Transcript ppt
Extracting Opinion Topics for Chinese Opinions using Dependence Grammar Guang Qiu, Kangmiao Liu, Jiajun Bu*, Chun Chen, Zhiming Kang Zhejiang University 浙江大學 SIGKDD Workshop on Data Mining and Audience Intelligence for Advertising ADKDD 07 Reporter: Chia-Ying Lee Advisor: Prof. Hsin-Hsi Chen Introduction Problem Definition: Determining opinion sentence, and extracting the topic from a opinion sentence. Advertisements promoting systems recommend without considering the sentiment polarity of the texts. A reasonable advertisement should be about that of a rival or solutions to user’s complain Kim, Soo-Min and Hovy, Eduard: Opinion is described as a quadruple including Topic, Holder, Claim and Sentiment. Most of previous work only focuses on sentiment classification assuming topics are given in prior. 2 Related Work (1/2) Sentiment Classification 1.Hatzivassiloglou and McKeown, 1997 Pairs of adjectives conjoined by and, or, but, either-or, or neither-nor 2.Wiebe, 2000 Focuses on subjectivity tagging which distinguishes opinion sentences 3 Related Work (2/2) 3.Esuli and Sebastiani, 2005 Text classification by glosses. 4.Ku, Liang, and Chen, 2006 Sentiment orientation of sentences can be concluded from that of words 5.Pang, Lee and Vaithyanathan, 2002 Machine learning methods:Naive Bayes, Maximum Entropy and SVM. 4 Method - Acquire sentiment words (1/2) Assumption: 1. Regard sentences with sentiment words as opinion ones 2. Topic is assured to exist in these opinion sentences Data set of sentiment words: WS1: NTUSD 2812 positive words and 8276 negative words D1: Emotion classification by Bruce and Wiebe 1256 positive blog articles and 1238 negative 5 Method - Acquire sentiment words (2/2) D2: Blog search results of Baidu 372 names of products as queries, 24146 snippets Manual label 1685 snippets of POS, NEG, NEU Select adj. correlated with of the adj. in D1. Calculate the probability of each word occurs in each sentiment category WS1+D1+D2: 3269 positive words and 9621 negative words 6 Method - Extracting the topics using rules (1/4) <ROLE_SENTI, RELA, ROLE_TOPIC> 7 Method - Extracting the topics using rules (2/4) 1. <VOB, SIBLING, SBV> 2. <DE, GRANDPARENT-SIBLING, SBV> 8 Method - Extracting the topics using rules (3/4) 3. <ATT, PARENT-SIBLING, SBV> 4.<HED, CHILD, VOB> 9 Method - Extracting the topics using rules (4/4) 5. <HED, CHILD, SBV> 6. <ADV, SIBLING, VOB> 7. <ANY, NEAREST, ANYNOUN> 10 Experiments and Result(1/2) Data: Blog search results of Baidu 372 queries 22461 snippets unlabeled as POS, NEG, NEU 51661 sentences Two annotators to annotate topics, and POS, NEG, NEU 570 sentences 250 for sentiment and 320 for neutral 11 Experiments and Result(2/2) Opinion sentence SVM (using unigram words as the features) Topic extraction 218 sentences are correctly extracted out of 250 opinion sentences, with the accuracy of 87.2%. Exectly match 12 Conclusion Proposed a rule-based approach to extracting topics in opinion sentences. Employ a syntactic parsing on sentences and take advantage of the syntactic roles of words and their dependency relationships to extract the topic. 13 Future Work Negation words Noise filtering method Co-reference resolution Enlarge current rules to cover more situations 14 HIT LTP system http://ir.hit.edu.cn/demo/ltp/ltp_v2.0.py 15 Thank you! 16