses 4 and 1 are the major senses. Sense 5 (i.e. conviction) is specific enough to appear only in the legal texts. Additional uses referring to referee’s judgment in sports events were observed from TW and SG data.
Sense and Example Legal (30) HK (30) BJ (30) TW (30) SG (30)1. [v.] the court decides on the outcome of a case
e.g. 本席判決上訴得直。 13.33 20.00 20.00 36.67 16.67
2. [v.] the court resolves an issue in a casee.g. 上訴法院判決受託人有權提出呈請。 3.33 6.67 0.00 0.00 3.33
3. [v.] the court gives an order or sentencee.g. 國際法庭未被授權判決罪犯死刑。 0.00 6.67 3.33 3.33 0.00
4. [n.] the decisions made by court, and related orderse.g. …一宗上訴案的判決… 56.67 66.67 76.67 53.33 70.00
5. [n.] conviction, the judgment of being guiltye.g. 以上是二項罪名定罪判決的立場。 26.67 0.00 0.00 3.33 0.00
*6. [v.] the action of judging by referee in sports eventse.g. 眾多裁判判決爭議的確曾阻擾比賽。 3.33
*7. [n.] the judgement of referee in sports eventse.g. …因為不滿裁判的一個判決而摔球拍… 10.00
Table 4 Sense distribution of “判決”
4.3 Regional Variation
An obvious difference among the three target words is that verb uses tend to dominate for “裁定”111 whereas “裁決” and “判決” are more used as nouns. However, from Table 2, it can be seen that BJ has a lot more nominal usages (sense 6) of “裁定”. This observation is nevertheless in line with the
111 In particular, sense 1 and sense 6 of “裁定” both refer to the same meaning but differ in syntactic category.
232
findings in Kwong and Tsou (2003) on verb-noun categorial fluidity in Chinese, where texts from BJ were shown to have about 18% of the verbs undergoing the verb-noun shift, compared to about 15% for TW and HK data. Hence this linguistic phenomenon might account for the dominance of sense 6 of “裁定” in BJ data.
There remain some interesting questions regarding the difference in sense distribution across the various regions:
1. Table 2 shows that “裁定” is mostly used in SG to state the order given by the court, i.e. sense 3. Does this tell us anything about the salience of court orders in SG reportage?
2. Table 2 also shows that “裁定” is least used in sense 3 in BJ, compared to other places. The frequency of its synonymous senses 裁決/2 and 判決/3 is also extremely low in BJ data. So is the concept missing or alternatively expressed?
3. 裁決/3 is absent from SG data, and so is its synonym 裁定/6. Do SG news reports not care about verdicts? Otherwise where have they absorbed the concept with other expressions?
The difference in sense distribution across various regions is on the one hand a result of the different linguistic norms and styles of language use, as exhibited by the dominance of nominal usages in BJ. Hence even though BJ does not use 裁定/3 or its synonyms, the relevant concepts might have been expressed via nominal uses such as 判決/4. On the other hand, it could reflect the varied approaches and perception by different communities regarding the judgemental concepts. For instance, these concepts are apparently less salient in BJ contexts given the relatively low frequency of the target words in BJ data. Moreover, since SG is found to use 裁定/3 heavily but not 裁決/3, it suggests that SG news tends to treat the conclusion (verdict) and the consequence (sentence and order) as a whole. Meanwhile, a preliminary survey of the words mined from the seed morpheme “判” reveals that words like “判刑”, “判監”, “判罰”, “判囚” and “判處” (all related to sentencing) are relatively more abundant in HK or TW than in SG, suggesting that HK and TW tend to distinguish between the verdict and the sentence more clearly. The use of these words and their relations with the target words calls for further investigation.
5 ConclusionThus in this study we have further analysed the usage and sense distribution of a set of closely related legal terms pertaining judgement in the Pan-Chinese context. Linguistic data reveals the variation in salience of these concepts in various Chinese-speaking communities and the distinctiveness of the SG data. Based on the subtleties among various uses, we have further probed the salience of these concepts in the various communities, which might be a result of the difference in legal system. For instance, the use of the target words in HK and SG might be more influenced by translation from English legal terms than in BJ and TW. Alternatives for expressing similar and related legal concepts should be further explored and the study should be expanded with other sets of closely related legal terms.
As mentioned in the beginning of this paper, the subtle differences among the target words may be insignificant for the general readers. However, when it comes to high quality translation, especially translations which bear legal implications, the preciseness therein will definitely be indispensable. Hence our analysis and study on the use of legal terms in various Chinese-speaking communities will provide useful information for the construction of a Pan-Chinese legal term lexical resource as we witness the growing maturity of the Chinese language in the legal domain. Such an enriched lexical resource should be useful to legal and paralegal professionals, and for legal document translation between English and Chinese, by machine or by human, as well as for many other natural language processing tasks.
233
AcknowledgementsThis work is supported by Competitive Earmarked Research Grants (CERG) of the Research Grants Council of Hong Kong under grant Nos. CityU1233/01H and CityU1317/03H.
References Cheng, C.C. (2001) 論同義詞的語意輕重與側重 . Presented at the Second Workshop on Chinese Lexical Semantics,
Peking University, Beijing, China. Gao, H. (2001) A Specification System for Measuring Relationship among Near-synonyms of Physical Action Verbs.
Presented at the Second Workshop on Chinese Lexical Semantics, Peking University, Beijing, China.Huang, C-R., Ahrens, K., Chang, L-L., Chen, K-J., Liu, M-C. and Tsai, M-C. (2000) The Module-Attribute Representation
of Verbal Semantics: From Semantics to Argument Structure. Computational Linguistics and Chinese Language Processing, 5(1): 19-46.
Kwong, O.Y. and Tsou, B.K. (2003) A Synchronous Corpus-Based Study of Verb-Noun Fluidity in Chinese. In Proceedings of the 17th Pacific Asia Conference on Language, Information and Computation (PACLIC 17), Singapore, pp. 194-203.
Miller, G.A., Beckwith, R., Fellbaum, C., Gross, D. and Miller, K.J. (1990) Introduction to WordNet: An on-line lexical database. International Journal of Lexicography, 3(4): 235-244.
Tsou, B.K. and Kwong, O.Y. (2003) When Laws Get Common: Comparing the Use of Legal Terms in Two Corpora. Language and Linguistics, 4(3): 609-629.
Tsou, B.K., Lai, T.B.Y. and Chow, K. (2004) Comparing Entropies within the Chinese Language. In Proceedings of the First International Joint Conference on Natural Language Processing (IJCNLP-04), Sanya, Hainan, pp. 475-481.
Tsou, B.K., Tsoi, W.F., Lai, T.B.Y., Hu, J. and Chan, S.W.K. (2000) LIVAC, A Chinese Synchronous Corpus, and Some Applications. In Proceedings of the ICCLC International Conference on Chinese Language Computing, Chicago, pp. 233-238.《法學詞典》編輯委員會 (1985) 《法學詞典（增訂版）》 上海辭書出版社
是次研究只集中在負面詞，至於正面詞的情況如何，則是我們將來的研究方向之一。參考 LEHRER, A. 1974. Semantic Fields and Lexical Structure. North Holland, Amsterdam and New York.HATZIVASSILOGLOU, V., AND MCKEOWN, K.R. 1997. Predicting the semantic orientation of adjectives.
Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and the 8th Conference of the European Chapter of the ACL, Madrid, Spain, 174-181.
HATZIVASSILOGLOU, V. AND WIEBE, J.M. 2000. Effects of adjective orientation and gradability on sentence subjectivity. Proceedings of 18th International Conference on Computational Linguistics (Coling’00), Saarbrücken, Germany.
Abstract To assign semantic roles in building Treebanks, there is a need for annotators having a guideline in determining semantic relations between phrasal head and its modifiers or arguments. Semantic roles are hard to have clear-cut definitions. It is not always easy to determine thematic relations between two concepts. This paper aims to introduce an integrated nominal modifier system. Basically we adopt other scholars' incisive idea in analyzing semantic roles that modify general nouns. We use the approach of building a fine-grain taxonomy of role system. The taxonomy of fine-grain thematic roles makes the role determination easier for human annotators, since the meaning of a fine-grain semantic role is self explanatory and a higher-level semantic role is described by its hyponyms. The proposed taxonomy has been attested during construction of Sinica TreeBank and HowNet definitions of nominal concepts and proven to be more applicable than conventional flat structures.
1 IntroductionTo assign semantic roles in building Treebanks, there is a need for annotators having a guideline in determining semantic roles. Semantic roles establish the semantic relations between phrasal head and its modifiers or arguments. The verbal modifiers describing the semantic roles of an event are much discussed in various kinds of research, and FrameNet is the most representative among them. However discussions regarding semantics roles modifying nouns are scattered.
Semantic roles are hard to have clear-cut definitions. For instance, it is not always easy to determine thematic relations between two concepts. Dowty (1991) proposed criteria for determining proto-agent and proto-patient which means sometimes it is hard even to differentiate between agent role and patient role. We use the approach of building a fine-grain taxonomy of role system. The taxonomy of fine-grain thematic roles makes the role determination easier for human annotators, since the meaning of a fine-grain semantic role is self explanatory and a higher-level semantic role is described by its hyponyms. Another important advantage of taxonomic approach is that any flat fine-grain role system suffers the problem of completeness of role coverage, but a taxonomic role system can always find a coarse-grain hypernym role if a proper fine-grain role doesn’t exist in the system.
2 BackgroundAccording to Pustejovsky(1991, 1995), the defining attributes of an object is named as the qualia structure. Pustejovsky argues that the qualia structure of a word tells us the concept it carries in a language. The qualia is further specified into four aspects: constitutive, formal, telic and agentive. He defines constitutive "the relation between an object and its constituents". Formal in qualia distinguishes the object within a larger domain. The role telic shows the purpose and function of the object. And agentive are factors involved in how the denotation of the word came into being. The structured representation which gives relational force of a lexical item is well-discussed in qualia by Pustejovsky(1995).
Barker(1998) put forth 20 noun modifier relationships (NMRs) in his semi-automatic
program115. However the individual NMR in this parallel framework tends to seem incomplete or fragmented. The firmness of the NMRs is to be challenged since one can add or delete a NMR easily. This is resulted from the scattered elements. Thus we decide to adopt a hierarchical order in our taxonomy of fine-grain semantic role for nominal modifiers.
3 Our analysisAt top level of taxonomy, we divide the semantic role for nouns into five categories: predication, quantifier, possessor, scene, and property. The following is our preliminary framework of fine-grain semantic role for nominal modifiers. The five categories are observable in real language and our data would not stray from these five. The subcategories in each role may not be exhausted or perfect, but they demonstrate a clearer picture on how the five roles function. We will introduce the skeleton of our analysis in the sections followed.
3.1 PossessorWithin the five, possessor and property are the two roles we get confused easily. We encounter some difficulties in assigning the semantic roles. The idea of possession seems to have less problem in examples like, wo de ya sui qian 我的壓歲錢 'my spring festival money gift' or zi ji de yi fu 自己的衣服 'one's own clothes'. The posessees in the two examples are separable from the possessors and the concept of owning something is vivid. Nonetheless, we hesitate to give the possessor role on some data like zu fu de yang zi 祖父的樣子 'the form of grandfather' and yu fu de shang huo 漁夫的生活 'the fisherman's life'. Does the grandfather own the form and does the fisherman have the life?
Heine (1997) compares many previous studies about possession in human languages. He uses criteria as alienable and inalienable, temporary and permanent, animate and inanimate, and abstract possession to subcategorize possession into seven. Hein's analyses are clear and ordered. But for us, some of the criteria are not so important as to be distinguished from the
115 The NMRs in Braker (1998) include "Agent (agt), Beneficiary (benf), Cause (caus), Container (ctn), Content (cont), Destination (dest), Equative (equa), Instrument (inst), Located (led), Location (loc), Material (matr), Object (obj), Possessor (poss), Product (prod), Property (prop), Purpose (purp), Result (resu), Source (src), Time (time), Topic (top)".
270
top. For instance, the physical possession, temporary possession, and permanent possession116
in his analysis are combined to ownership in ours. The ownership means the alienable possession with an inanimate possessee. Heine leaves the inalienable possession117 to a single category. However, we consider the body-parts and relative relation which are described as inalienable possession in his study to be different; on the contrary, we think that body-parts (inalienable possession) and part-whole relationship (inanimate inalienable possession118) are similar, since they denote parts of their respective substances. Thus we have whole-part relation, head noun being a part of the modifier as well as relative in our study. In Heine's discussion, possession includes the inanimate alienable possession, such as the sentence The tree has crows on it. Yet we settle this group of data in scene for "the tree" is actually a position and the possessing connection between the two is not so close.
3.2 Property
We subdivide the property into constitutive and formal, following Pustejovsky's work. The constitutive includes the material and component parts of the noun. The phrase yu mi nong tang 玉米農湯 'corn puree' is on of the constitutive examples. Formal is describing the target noun's evaluation, attribute, apposition, gender or name. For instance, in kun nan de ren wu 困難的任務 'difficult task', kun nan is a difficulty evaluation of formal. We refer to Dixon (1982) on the features of attributes and evaluations of nouns119.
3.3 QuantifierGil (2001) distinguishes the internal and external quantification. In his studies, the external quantifiers is in the scope of syntax and thus beyond our discussion. His analysis of internal quantification is briefed as follows: 1. mass vs. count quantifiers 2. existential vs. universal quantifiers 3. numeral 4.others. Similarly our quantifier is subdivided into four, including existential, universal, mid-range quantifiers and numerals. Existential quantifiers mean that there exists something or someone to make the proposition real. The word you ge 有個 'some, certain' can be an existential quantifier which modifies nian tou 念頭 'idea'. Universal quantification refers to the range within the group, as the word quan bu 全部 'all' modifying sai cheng 賽程 'game'. The massive degree in-between existential and universal is called mid-range quantification, like shao shu 少數 'few', da duo 大多 'most' and others. Then the last type of the quantification is the numeral quantifiers. As a classifier language, the numeral quantifiers in Mandarin usually denote exact numbers along with classifiers as si tiao 四條 'four' is a possible numeral
116 According to Heine the definition of the three is as follows. Physical possession: the possessor and the possessee are physically associated with one another at reference time. Temporary possession: the possessor can dispose of the possesee for a limit time but cannot claim ownership to it. Permanent possession: the possessee is the property of the possessor, and typically the possessor has a legal title to the possessee.117 Inalienable possession: the possessee is conceived of typically as being inseparable from the possessor, e.g. body-parts, relative..118 Inanimate inalienable possession: part-whole relationship, inanimate possessor. The possessee and the possessor are conceived of as being inseparable119Dixon (1982) classifies English adjectives as dimension, age, value, color, physical property, human propensity, speed, difficulty, similarity, qualification.
271
quantifier of yu魚 'fish'.
3.4 SceneLu (2001) uses "scene" in depicting the eventive case roles120. We adopt Lu's idea of scenic description and diminish it to adjust the nominal modifiers. The scenic description of location, time and domain are three self-explaining roles of nominal modifiers. The word Taipei is a locative use in Taipei de jie yun 台北的捷運 'Taipe's MRT', while the san xioa shi 三小時 'three hours' describes the duration of dian ying 電影 'movie'. Meantime, in xuan mei huo dong 選美活動 'beauty contest', we see that xuan mei is the domain of huo dong.
3.5 PredicationThe predication denotes the head noun would play an argument role rather than adjunct. The predication is an eventive modifier of the head noun. Pustejovsky's agentive aspect of qualia falls in the predication role of our system. Examples of predication are as below: ni zuo de huo gou 你做的火鍋 'the hot pot you made'. We can see that ni zuo de你做的 'you made' shows how the huo guo 火鍋 'hot pot' come into being, and also 'the hot pot' is the object in the clause of 'you made'. As to the telic use of predication, usually means that there is an omitting verb between the modifier and the head noun, such as shui guo pan 水果盤 actually means the plate used for placing fruits. The simplified phrase is without the predicate "used for placing". We arrange this kind of modifiers in telic use. Thus form the two predication use of the head noun.
4 Future workWe consult previous studies on each subcategory of the semantic roles for nouns, look for evidence form corpus and then modify the framework to fit Mandarin. The proposed taxonomy has been attested during construction of Sinica TreeBank and HowNet definitions of nominal concepts and proven to be more applicable than conventional flat structures. In the future we expect to make a complete fine-grain role system for not only nominal but also eventive semantic roles. Furthermore, we would like to compare and contrast even unify our work with other similar projects, such as FrameNet and HowNet.
References Barker, Ken & Stan Szpakowicz. 1998. Semi-Automatic Recognition of Noun Modifier RelationshipsProceedings of COLING-ACL '98. Montréal: 96-102.Chappell and McGregor. 1996. The grammar of inalienability: A typological perspective on body part terms
and the part-whole relation. Berlin, New York: Mouton de Gruyter.Dixon, R.M.W. 1982. Where have all the adjectives gone? Mouton Publishers, Berlin.Dowty, David R. 1991. Thematic Proto-roles and Argument Selection. Language, 67(3):547-619.Gil, David. 2001. Quantifiers. In Language typology and language universals : an international handbook,
edited by Martin Haspelmath et al.Heine, Bernd. 1997. Possession: cognitive sources, forces and grammaticalization. Cambridge: Cambridge
University Press.Li, N Charles and Sandra A. Thompson. 1989. Mandarin Chinese: A Functional Reference Grammar.
Berkeley: University of California Press.魯川 (Lu, Chuan). 2001. 漢語語法的語意網絡. 北京：商務印書館.Nastase, Vivi and Stan Szpakowicz. 2003. Exploring Noun-Modifier Semantic Relations. Proceedings of
International Workshop on Computational Semantics. Tillburg, Netherlands.Nichols. 1988. On alienable and inalienable possession. In In honor of Mary Haas: from the Haas festival 120See Lu (2001) for further discussion.
272
conference on native American Linguistics, ed. by Shipley. Berlin: Mouton de Gruyter.Pustejovsky, James. 1995. The Generative Lexicon. Cambridge: MIT Press.
Department of PsychologyChung-Yuan Christian University, Taiwan
INTRODUCTIONDevelopmental dyslexia refers to a specific deficit on reading and writing. Dyslexic
subjects have normal intelligence and adequate leaning opportunity, yet cannot acquire age-appropriate reading skills. Studies done on many languages repeatedly show dyslexic readers suffer from phonological-related deficits (Bradley & Bryant, 1985; Mann & Brady, 1988). They have poor phonological awareness. Their phonological short term memory is significantly poor, even though their general memory is intact. They have problem on phonological encoding, and their phonological representation of is fuzzy. There are also other accounts being proposed. Magnocellular hypothesis illustrated dyslexia has visual processing deficit (Livingstone, Rosen, Drislane, & Galaburda, 1991; Chen, 2000). Connectionism model suggested dyslexics can also suffer from orthographic processing deficits (Harm & Seidenberg, 1980; 1999).
Most of the dyslexia studies in literature were conducted on alphabetic writing sytems (e.g, English). As logograph, Chinese has very different orthography, and the learning processes are also quite different. This leads us to a question: will Mandarin-Chinese dyslexic subjects show the same behavioral pattern as English readers? Researches seemed to suggest phonological ability is as important for Chinese readers as for English readers. 陳淑麗 and 曾士杰 (1999) demonstrated phonological awareness is critical to character perception, verbal achievement, and can be a good predictor on reading comprehension. Other studies also suggested the important role that phonological ability/awareness plays in reading Chinese (柯華葳&李俊仁, 1996; 洪慧芳, 1993; 黃秀霜, 1997; Hu & Catts, 1998; McBride-Chang & Ho, 2000). We can then conclude Mandarin-Chinese dyslexic readers also suffer from phonological deficit. Therefore, the purpose of current study was to investigate whether dyslexics also have difficulty on semantic-related lexical access.
METHODSThree experiments were conducted. There were three groups of subjects: 25 Dyslexia
subjects (D, average of 13.4 yr), 22 Reading Level Controls (RL, average of 10 yr), and 17 Chronological Age Controls (CA, average of 13.5 yr). The first study was a naming speed test. Dyslexic children usually have a very slow reading speed. Therefore, we would like to eliminate the possibility that their problem was due to a more general slow reaction capacity. Four different types of stimuli were introduced: objects, numbers, juyin symbols, and colors. There were 50 items for each type, presented visually. Table 1 was the results. The result between D and RL were not significantly different on all conditions, but both groups were slower than CA group. This result seemed to suggest the slower naming speed in this experiment cannot be purely accounted for by reading ability of D group.
Table 1. Naming Speed in Experiment 1 (in second).Stimuli D RL CA comparison
Color naming 41.7 39.1 37.8 n.s. among groupsObject naming 46.3 43.3 35.4 RL > CA*, D > CA *Digit naming 22.9 20.0 17.3 D > CA *Juyin naming 35.9 31.4 27.3 D > CA*** significant at .05 level. ** significant at .01 level.
We then asked whether dyslexic readers have orthographic rules. There were totally 16 low frequency characters, 16 pseudo-characters, 32 non-characters in study 2. 24 items were selected after pilot study. Subjects were to decide do they think each item look like Chinese characters. All three groups yielded almost perfect scores (D = 22.1, RL = 22.3, CA = 23.4) out of 24 points. No significant difference was found. This indicated that the problem of dyslexia was not due to lack of orthographic rule alone. In other words, the result of experiment 2 suggested dyslexics may have intact orthographic rule, other factors (phonological representation and/or semantic similarity) may play crucial roles.
Now, we all know dyslexic subjects have reading problem. And experiment two showed their orthographic rule may be intact. This leaded us to another question: what went wrong in the process from script to meaning? We already knew they do have phonological processing problem (from previous research), therefore, one question remained: do they also suffer from semantic-related processing. We would investigate this in study three with a semantic-related deficit experiment. A 2 X 2 complete within subjects design was employed. There were totally 40 trials, 10 in each condition. Half of the questions were “yes” and the other half were “no” questions. In each trial, experimenter would read a phrase, and a character would be presented in front of the subject visually, and the subject was to make a Yes/No judgment. For example, the phrase would be “彩虹的虹？” (Is this the “hong” in the word of rainbow?), and subject saw “紅” (red), and the correct answer would be “no”. Based on the relation of two characters being compared (one by auditory, one by visual presentation), there were two independent variables: Physical Similarity (sharing a component or not), and Semantic Relatedness (Table 2). In all four conditions, the two characters for comparison were homophones. Reaction time and accuracy were the recorded. The ANOVA showed two significant main effects (Group, F(2,60)=23.424, p<,001; Semantic Relatedness, F(1.60)=7.501, p<.05). Semantic also showed significant interaction with Group, and Physical Similarity. Pair-wise comparisons (table 2) were conducted. Besides, regression analysis for D group, Semantic Relatedness can account for 31.7% of totally variance, and Physical Similarity had no significant effect. For RL group, Semantic Relatedness can account for 40.5% of totally variance, and Physical Similarity can account for only 4.3% of total variance. For CA group, Semantic Relatedness can account for 31.7% of totally variance, and Physical Similarity can account for 9.8%. In sum, this paradigm indicated the poor performance of dyslexia can be accounted for by semantic relatedness.
Table 2. Design and Results in Experiment 3 (number of accurate responses out of 5 trials).Physical Similarity No Physical Similarity
* significant at .05 level. ** significant at .01 level. *** significant at .001 level.
DISCUSSION AND CONCLUSIONSExperiment 1 suggested naming speed was not the main problem of dyslexics.
Experiment 2 suggested Mandarin-Chinese dyslexics also have certain orthographic rule and ability. In experiment 3, since all target characters were homophones, the well-known phonological impact has been kept constant. We can therefore conclude the semantic relatedness between the two characters did reveal some of the difficulty dyslexic children encounter. Dual Deficit Hypothesis (Bowers & Wolf, 1993; Wolf & Bowers, 1999; Wolf, Bowers, & Biddle, 2000) was then discussed for a theoretical interpretation of the result.
mechanisms, and orthographic skill in dyslexia. Reading and writing：An
Interdisciplinary Journal, 5（1）, 69-85.
Bradley, L., & Bryant, P.（1985）. Rhyme and reason in reading and spelling. Ann Arbor：
284
University of Michigan Press.
Harm, M. W., ＆ Seidenberg, M. S.（1999）. Phonology, Reading Acquisition, and
Dyslexia：Insights From Connectionist Models. Psychological Review, 106（3）,
491-528.
Hu, C.F., ＆ Catts, H.W.（1998）.The role phonological processing in early reading ability：What we can learn from Chinese？ Scientific Studies of Reading, 2,55-79.
Livingstone, M. S., Rosen, G. D., Drislane, F. W., & Galaburda, A. M.（1991）.
Physiological and anatomical evidence for a magnocellular defect in developmental
dyslexia. Proceeds of the National Academy of Science, 88, 7943-7947.
Mann, V. A., ＆ Brady, S.（1988）.Reading disability：The role of Language Deficiencies.
Journal of Consulting and Clinical Psychology, 56（6）, 811-816.
McBride-Chang, C. & Ho, C. S. H.（2000）. Developmental issues in Chinese children’s
character acquistion. Journal of Experimental Child Psychology, 58, 112-133.
Wolf, M., Bower P. G., ＆ Biddle, B.（2000）. Naming-Speed Processes, Timing, and
Reading：A Conceptual Review. Journal of leaning disabilities, 33（4）, 387-407.
AbstractThe Gross Control Model[GCM]of Semantic Vocabulary as Dictionary with
Examples[SVDE]is a new theory or method or tool。There are tow kinds of GCM in man-com-net including the Gross Text Control Model[GTCM]and the Gross Syllable Control Model[GSCM]on Language Understanding。GTCM means 0~16 levels with all kinds of texts in GLPS。GSCM means 1~n levels with all kinds of syllables in SVDE。The meanings of SVDE controlled by geminate numbered list("1...n" to "1...n")。SVDE with mother tongue is made by the way of that the meanings of words are paraphrased by phrases in the same language such as Chinese according to the law of "1" to "n" as "word" to "its paraphrased phrases"。Bilingual SVDE is based on Co-ordinate Bilingual or Compound Bilingual according to the law of "1" to "1"("1...n" to "1...n") as "the numbered list of all meanings "to" the numbered list of all the paraphrased phrases both in Chinese and in English"。Quality currency 一、绪言在人机协作网络（融智[1]系统）中有两种总量控制模型，即：关于自然语言理解的
到的情形。因此字典的編纂，若是能盡量考慮到不同地區的發音，在查詢到與編纂者不同語音腔調時能引導讀者翻閱至與其語義相對應的頁面，如此也能讓辭典更有系統，也大大減少不同地區的閩南人查閱上的困難。另外，在閩南語有音無字方面，一些字典自創本字，不同的字典自創不同的本字，常造成讀者的疑惑。由於各地閩南方言的差異和生活習慣的不同，所以除了這二十幾種閩南語手部動作之外，一定還有更多豐富的辭彙未收錄進來。除此之外，在決定以什麼為語義分類依據時，這些手部打人動作中，是否有程度上的差異？這方面字典上並沒有提及，這幾點值得以後再深入研究。五、參考書目Lai, Huei-ling 2000 The Lexicalization Patterns of Verbs of Hitting in Hakka, 《國立政治大學教師語言學研究成果發表/研討會論文集》。
參考文獻：Chinese characteristics(1894) ， Arthur H. Smith， New York : Fleming H. Revell, c
陳融(1986) ，「面子‧留面子‧丟面子」，《外國語》第四期Politeness：Some universals in language usage(1987)Brown, Penelope./Levinson, Stephen C./Penelope Brown andStephen C. Lev/Cambridge University Press
它们的逻辑关系可表示为：（p ∧∧ q）→ r （→表蕴涵）而且只有当 p尤其是 q存在时，才可能存在 r 这样的判断，因此，Ⅰ式完整的逻辑公式应
为：p∧q ∧r(（p ∧∧ q）→ r ) （∧表并列关系）2.1.2 表转折关系的Ⅱ式2.1.2.1句法结构特征（3）他不懂他为什么知道是不能的，却偏要去向往，简直是搬起石头砸自己的脚。 C D ZⅡ 式中，Z仍然是一个表结果的主观结论句，它和前面的 SnP+CONNt-conn也依然是一种因果关系，只是 SnP+CONNt-conn所包含的两个语义成分之间的关系与Ⅰ式的迥然不同。C是一个客
A F Z在假设条件句型中，SnP+CONNt-conn一般只包含Ⅰ式的客观背景句 E和假设条件句 F。客观背景句 E 为下面“简直”句型的生成提供了一个可发生的前提。假设条件句 F，表示在 E情况下的假设；Z同样是一个主观结论句，是在 E背景下，从 F假设中推倒出来的结论。这个句型的联结标记一般是“如果，那么，要不，若”等。用符号表示为：J＝E＋（R＋Z）
Ellen F.Prince, The ZPG Letter: Subjects, Definiteness, and Information-Status, In Thompson, S. and Mann, W., eds. Discourse description: diverse analyses of a fund raising text. Philadelphia/Amsterdam: John Benjamins B.V. pp. 295-325. 1992.Jeanette Kohn Gundel, Nancy Hedberg, Ron Zacharski, Topic-Comment Structure, Syntactic Structure and Prosodic Tune, Workshop on Prosody and Grammar in Interaction, Helsinki, Finland, August 13-15.Mira Ariel, 2001, Accessibility Theory: An Overview. In Ted Sanders, Joost Schilperoord and Wilbert Spooren eds. Text representation: Linguistic and psycholinguistic aspects. Amsterdam: John Benjamins.29-87.Perter C.Gordon, Randall Hendrick, Comprehension of Coreferential Expressions, Meeting of the Association for Computational Linguistics: Proceedings of the Workshop on the Relation of Discourse/Dialogue Structure and Reference. pp 82-89. New Brunswick, NJ: Association for Computational Linguistics.Tao, Hongyin, 1996, 《Units in Mandarin conversation: prosody, discourse, and grammar》, Amsterdam; Philadelphia: John Benjamins Pub. Co., 1996.陈平，1991，《现代语言学研究：理论‧方法与事实》，重庆：重庆出版社，1991年。符淮青，2000，《词义的分析和描写》，北京：语文出版社，2000年。胡壮麟，1994，《语篇的衔接与连贯》，上海：上海外语教育出版社，1994年。廖秋忠，1992，《廖秋忠文集》，北京：北京语言学院出版社，1992年。王寅，2003，“认知语言学与语篇分析”，见人大复印资料 2003年第 8 期，原载《外语教学与研究》2003年第 2 期 83－88页，北京。徐赳赳，1995，“话语分析二十年”，见《外语教学与研究》1995年第 1 期。詹卫东，2003，“框架语义学”，《语言学论丛》第 27辑，2003年，北京：。郑贵友，2002，《汉语篇章语言学》，北京：外文出版社，2002年。朱永生、严世清，2001，《系统功能语言学多维思考》，上海：上海外语教育出版社，2001年。
322
《人民日报》语料库专有名词分类的研究The Chinese Named Entity categorization on People’s Daily Corpus于浩 西野文人
YU Hao NISHINO Fumihito富士通研究开发中心有限公司Fujitsu Research & Development Center Co.,LTD.
Abstract In order to explore the interpsychological force-dynamic interaction, this study, adopting Talmy’s (2000a) force-dynamic framework, examines (i) how FORCE is manifested in verbs of urging in Hakka, and (ii) the correlation between the semantic content FORCE and the syntactic manifestation. In adopting these verbs to study the interpsychological force-dynamic interaction, this study modifies Talmy’s framework and also provides a fine-grained analysis for verbs of urging in Hakka.
1 IntroductionForce-dynamics, which concerns the exertion and resistance of a force, the overcoming of such a resistance, and the like, figures in an essential place in language structure. Talmy (2000a) develops a system of diagrams to analyze three different domains of force-dynamic patterns, including psychophysical, intrapsychological, and interpsychological domains. Among them, the interpsychological force-dynamic interaction, involving force interactions between sentient entities is especially complicated. Verbs of urging in Hakka can be used as a prototypical case of this domain. Therefore, this study, adopting and modifying Talmy’s (2000a) force-dynamics framework, has the following two purposes. First, it demonstrates the interpsychological force-dynamic patterns of the verbs of urging in Hakka. Second, it examines the correlation between the semantic component FORCE and the syntactic manifestation of the verbs.
This paper is sequenced in the following way. Talmy’s force-dynamics model will be outlined in Section 2. The analysis will be presented in Section 3. And finally, Section 4 will conclude this paper.
2 Theoretical FrameworkAs encoded in language, force interaction preponderantly involves two forces opposing each other. The focal force entity is called the Agonist and the opposing force is called the Antagonist by Talmy (2000a). Talmy makes extensive use of diagrams to represent different force-dynamic patterns. The basic elements of this framework are shown in (1) below.(1) a. Force entities b. Intrinsic force tendency Agonist (Ago): toward action: >
Antagonist (Ant): toward rest: �
c. Balance of strengths d. Resultant of the force interaction stronger entity: + action: -------->-------
Moreover, in Talmy’s model, force-dynamic interactions can be divided into EXTENDED
329
pattern and ONSET pattern. The EXTENDED pattern is a steady-state force-dynamic pattern with the Antagonist impinging steadily on the Agonist. In contrast, the ONSET pattern indicates that the Antagonist has come into position against the Agonist or has moved out of a condition of impingement on the Agonist.
3 AnalysisIn this section, the data description will first be presented in Section 3.1. Next, the semantic elements for verbs of urging in Hakka will be shown in Section 3.2. Then the cases concerning hard-to-resist urging and easy-to-resist urging will be discussed in Section 3.3 and Section 3.4, respectively.
3.1 Data descriptionVerbs of urging in interpsychological domain involve the force-dynamic interactions
between two sentient entities---the urger producing the force of urging with a piece of strong advice, and the urgee refusing to perform the requested action by producing the force of resistance. Nine verbs of urging in Hakka are found and listed in (2) below.(2) Urging
harder to resist easier to resist
hak2 cui24
cat2 kien55
giuk2 bet2
dui31 gam24
hem24
We can see that verbs of urging in Hakka are located toward different ends of the scale. Hard-to-resist urging refers to the act of urging which is difficult for the urgee to resist. On the other hand, easy-to-resist urging refers to the kind of urging which is not too difficult for the urgee to resist.
3.2Semantic elements for verbs of urging in Hakka
Some modified elements of Talmy’s (2000a) model will be adopted to examine verbs of urging in Hakka. The basic elements of the force-dynamic interaction are depicted in (3) below.(3) Semantic elements of the force-dynamic framework for urging
a. Agonist b. Antagonist (urgee) (urger)
the urgee’s tendency: does not want to perform the urger’s intention: to force the urgee to perform some action a particular action
c. Volition for the urger and the urgee d. The resultant state of urging ------------------------- (the requested
action is performed) ----------�------------- (the requested
action is disobeyed)
3.3 Hard-to-resist urging
� >
V
330
As shown in the scale, the verbs hak2, cat2, giuk2, and dui31 belong to this type of urging. This force-dynamic interaction is diagrammed in (4) below.
(4)a. b. c.
-----●-------
This force-dynamic script shows that only phase (a) is windowed, while phase (b) to (c) is gapped (shown by the gray background of the surrounding rectangular). The following two examples show that only phase (a) is placed into the foreground of attention, and the occurrence of phase (b) to (c) is unknown. See the following example.(5) A24-min11 giuk2 gia11 moi55-e31 hi55 hok5-gau31.
“A-min urged his daughter to go to school.”In example (5), we do not know whether A-min’s daughter went to school or not at that moment.
This example shows that after verbs of hard-to-resist urging is used, we just focus our attention on the beginning phase. Moreover, the example below illustrates that the urger’s force of urging will become stronger than the urgee’s force of resistance, and the requested action needs to be fulfilled, which matches the urger’s desired result.(6) Bun24 gia11 ba24 gin31 cat2, A24-min11 ciet5-do55 voi55 ten11-den31 zon31-vuk2-ha24.
“Urged iteratively by his father, A-min definitely will return home right away.”
The felicitous use of the adverb of judgment ciet5-do55 ‘definitely’ in (6) to confirm the speaker expectation is based on the lexicalized implicature of the urging verbs cat2---the force of the urger’s urging becomes too strong for the urgee to resist.
The windowed phase of this interpsychological force-dynamic interaction belongs to the ONSET DESPITE pattern because in this beginning phase, despite the force of urging, the real result does not match the urger’s desired result. However, in the gapped final phase, it is an ONSET CAUSATION pattern because in this phase, the real result matches the intention of the urger---causing the urgee to perform certain kind of action.
3.4 Easy-to-resist urging
Verbs that are located toward the other end of the scale include: kien55, cui24, bet2, gam24, and hem24. Whether it is resistible or not depends on the relative power concerning the social status or the vicious power struggle between the urger and the urgee, and also the personality of the urgee himself.
Easy-to-resist urging can be categorized into two subtypes, one concerning a stronger force of urging, and the other a stronger force of resistance. Because of the limitation of the length of the paper, we will only show the diagrams of these two subtypes of forceful interactions with several illustrative examples.
(8) force of the urger’s urging (+) > force of the urgee’s resistance (-)
a. b. c.
------�------
When the urger has greater power over the urgee, and when the urgee is receptive or obedient, the force of the urger’s urging will be stronger than the force of the urgee’s resistance, and thus the
+�> +
�
>�
>+
VV V
VV V
+�> +
�
>�
>+
VV V
VV V
331
requested behavior will be accepted and performed successfully by the urgee. It is shown in this force-dynamic script that only phase (a) is windowed, while phase (b) to (c) is gapped (e.g. Gia11 me11 kien55 gi11 zo31-diam31 soi55-muk2.). Although phase (b) to (c) is placed into the background of attention, it can be predicted that after the force of the urger’s urging gets greater, finally this force will even become stronger than the force of resistance. As a result, in phase (c), the urge will be obeyed and the requested behavior will be fulfilled, shown by the arrowhead on the resultant line (e.g. A24-min11 an31 hien11-hau55, bun24 gia11 me24 kien55-kien5-a31-le11, ziin55-gin24 ciu55 goi31-go55-hiong55-san55.).
The diagram of the second subtype is shown below.(8) force of the urger’s urging (-) < force of the urgee’s resistance (+)
a. b. c.
------�------It is shown in this force-dynamic script that only phase (a) is windowed, while phases (b) and (c) are not only gapped but also blocked. More specifically, phases (b) and (c) do not occur, and the non-occurrence of these phases is placed into the background of attention. Since phases (b) and (c) do not occur, the urgee’s force of resistance remains stronger despite the urger’s force of urging. This stronger force of resistance is derived from the urgee’s personality involving rebelliousness. See the following example.(9) A24-min11 o24-sii31-mang11-ngang55, but2-gon31 gia11 me24 ngiong31-e11 kien55, ciet5-do55 han11-he55 sii31-cien11-siong55-sii31-sui31.
“A-min, young and inexperienced, absolutely will continue to waste money regardless of his mother’s urge (not to).”
4 ConclusionA scale is proposed to depict the correlation among the urger’s social status and power of persuasion, the urger’s force of urging, the urgee’s personality of rebelliousness, and the urgee’s force of resistance, as given in (10).(10) Urger’s social Urger’s force Urgee’s force Urgee’s
status and power of urging of resistance personality ofof persuasion rebelliousness
+ +
ONSET CAUSATION
ONSET DESPITE + +The analysis proposes that verbs of urging in Hakka are located toward
different ends of the scale, one concerning hard-to-resist urging, and the other easy-to-resist urging. As for the hard-to-resist urging, the beginning phase of it belongs to the ONSET DESPITE pattern, while the final phase belongs to the ONSET CAUSATION pattern. The easy-to-resist urging, on the other hand, can be further grouped into two subtypes, one concerning a stronger force of urging, and the other a stronger force of resistance. Because the first subtype of easy-to-resist urging involves the shift of the stronger force, the beginning phase of it is the ONSET DESPITE pattern while the final phase is the ONSET CAUSATION pattern. Next, because in the second subtype of easy-to-resist urging, the force of resistance
+�> +
�
>�
>+
VV V
VV V
332
remains stronger than the force of urging, only the ONSET DESPITE pattern is involved. Generally speaking, for verbs of urging in Hakka, the relative strengths between the urger and the urgee lie in the relative social status and power between the urger and the urgee and also the urgee’s personality of rebelliousness. Moreover, some syntactic manifestations including Hakka BUN-constructions, the adverbs of duration, evaluation and judgment, and the resultative complements are used to examine different patterns in the interpsychological domain.
After examining these verbs from a force-dynamic perspective, this study not only provides a more fine-grained and extensive analysis for the verbs of urging in Hakka but also modifies Talmy’s framework to some extent.
