Yahoo! Answers HackU 2012
- 1. Pramit Roy IIT Delhi, August 2012
- 2. Community Question Answering (CQA) “Just because Google exists does not mean you should stop asking things”, A. Totsis, TechCrunch, 2010 Home > All Categories > Arts & Humanities > Philosophy
- 3. • Largest Community driven question/answer site with more than 1B answers ( http://bit.ly/am1LwL ) • 4th most popular property on yahoo network • Average Dynamics: 3.5 answers/question, 11 minutes to first answer • Community Moderated, no editorial control • Supports 12 languages • Model: – Content-driven (different from Quora , which is social-driven) – Subjective quality model (different from WikiAnswers or StackOverflow) • By default broadcast questions to *all* potential users, not to a group of specific focus • Asker picks best answer (subjectivity rules): “Quality in the eye of the asker” even if the community disagrees
- 4. • User Interest Analysis – Seekers – Knowledge Experts • Sentiment Analysis
- 10. • Problem: About 15% of incoming questions are unanswered • Key observation: Certain questions are recurrent and prevalent – What are the symptoms of cervical cancer? – How much weight Should I lose? – How do you train your cat to use the litter box? – What is a black hole? • Possible Solution: use past answered questions to automatically answer new questions.
- 11. Title: • Non-informative, ambiguous title How often should i jump? • Complex information need Body: I have a 6yo cob, and we've started • Detailed and personal jumping, he's done it before, we're only jumping around 1-1,8 foot. He isn't fat and he's quite fit. I jumped him on • Multiple questions posed in one Friday, but did just 3 small jumps, i did flatwork on Saturday, and jumped him 3 times today then hacked out, tomorrow, • Grammatical errors, slang should i just flatwork him or would he be ok to jump a few small singles? Like 3? I won't ride him again until • Non factual, rather opinion and Wednesday/Thursday to give him a recommendation is expected break, what do you think? He is fine doing what he has done, he doesn't get sweaty much or tired, i myself am not • Extensive variability among ready to go any higher and would like to just do maybe 1 or 2 jumps? Is he ok to questions: in language style, jump tomorrow? If not when next? Thank cultural aspects, degree of detail you for the help! I appreciate it all! :) xx
- 14. • Users want to answer new questions – No social information on such questions (item cold-start) – This known scenario in recommendation systems is our typical case • Many users are new – Hardly any answering history (user cold-start) – The majority of registered users • Current solution: show most recent questions in the category • Goal: a question recommendation model that fits all user types – Active users, – new users, – surfers
- 15. • Recommender system approach, (like for movie recommendation but in a much larger and very sparse space) • Learn from past interactions with users and push relevant open questions I want to answer, who has a question? Yahoo! Answers Recommender System ” - G. Dror, Y. Koren, Y. Maarek and I. Szpektor, KDD’2011, San-Diego, CA
- 17. • Why do people ask questions on Answers? – Users more likely to register with Y!A to ask things they can‘t ask elsewhere (“conversational”, ”personal”, “opinion based”) – When you can’t find what you are looking for by search • Different (demographic) groups use it differently – Women ask more conversational questions – Older people ask more informational questions – Celebrities or professionals ask question to spread awareness, get survey results etc
- 18. Credits: Giovanni Gardelli, Ingmar Weber, Antti Ukkonen, B. Barla Cambazoglu – Y! Research Barcelona Onur Kucuktunc, The Ohio State University, US Hakan Ferhatosmanoglu, Bilkent University, Turkey
- 19. • Relating answers activity to activities around the world. – Huge drop in science & math category in mid- December every year. This category is heavily used for homework help -- effect of Christmas break? • Android/iOS apps. • ~ 35 small and big hacks during the 2011 Y! hack day event @ blr.
- 20. • Person Genome – Everyone has a set of interests. Good contributors in Answers usually ask or answer most seriously in certain specific topics, categories. answers.yahoo.com/activity/qa?show=c8BPMeHBaa&link=question
- 22. • Create granular profile of users – is_a_ps3_power_user, _is_hair_style_conscious etc • Create a graph of profiles – Graph of related users Group of users with related interest • Create question profile and match with user profile Recommend question for users
- 23. • We always have some song we want lyrics for. – How can you ask that in Answers? – Post a link to some streaming URL or YouTube video – Replier will have to go to that page and come back for answering Embed the image and video directly in Answers page for direct preview Add gallery of images to support where asker can create poll (about which shoe to buy, which hair style etc.)
- 24. Answers API •developer.yahoo.com/answers/ •Provides capabilities of searching and reading questions and answers
- 26. • Yahoo Research Labs – Haifa, Israel. Gideon Dror, Yehuda Koren, Yoelle Maarek, Dan Pelleg, Idan Szpektor, Oleg Rokhlenko • Yahoo! Research Labs, Barcelona Giovanni Gardelli, Ingmar Weber, Antti Ukkonen, B. Barla Cambazoglu
Editor's Notes
- #3: Y! A still the largest community question answering site with 1B answers – less “fashionable” maybe than newcomers like Quora or less specialized than StackOverflow but does achieve its goal which is to satisfy askers whatever their intent is – side effect perceived “poor quality” – our approach since the quality is in the eye of the asker – simply don’t show to potential askers questions that don’t resonate with him – each set of questions has its own community – and quality is totally subjective Yahoo gets ~ 600 mil uu per month