Enabling Opinion Driven Decision Making - Kavita Ganesan, GitHub
Download as PPTX, PDF1 like97 views
The document discusses a system called FindiLike that leverages user reviews to enable preference-based decision making. It describes FindiLike's search and summarization components. The search component recommends relevant entities based on user preferences and reviews. Key challenges include uneven review lengths, matching similar concepts, and handling negations. The summarization component generates natural language opinion summaries using a graph-based sentence compression approach to capture redundancies in an ad hoc manner. The summarization technique aims to empower consumers and help users make faster decisions.
Enabling Opinion Driven Decision Making - Kavita Ganesan, GitHub
- 1. Enabling Opinion-Driven Decision Making Kavita Ganesan Senior Data Scientist, GitHub Inc. @kav_gan #Activate18 #ActivateSearch
- 2. Agenda • Problem Intro • Demo of FindiLike – System leveraging user reviews • FindiLike - Search component • FindiLike - Summarization component
- 3. What people think? INFLUENCES DECISION MAKING
- 4. What companies to work for?
- 5. Which cool places to visit?
- 6. What data science tools to use?
- 7. Before Web became popular – life was simple Friends and relatives Consumer Reports
- 8. Crazy WWW What people think? - e-commerce sites - online directories - blog articles - review sites - social networking sites …more
- 9. TOO MUCH INFORMATION TOO FEW TOOLS FOR DECISION MAKING
- 10. Looking for a hotel in an unknown city
- 11. Looking for a hotel in an unknown city 1 Safe ! - No bullets
- 12. Looking for a hotel in an unknown city 1 Safe ! - No bullets 2 Spacious room - Claustrophobic
- 13. Looking for a hotel in an unknown city 1 Safe ! - No bullets 2 Spacious room - Claustrophobic 3 Tempurpedic beds
- 14. Can we currently find hotels matching these criteria? 1 Safe ! - No bullets 2 Spacious room - Claustrophobic 3 Tempurpedic beds
- 15. Maybe by reading 2000 reviews?
- 16. Information is there – somewhere Read reviews & perform own data mining
- 17. FindiLike – Preference Based Hotel Search • Research prototype • Showcase algorithms leveraging user reviews
- 18. Different tools to help with decision making based on opinions
- 19. Can we use this in other domains?
- 21. Healthcare provider search Family Doctor: "speaks fluent spanish" "good with kids" "calming clinic"
- 22. How to implement such a system?
- 23. 1. Preference Based Search How to recommend relevant entities using user reviews and a users’ unstructured preference?
- 24. 2. Query Understanding "good for kids" == "child friendly" ? "baggy" == "oversized" ?
- 25. 3. Summarization How do you provide concise views of user reviews?
- 26. 4. Combine structured + unstructured info How do you combine structured + unstructured info to provide a powerful search experience?
- 27. Let's zoom into two components Preference Based Search Summarization Query Understanding Structured + Unstrcutred Search
- 29. Click to add text Query (short) Document <content> Document <content> Document <content> Traditional Search Click to add text Query (Preferences) Entity <content> Entity <content> Entity <content> Preference Based Search
- 30. Each entity represented by entity document Hotel Los Angeles (Entity) Vagabond Inn (Entity) Holiday Inn Los Angeles (Entity) Entity Document (content) review1: this hotel was absolutely... review2: nice pet friendly hotel... Entity Document (content) review1: extremely old and dirty... review2: nice little place although... Entity Document (content) review1: the continental breakfast was... review2: I will never stay at this place..
- 31. Use regular search technologies to index Entity Document review1: this hotel was absolutely... review2: nice pet friendly hotel... Entity Document review1: extremely old and dirty... review2: nice little place although... Entity Document review1: the continental breakfast was... review2: I will never stay at this place.. Hotel Los Angeles (Entity) Vagabond Inn (Entity) Holiday Inn Los Angeles (Entity)
- 32. more keyword matches = better entity ranks Entity Document Review 1: I loved that this hotel was pet friendly... Review 2: nice pet friendly hotel... Review 3: pets are welcome at this hotel... Review 4: I brought my puppy along since this was... Pet friendlyPreference / query: Hotel Los Angeles
- 33. #1: Uneven Entity Document Length extremel Entity Document (Hotel A) Review 1: This hotel was absolutely... Review 2: Nice pet friendly hotel... Review 3: Beautiful hotel with friendly... Review 4: My pets were really happy... Review 5: Our stay was fantastic. We... ...................................... ...................................... ...................................... ...................................... ...................................... ...................................... Review 1000: Unfortunately, we had a bad... Entity Document (Hotel B) Review 1: We had muffis for bfast... Review 2: Dull hotel with very little... ...................................... ...................................... ...................................... Review 10: Unfortunately, we had a bad... more matches ==> relevance goes up ↑
- 34. #1: Uneven Entity Document Length Solution: similarity models that have saturation points
- 35. #2: Matching similar concepts
- 36. #2: Matching similar concepts extremel pet friendly Entity Document Review 1: I loved that this hotel was pet friendly... Review 2: nice dog friendly hotel... Review 3: pets are welcome at this hotel... Review 4: Brought my kitty along since this is an animal friendly facility... Preference / query:
- 37. #2: Matching similar concepts extremel pet friendly Entity Document Review 1: I loved that this hotel was pet friendly... Review 2: nice dog friendly hotel... Review 3: pets are welcome at this hotel... Review 4: Brought my kitty along since this is an animal friendly facility... Preference / query: Expand query with related concepts
- 38. #3: Dealing with negations
- 39. #3: Dealing with negations extremel pet friendly This hotel was NOT pet friendly so I could not bring my dog Preference / query:
- 40. #3: Dealing with negations extremel pet friendly This hotel was NOT pet friendly so I could not bring my dog Preference / query: Nearby negations should be considered
- 41. Opinion-Based Entity Ranking – Ganesan & Zhai 2012 http://kavita-ganesan.com/opinion-based-entity-ranking/
- 42. Let's zoom into summarization of opinions Preference Based Search Summarization Query Understanding Structured + Unstrcutred Search
- 43. What can review summaries provide that search cannot?
- 44. Unexpected Information That can make or break the deal
- 46. Deal breaker!
- 47. Lots of summarization options Concise or aggregated view of what people are saying Cleanliness Comfort Staff Views 4.0/5 2.0/5 3.0/5 4.5/5
- 48. Natural Language Opinion Summary • Aggregation of what people think • Concise, conveys essential info • Analytics
- 49. Opinosis: Graph Based Sentence Compression Approach (Ganesan & Zhai Coling 2010)
- 50. Sentence Compression Approach “The bed was comfortable, I loved it!” “The bed was really comfortable..” “The tempurpedic bed was super comfy”
- 51. Sentence Compression Approach “The bed was comfortable, I loved it!” “The bed was really comfortable..” “The tempurpedic bed was super comfy” bed was comfortable (3)
- 52. Opinosis: Uses a Word Graph “The bed was comfortable, I loved it!” “The bed was really comfortable..” “The tempurpedic bed was super comfy” bed I really comfortable . tempurpedic the , loved itwas super . WORD GRAPH
- 53. What's nice about this technique? 1. Really speedy for ad hoc summarization • No need to pre-compute • Generate summaries for specific topics
- 54. What's nice about this technique? 1. Really speedy for ad hoc summarization 2. Captures redundancies naturally • Automatically provides the analytics component
- 55. What's nice about this technique? 1. Really speedy for ad hoc summarization 2. Captures redundancies naturally 3. Extract phrases from existing sentence structure • Limiting external dependencies
- 56. Try out Web API or JAR File www. kavita-ganesan.com/opinosis/
- 57. In summary…
- 59. Leverage text data to empower consumers…
- 60. Help users make faster decisions...
- 61. Increase conversion rates & loyalty
- 62. Thank you! Kavita Ganesan @kav_gan kavita-ganesan.com #Activate18 #ActivateSearch