Post on 18-Dec-2015
transcript
Overview
Introduction
What is Natural Language Generation (NLG)? Usages of Natural Language Generation
When NLG Systems are appropriate?
Applications of NLG
Example NLG System Architectures for NLG How to Evaluate NLG Systems Conclusions Demo on ILEX System
What is NLG
“Natural language generation is the process of deliberately constructing a natural language text in order to meet specified communicative goals”.
[McDonald 1992] from [Dale and Reiter 1999]
“Natural Language Generation (NLG) is the process of constructing natural language outputs from non-linguistic inputs”.
[Jurafsky and Martin 2000]
What is NLG (contd.)
Non-linguistic Input NLG System Output Text Goal:
produces understandable and appropriate texts in English or other human languages
Input: some underlying non-linguistic representation of information, e.g. Meteorological maps, Airline/Railway schedule databases
Output: documents, reports, explanations, help messages, and other kinds of texts
Knowledge sources required: knowledge of language and of the domain
[Dale and Reiter 1999]
Text vs. Graphics
which medium is better? Computer Generation vs. Human Authoring
is the necessary source data available?
is automation economically justified? NLG vs. Simple String Concatenation
how much variation occurs in output texts?
[Reiter and Dale 1999]
When NLG Systems are Appropriate?
Applications of NLG
Automated Document Production
weather forecasts, summarizing statistical data, answering questions etc.
Information Presentation medical records, weather forecast etc.
Entertainmentjokes, stories, poetry etc.
Teaching Dialog Systems [Rambow et al., 2001]
Applications of NLG (contd.)
Two Types of NLG Systems
The system produces a document without human help
summaries of statistical data, generating weather forecast etc.
The system helps human authors to create documents
customer-service letters, patent claims, technical documents, job descriptions etc.
[Reiter and Dale 2000]
NLG System: FoG Reiter and Dale give the description of the FoG System as follows Function:
Produces textual weather reports in English or French Input:
Graphical weather depiction User:
Environment Canada (Canadian Weather Service) Developer:
CoGenTex Status:
Fielded, in operational use since 1992 [Reiter and Dale 1999]
NLG System Architectures:
Text Planner
Linguistic Realiser
Sentence Planner
From [Jurafsky and Martin 2000]
Goal
Text Plan
Sentence Plan
Surface Text
From [Reiter and Dale 1997]
This component starts with a communicative goal and makes choices of Content selection Discourse Plan Lexical selection Micro planning
Aggregation Referring expressions
It selects the content from the knowledge base and then structures that content appropriately
The resulting discourse plan will specify all the choices made for the entire communication
Discourse Planner
Content Selection
Content Selection: is the process of deciding what information should be communicated in the text
Creating a set of MESSAGES from the underlying data sources
Message-creation process and the form and content of the messages created are highly application-dependent
Generally messages are expressed in some formal language (e.g., Sentence Planning Language) with the notion of ENTITIES, CONCEPTS and RELATIONS in domain
Content Selection (contd.)
For Example, specific trains, places and times as entities, the property
of being the next train as a concept, and departure as relation between
trains and time.
Message-id: msg01Relation: IDENTITY
Arguments: arg1: NEXT-TRAIN
arg2: CALEDONIAN-EXPRESS
The next Train is the Caledonian Express
Message-id: msg02Relation: DEPARTURE
Arguments: departure-entity: CALEDONIAN-EXPRESS
departure-location: ABERDEEN
departure-time: 1000
The Caledonian Express leaves Aberdeen at 10 am
Discourse Plan
Discourse Planning is the task of structuring the messages produced by the Content Selection process
Two predominant mechanisms for building discourse structures: Text Schemata Rhetorical Relation
Text Schemata
Figure: Knowledge Base Representation for saving a file as a simple procedural hierarchy
Figure: A Schema for expressing procedures
From [Jurafsky and Martin 2000]
Drawbacks of Text Schemata
Impractical when the text being generated requires more structural variety and richness of expressions.
For example, we express certain segments of the text in a different manner or in different order.
No higher-level structure relating the sentence together.
For example, if we explained a process in some detail, we might not want to do it again.
Rhetorical Relation Rhetorical Structure Theory (RST), is a descriptive
theory of text organization based on the relationships that hold between parts of the text. Example:i) I love to collect classic automobiles. My favorite car is my 1899 Duryea.ii) I love to collect classic automobiles. My favorite car is my 2001 Toyota.
It designates a central segment of text, called nucleus, and a more peripheral segment, called the satellite
RST relations are defined in terms of the constraints placed on the nucleus, on the satellite, and on the combination of both
Rhetorical Relation (contd.)
Name: Expand PurposeEffect: (COMPETENT hearer (DO-ACTION ? action))Constraints: (AND (c-get-all-substeps ?action ?sub-actions) (NOT (singular-list? ?sub-actions))Nucleus: (COMPETENT hearer (DO-SEQUENCE ?sub-actions))Satellites: (((RST_PURPOSE (INFORM s hearer (DO ?action)))
Name: Expand Sub-ActionsEffect: (COMPETENT hearer (DO-SEQUENCE ? actions))Constraints: NILNucleus: (foreach ?actions (RST-SEQUENCE (COMPETENT hearer (DO-ACTION ?actions))))Satellites: NIL
From [Jurafsky and Martin 2000]
Lexical Selection
Lexical selection: the process of deciding which words and phrases should be used in order to transform the underlying messages into a readable text.
Handling lexical selection requires that the generation system deal with two issues:
It must be able to choose the appropriate lexical item when more than one alternatives exists.
The generation system must be able to choose the appropriate grammatical form for the expression of the concept.
[Jurafsky and Martin 2000]
Micro Planning
AggregationThe process of grouping messages together into sentences; not always necessary. Example:
"You’ve just compiled and run a simple C program."
Without aggregation:
"You’ve just complied a simple C program. You’ve just run a simple C program."
Referring ExpressionsSelecting words and phrases to identify entities (e.g. Caledonian Express or it or this train), generating deictic expressions.
Surface Realization
This component receives the fully specified discourse plan and generates individual sentences as constrained by its lexical and grammar
If the plan specifies multiple-sentence output, the surface realizer is called multiple times
No general consensus as to the level at which the input to the surface realizer should be specified
Approach for Surface Realizations
Functional Unification Grammar
Functional Unification Grammar
Functional Unification Grammar uses unification to manipulate and reason about feature structure
Unify the available grammar with an input specification which is represented with the same feature structure
The unification process then takes the features specified in the input and unify with those in the grammar, producing a full feature structure which can then be linearized to form sentence output
Functional Unification Grammar (contd.)
Sample Output: The system will save the document
Propositional content specification: a saving action done by a system entity to a document entity
Specification of the grammatical form: a future tense assertion and lexical items (“save”, ”system”, and “document”).
Functional Unification Grammar (contd.)
Input (functional description)
CAT S
ACTOR [HEAD [LEX SYSTEM] ]
PROCESS HEAD [LEX SAVE ] TENSE FUTURE
GOAL [HEAD [LEX DOCUMENT] ]
From [Jurafsky and Martin 2000]
Reusable Surface Realization Packages
FUF: is a reusable package to generate English grammar This package is developed using functional unification
structures If the grammar and the input are specified then the
system will construct the syntactically correct sentence output
Drafter is a system to support the production of software documentation in English and French.
Drafter [Power et al., 1998] is built using the FUF for surface realization
It uses Rhetorical Structure Theory (RST) based planning for Discourse planning
Evaluating Generation Systems
In early work, the quality of the NLG system was assessed by the system builders themselves. If the system gives correct output then the system was judged as success.
Currently Convene a panel of experts to judge the output of the
generator in comparison with text produced by human authors
Judge how effective the generated text is at achieving its goal.
[Jurafsky and Martin 2000]
Conclusions
Many NLG applications being investigated but all are not successful. However, few systems are in use e.g., FoG
Currently the evaluation process of NLG systems has received much attention
In late 1980s and early 1990s the trend was to construct reusable NLG system e.g., FUF. Now the trend is to port the systems to other languages and platforms
ReferencesJurafsky D., and Martin J.H. 2000. “Speech and Language Processing, An
Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition”. Prentice Hall.
Reiter E., and Dale R., 1997. “Building Applied Natural Language Generation”. Cambridge University Press.
Reiter E., and Dale R., 2000. “Building Natural Language Generation Systems”. Cambridge University Press.
Bateman J., and Zock M., 2001. “The B-to-Z of Natural Language Generation: an almost complete list.” Oxford Handbook of computational Linguistics.
Rambow O., Bangalore S., and Walker M., 2001. Natural Language Generation in Dialog Systems.
Reiter E., and Dale R., 1999. Building Natural Language Generation System. www.csd.abdn.ac.uk/~ereiter/papers/eacl99-tut.ppt
Power R., Scott D., and Evans R., 1998. What You See Is What You Meant: direct knowledge editing with natural language feedback.
Elhadad M., 1993. FUF: the Universal Unifier User Manual Version 5.2FUF: http://www.cs.bgu.ac.il/research/projects/surge/index.html