An Opinion-aware Approach to Contextual Suggestion
0 likes92 views
This document describes an opinion-aware approach for contextual suggestion. It discusses gathering candidate suggestions from sources like Yelp, modeling suggestions and user profiles, ranking candidates, and generating personalized descriptions. Candidate suggestions are crawled from various sources and categories. User profiles are modeled based on the opinions expressed in reviews of example suggestions. Candidate suggestions are ranked based on comparing their reviews to the user's positive and negative profiles. Descriptions are generated with an opening, official introduction, highlighted reviews tailored to the user, and conclusion. Preliminary results found the opinion-based methods to be more effective.
![Representation of User Profiles
Unique Full
Reviews
Unique terms from the original review excluding stop words
Review
Summaries The review summaries generated by Opinosis [1].
21
… From the stunning architecture to the croissant and latte
served up in the food court downstairs. Go to this place and
ask why all train stations can't be like this!
Wow, over 100 tracks. Unbelievable architecture. Shopping,
food. Etc. it is amazing. We ate at the oyster bar last time
and that was a treat. The oyster pots are quite something.
Original review
K. Ganesan, C. Zhai, and J. Han. Opinosis: a graph-based approach to abstractive summarization of highly redundant
opinions. In Proceedings of the 23rd International Conference on Computational Linguistics, COLING ’10, pages 340–348,
Stroudsburg, PA, USA, 2010. Association for Computational Linguistics.](https://image.slidesharecdn.com/trec2013-contextualsuggestion-talk-180314040645/85/An-Opinion-aware-Approach-to-Contextual-Suggestion-20-320.jpg)
An Opinion-aware Approach to Contextual Suggestion
- 1. An Opinion-aware Approach to Contextual Suggestion TREC 2013 Contextual Suggestion Track Peilin Yang and Hui Fang University of Delaware 1
- 2. Overview of Our Methods 2 Candidate Suggestion Gathering Example Suggestions Suggestion Modeling User Profile Modeling Candidates Ranking Description Generation Candidate Suggestions Opinion matters! Personalization is useful!
- 3. Overview of Our Methods 3 Candidate Suggestion Crawling Example Suggestions Suggestion Modeling User Profile Modeling Candidates Ranking Description Generation Candidate Suggestions
- 4. Crawling Suggestion Candidates • Source : Yelp • Strategy : At most 100 pages per top category (arts, shopping, food and etc.) • Total number of crawled suggestions : 105,871 – Average number of suggestions per context : 2,117 – Max: 8410 (i.e., Washington D.C.) – Min: 302 (i.e., Crestview) 4
- 5. Overview of Our Methods 5 Candidate Suggestion Crawling Example Suggestions Suggestion Modeling User Profile Modeling Candidates Ranking Description Generation Candidate Suggestions Peilin Yang and Hui Fang. Opinion-based User Profile Modeling for Contextual Suggestions. In Proceedings of the 4th International Conference on the Theory of Information Retrieval, 2013. (ICTIR’13)
- 6. Central Park The MET Sushi Yasuda Angel's Share New York City ? User Profile Modeling Profile Places from Philadelphia 7 A Motivating Example Positive Negative Operison Hotel Liberty Bell
- 7. Description-based Profile Modeling Central Park The MET Sushi Yasuda Angel's Share New York City Profile Ranker Central Park is a public park … … remarkable food in unique environment … largest art museum in the US … … Jazz music … 1. Liberty Bell 2. Operison Hotel America's most historic areas … Downtown public art circuit tour .. At Operison the focus is on detail - and the guest is always at the center of attention. Can not be generalized! 8 Positive Negative Candidate suggestion s
- 8. Central Park The MET Sushi Yasuda Angel's Share New York City Profile Ranker Landmarks Japanese Restaurant Art Museum Night Bar Landmarks Hotel 1. Liberty Bell 2. Operison Hotel 9 Category-based Profile Modeling Positive Negative Still not quite right Candidate suggestion s
- 9. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 10
- 10. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 11
- 11. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 12 Positive Negative User Profile
- 12. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 13 Positive Negative User Profile
- 13. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 14 Positive Negative User Profile
- 14. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 15 Positive Negative User Profile
- 15. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 16 Positive Negative User Profile
- 16. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 17 Positive Negative User Profile
- 17. From “What” to “Why” Central Park The MET Sushi Yasuda Angel's Share New York City 18 Positive Negative User Profile
- 18. From “What” to “Why” New York City Assumption: A user’s profile is constructed based on reviews of other users who share the similar opinions on the example suggestions. 19 Central Park The MET Sushi Yasuda Angel's Share Positive Negative User Profile
- 19. Central Park The MET Sushi Yasuda Angel's Share New York City Profile Ranker … in the center downtown of NYC … Japanese inspired decor, dim lighting, and a clean setting … … it is fantastic but some areas are crowd … The cocktails are very well crafted … A little bit far away from downtown… … it is crowd and you need to take bus to there … … The hotel is very close to the train station … The neat and clean environment is desirable… 1. Operison Hotel 2. Liberty Bell 20 Opinion-based Profile Modeling Positive Negative
- 20. Representation of User Profiles Unique Full Reviews Unique terms from the original review excluding stop words Review Summaries The review summaries generated by Opinosis [1]. 21 … From the stunning architecture to the croissant and latte served up in the food court downstairs. Go to this place and ask why all train stations can't be like this! Wow, over 100 tracks. Unbelievable architecture. Shopping, food. Etc. it is amazing. We ate at the oyster bar last time and that was a treat. The oyster pots are quite something. Original review K. Ganesan, C. Zhai, and J. Han. Opinosis: a graph-based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd International Conference on Computational Linguistics, COLING ’10, pages 340–348, Stroudsburg, PA, USA, 2010. Association for Computational Linguistics.
- 21. Ranking candidate suggestions 22 ),(),( pospos CSUSIMCSUS ×= α Why the user dislikes example suggestions? Why the user likes example suggestions? Candidate Suggestion Why other users dislike candidate suggestions? Why other users like candidate suggestions? Positive profile Negative profile Positive reviews Negative reviews
- 22. 23 Preliminary Results on last year’s data: Opinion-based methods are more effective.
- 23. Overview of Our Methods 24 Candidate Suggestion Crawling Example Suggestions Suggestion Modeling User Profile Modeling Candidates Ranking Description Generation Candidate Suggestions
- 24. Personalized Description Generation • Opening Sentence • “Official” Introduction • Highlighted Reviews • Concluding Sentence 25 What is this place? Why do other people like it? Why is it recommended for YOU? Example suggestions Web site Category Reviews User Candidate suggestion
- 25. • Opening Sentence • “Official” Introduction • Highlighted Reviews • Concluding Sentence 26 What is this place? Example suggestions Web site Category Reviews User Candidate suggestion Personalized Description Generation
- 26. • Opening Sentence • “Official” Introduction • Highlighted Reviews • Concluding Sentence 27 Why do other people like it? Example suggestions Web site Category Reviews User Candidate suggestion Personalized Description Generation
- 27. • Opening Sentence • “Official” Introduction • Highlighted Reviews • Concluding Sentence 28 Why is it recommended for YOU? Example suggestions Web site Category Reviews User Candidate suggestion Personalized Description Generation
- 28. "The Olive Room is a bar. HERE ARE THE DESCRIPTIONS FROM ITS WEBSITE: Here at the olive room, you will receive the finest cuisine montgomery has to offer. HERE ARE REVIEWS FROM OTHER PEOPLE: If you are looking for a unique dining experience, with excellent food, service, location, and outstanding ambiance, look no further! THIS PLACE IS SIMILAR TO OTHER PLACE(S) YOU LIKED, i.e. Tria Wine Room." 29 An Example of Generated Description What is this place? Why do other people like it? Why is it recommended for YOU?
- 29. Description of Our Two Runs 30 Runs User Profile Description UDInfoCS1 Review Summaries Opening Sentence + Meta Description + Web Site Sentences + Highlighted Reviews + Concluding Sentence UDInfoCS2 Unique Full Review Opening Sentence + Meta Description + Highlighted Reviews + Concluding Sentence
- 30. Effectiveness of the runs (from the CS overview paper) 31
- 31. Effectiveness of description generation 34 UDInfoCS1 UDInfoCS2 Accuracy 0.803 0.811 Precision 0.904 0.902 Recall 0.808 0.821 One observation regarding relevance assessment: Among the 569 suggestions returned by both runs, 27.59% (157) of them have inconsistent relevance labels for their websites, and 12.13% (69) of them have inconsistent relevance status.