Sentiment Analysis Based
on Chinese Thinking Modes
Yang LiangYang Liang
2 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Introduction
Thinking models
Description of Chinese Sentiment Expression Model
Implicit Chinese Sentiment Expression Mining Based on LSA
Experiment Setting and Evaluation
Outline
3 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Sentiment analysis ——a popular research topic in NLP in recent years
◆ Blog ,twitter, comment
Sentiment analysis in China
◆ phase level, sentence level(COAE)
◆ Less work in passage level
Our work
◆ relationships between thinking modes and language
◆ Chinese sentiment expression model and LSA
Introduction
4 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Introduction
Thinking models
Description of Chinese Sentiment Expression Model
Implicit Chinese Sentiment Expression Mining Based on LSA
Experiment Setting and Evaluation
Outline
5 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Spiral Graphic Mode and Straight Line Mode
Concreteness and Abstractness
Scatter view and Focus view
Thinking models
6 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Spiral Graphic Mode
◆ reflects in the passage organization
◆ the topic of the passage is discussed after examples
Straight Line Mode
◆ focus on deduction and thinking in a straight line way
◆ tend to state their views directly and frankly
Spiral Graphic Mode and
Straight Line Mode
7 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Example
◆ (1) Chinese:
“他被眼前的一幕震惊了。”
English:
“He was shocked by what he saw.”
◆ (2) Chinese:
“经过反复的思考,我终于得到了完美的答案 ”。
English:
“I got a perfect answer after deeply thinking.”
Spiral Graphic Mode and
Straight Line Mode
8 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Spiral Graphic Mode and Straight Line Mode
Concreteness and Abstractness
Scatter view and Focus view
Thinking models
9 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Concreteness
◆ Chinese uses quantities of specific words, shapes, sounds and description to illustrate abstract things
Abstractness
◆ English tend to implement general vocabularies and their variants to express abstract feelings or opinions, such as “-ion”, “-ance” and “-ness”
Concreteness and Abstractness
10 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Example
◆ (1) Chinese:
土崩瓦解。
English:
Disintegration
◆ (2) Chinese:
有志者,事竟成 English:
When there is a will , there is a way.
Concreteness and Abstractness
11 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Spiral Graphic Mode and Straight Line Mode
Concreteness and Abstractness
Scatter view and Focus view
Thinking models
12 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Scatter View
◆ Chinese tend to emphasize unified whole
◆ Example : more than one verb is used in one Chinese sentence
Focus View
◆ English pay more attention to logical reasoning or deduction
◆ express their feelings or emotions briefly thinking
Scatter View and Focus View
13 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Example
◆ (1) Chinese:
他拿着课本走进了教室。
English:
He walked into the classroom with a textbook in hands.
◆ (2) Chinese:
他们俩青梅竹马,两小无猜
English:
The boy and the girl were playmates in their childhood.
Scatter View and Focus View
14 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Introduction
Thinking models
Description of Chinese Sentiment Expression Model
Implicit Chinese Sentiment Expression Mining Based on LSA
Experiment Setting and Evaluation
Outline
15 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Quantification of “Spiral Graphic Mode”
◆ emotion-determining words mostly locate the end part of Chinese sentences
◆ the closer ai locates the tail, the larger the score(ai) will be
Description of CSE Model
1 |i ii
score A position a count a A score a
16 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Quantification of “Concreteness”
◆ Chinese sentiment expression, verb also plays an important role,“脸红”“溃败”
◆ the highest priority to adjective, then the higher priority is given to the verb, finally other words are processed
Description of CSE Model
n m l
adj verbi j k
i j k
score a W W W
17 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Quantification of “Scatter View”
◆ View window to simulate the Scatter View
◆ fixed at 6
Extend Resource
◆ CRF, COAE
Description of CSE Model
18 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Quantification of Similarities between Thinking Modes
◆ The similarities between Chinese and English
◆ Example:
1)Chinese: “这个酒店什么都好,就是服务让人失望。”
English: “Every aspect about the hotel is ok except the disappointing service.”
2)脸红, blush
Description of CSE Model
19 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Introduction
Thinking models
Description of Chinese Sentiment Expression Model
Implicit Chinese Sentiment Expression Mining Based on LSA
Experiment Setting and Evaluation
Outline
20 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
The Criterion of Implicit Emotion Sample
◆ Emotion expressed in an indirect way
◆ the implicit emotion articles ,low scores
◆ group of samples are chosen to determine the threshold,(from DUTIR Emotion Ontology )
Implicit Emotion Classification Based on LSA
◆ LSA, relationships between the implicit samples and the seed samples
Implicit Chinese Sentiment Expression Mining Based on
LSA
,j i i jk jkk
score A sim A s score s
21 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Introduction
Thinking models
Description of Chinese Sentiment Expression Model
Implicit Chinese Sentiment Expression Mining Based on LSA
Experiment Setting and Evaluation
Outline
22 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Experiments of Chinese Thinking modes in different domains
◆ Corpus , three domains, 4000 hotel reviews, 1608 electronics reviews and 1047 stock reviews
◆ result
Experiment Setting and Evaluation
23 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Elec
◆ adding each Chinese thinking mode the precision has increased.
◆ elec-pos reviews is not increasing obviously——too much noun
Experiment Setting and Evaluation
24 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Stock
◆ stock-neg data set is not good
◆ great number of specialized words exist and part of them does not appear in DUTIR emotion ontology
Experiment Setting and Evaluation
25 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Experiment of Chinese sentiment expression model and LSA
◆ Experiment results on ChnSentiCorp
◆ Statistic data of implicit samples
Experiment Setting and Evaluation
Lexicon Semantic CSE
Pos 1575/1828(86.16%) 1502/1828(82.17%) 1575/1828(86.16%)
Neg 1727/2163(79.84%) 1811/2163(83.73%) 1827/2163(84.47%)
Total 3302/3991(82.74%) 3313/3991(83.01%) 3402/3991(85.24%)
Lexicon CSEPos-Implicit 751/1575(47.68%) 640/1575(40.63%)Neg-Implicit 836/1827(45.76%) 624/1827(34.15%)
Total 1587/3402(46.65%)
1264/3402(37.15%)
26 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Secondary classification by implementing LSA
◆ Macro-Average-Precision of different methods in ChnSen
Experiment Setting and Evaluation
27 23/4/21ir.dlut.edu.cnir.dlut.edu.cn
Thanks for your attention