Opinion Sentence Search Engine on Open-domain Blog Osamu Furuse, Nobuaki Hiroshima, Setsuo Yamada, Ryoji Kataoka NTT Cyber Solutions Laboratories, NTT Corporation IJCAI 2007 Reporter: Chia-Ying Lee Advisor: Hsin-Hsi Chen Introduction A user want to know others’ opinion about a product or something Input: query phrase Output: relevant opinion sentence Explicitly stated opinion sentence Exclude quoted or implicational opinions Including positive, negative and neutral Opinion clues 2008/01/15 Chia-Ying Lee 2 Opinion Sentences to be Searched Opinion clues [Hiroshima et al. 2006] 1. Evaluative adjectives placed in the predicate part This house is beautiful. 2. Subjective sentential adverbs(全句副詞) Amazingly, few people came to my party. 3. Idiomatic collocations (慣用語搭配) in main clause My wish is to go abroad. 2008/01/15 Chia-Ying Lee 3 Architecture of Opinion Sentence Search(1/3) 2008/01/15 Chia-Ying Lee 4 Architecture of Opinion Sentence Search(2/3) High proportion off-line processing Presented in a blog page unit Ranked according to The number The ratio Total strength 2008/01/15 Chia-Ying Lee 5 Architecture of Opinion Sentence Search(3/3) 2008/01/15 Chia-Ying Lee 6 Opinion Sentence Extraction(1/4) Opinion clue expression collection Top 20 web pages with 40 queries 13,363 opinion sentences judged by all 3 evaluator 2,936 opinion clues judged by human Japanese predicates are in principle placed in the last part of a sentence. 2,514 clues in predicates ; 422 not in. 2008/01/15 Chia-Ying Lee 7 2008/01/15 Chia-Ying Lee 8 Opinion Sentence Extraction(3/4) Augmentation by semantic categories Opinion clue expressions and cooccurring words (X) The sky is high. (O) The quality of this product is high. 2008/01/15 Chia-Ying Lee 9 Opinion Sentence Extraction(4/4) Binary classifies sentences using SVM Features: 2,936 opinion clue expressions 2,715 semantic categories 150 frequent words 13 parts of speech Number Train set Test set Query 72 18 Total sentence 23,800 5,686 Sentence at least one judged opinions 8,050 1,791 2008/01/15 Chia-Ying Lee 10 Query-relevant Sentence Extraction Accepting weak query relevance Strategies about query relevant (a) A query phrase occurs in the sentence or within some number of sentences before the sentence. (b) A query phrase occurs in the sentence or within the chunk opinion sentences consecutively appear in. 2008/01/15 Chia-Ying Lee 11 Experiment ─ Opinion Sentence Extraction(1/2) Baseline: Regards a sentence contain more than 4 opinion clues as a opinion sentence 2008/01/15 Chia-Ying Lee 12 Experiment ─ Opinion Sentence Extraction(2/2) Quasi predicate part: Features are permitted only within the last ten words in the sentence Recall: 74.3 % to 3 judged 62.0 % to 2 judged 44.4 % to 1 judged 2008/01/15 Chia-Ying Lee 13 Experiment ─ Query Relevance Baseline: A query phrase occurs in the sentence Data set: 2,868 opinion sentences 2008/01/15 Chia-Ying Lee 14 Experiment ─ total performance Data set: 429 query-relevant opinion sentence out of 5,686 sentences (7.5%) Opinion sentences tend to be more query-relevant than non-opinion sentences. 2008/01/15 Chia-Ying Lee 15 Conclusion and Future Work The experiments suggested that the system is a practical application Improve query-relevant strategy Classify opinion sentences in terms of emotion, sentiment, requirement, and suggestion Summarize 2008/01/15 Chia-Ying Lee 16 Thank You! 2008/01/15 Chia-Ying Lee 17