Q&A over Social Networks KSE 801 Uichin Lee Aardvark: The Anatomy of a LargeScale Social Search Engine Damon Horowitz, Aardvark Sepandar D. Kamvar, Stanford University WWW '10 Original slides by Hailong Sun, April 13, 2010 Web Search vs. Social Q&A • Google is about traditional Web search – Give me keywords, I will provide contents – Search for the most suitable contents • While Aardvark is about social search – Users can ask questions in natural language, not keywords – Content is generated “on-demand”, tapping the huge amount of information in peoples’ heads (i.e., everyone knows something) – The system is fueled by the goodwill of its users Problem in Aardvark Search Engine • How to find the user who can best answer a given question? Aardvark Architecture 2 1 Question? 3 Edit question? Determine the appropriate topics for the question 3 (Routing Suggestion Request) -- find a list of candidate answerers (and rank them) 4 Ask one by one Social graph indexing User’s topic parsing VarkRank: Relevance and Connectedness Question User User User Topic User User User P(ui|t) Used for measuring relevance score of a user’s question (query dependent) User User User P(ui|uj) Used for measuring connectedness score between users (query independent) VarkRank: Relevance and Connectedness • Given a question q, the probability that a user ui can answer it (relevance score) • Score that user ui can answer a question from uj: (query dependent user’s query relevance score * query independent user connectedness score) – Query independent user quality score = p(ui|uj): i.e., user i delivers a satisfying answer to user j (simply measured using connectedness) Relevance Scores (Expertise) • How to find experts on a given topic? • Expertise score: p(ui|t), and w/ Bayes’ law, we have: • For each user, we need to profile a user’s interest in a given topic by using the following information sources: – – – – – 3+ topics provided by a user Topics provided by friends of a user Online profiles Online unstructured data Status message updates 9 Connectedness Scores • Connection strengths between people i.e., p(ui|uj) are computed using a weighted cosine similarity over this feature set (normalized) • Utilize existing social networks – Facebook, Twitter, LinkedIn… • Feature set measures similarities in demographics/behavior – – – – – – – – Social connection (common friends and affiliations) Demographic similarity Profile similarity (e.g., common favorite movies) Vocabulary match (e.g., IM shortcuts) Chattiness match (frequency of follow-up messages) Verbosity match (the average length of messages) Politeness match (e.g., use of “Thanks!”) Speed match (responsiveness to other users) Analyzing Questions: TopicMapper • Question classification: – – – – Question or not? Inappropriate question? Trivial question? Location sensitive question? • Map a question to a topic (weighted linear sum of the following features) – Keyword matches w/ a user’s profile topics? – Classifies the question text into a taxonomy of roughly 3000 popular question topics (using an SVM trained on an annotated corpus of several million questions) – Extracting salient phrases from questions and find semantically similar user topics Ranking Algorithm • Topic and connectedness matching availability • Routing Engine prioritizes candidate answerers – Optimize the chances that the present question will be answered – Yet, preserving the available set of answerers (i.e., the quantity of “answering resource” in the system) as much as possible by spreading out the answering load across the user base • Considering factors: currently online users (e.g., via IM presence data, iPhone usage, etc.), user’s daily activity history, and user’s response history (lowering scores of non-responsive users) • Conversation Manager serially inquiring whether candidates would like to answer the present question; and iterating until an answer is provided and returned to the asker. User Interface IM Email Twitter Aardvark Deployment Status • Aardvark is actively used – Users: from 2,272 to 90,361 – 55.9% active users, 73.8% passive users – 3,167.2 questions/day – 3.1 queries/month • Mobile users are particularly active – – – – Average 3.6322 sessions/month Comparison: Google Desktop v.s. mobile users: 3 Mobile users: 5.68 sessions/month Categories of Questions in Aardvark • Questions are highly contextualized – Average query length: 18.6 words • 2.2~2.9 for Web search – 45.3% are about context • Questions often have a subjective element – What are the things/crafts/toys your children have made that made them really proud of themselves? Answers • Answers – 57.2% received answers in less than 10 minutes – A question receives 2 answers averagely – The quality of answers are good • 70.4% are “good”; 14.1% are “OK”; 15.5% are “bad” Distribution of questions and answering times Distribution of questions and number of answers received Topic Distribution • People are indexable – 97.7% have 3+ topics 17 Comparative Evaluation with Google • Experimental setup – Randomly select a group of questions – Insert a tip “do you want to help Aardvark run an experiment?” – Recording response time and quality of answers from Google and Aardvark • Experimental results – Aardvark: 71.5% answered; rating: 3.93 (σ=1.23) – Google: 70.5% answered; rating: 3.07 (σ=1.46) • Aardvark is more suitable for subjective questions What Do People Ask Social Networks? Meredith Ringel Morris, MSR Jaime Teevan, MSR Katrina Panovich, MIT http://www.yelp.com good restaurants in atlanta Questions About People’s Questions • What questions do people ask? – How are the questions phrased? – What are the question types and topics? – Who asks which questions and why? • Which questions get answered? – How is answer speed and utility perceived? – What are people’s motivations for answering? What Is Known About Question Asking • • • • Collaborative search [Morris & Teevan] Searching v. asking [Evans et al.; Morris et al.] Expertise-finding [vark.com; White et al.; Bernstein et al.] Online question answering (Q&A) tools – Question type [Harper et al.: conversational v. informational] – Response rate [Hseih & Counts: 80%] – Response time [Zhang et al.: 9 hours; Hseih & Counts: 3 hours] – Motivation [Raban & Harper; Ackerman & Palen; Beenan et al.] Survey of Asking via Status Messages • Survey content – Used a status message to ask a question? • Frequency of asking, question type, responses received • Provide an example – Answered a status message question? • Why or why not? • Provide an example • 624 participants – Focus on Facebook and Twitter behavior Questions About People’s Questions • What questions do people ask? – How are the questions phrased? – What are the question types and topics? – Who asks which questions and why? • Which questions get answered? – How is answer speed and utility perceived? – What are people’s motivations for answering? Questions: Phrasing • Questions short (75 characters, 1 sentence) • 18.5% of phrased as a statement I need a recommendation on a good all purpose pair of sandals. • Often scoped – 1 out of 5 directed to “anyone” Anyone know of a good Windows 6 mobile phone that won’t break the bank? – Network subset Hey Seattle tweeps: Feel like karaoke on the Eastside tonight? Questions: Types Type % Example Recommendation 29% Building a new playlist – any ideas for good running songs? Opinion 22% I am wondering if I should buy the Kitchen-Aid ice cream maker? Factual 17% Anyone know a way to put Excel charts into LaTeX? Rhetorical 14% Why are men so stupid? Invitation 9% Who wants to go to Navya Lounge this evening? Favor 4% Need a babysitter in a big way tonight… anyone?? Social connection 3% I am hiring in my team. Do you know anyone who would be interested? Offer 1% Could any of my friends use boys size 4 jeans? Questions: Topics Topic % Example Technology 29% Anyone know if WOW works on Windows 7? Entertainment 17% Was seeing Up in the theater worth the money? Home & Family 12% So what’s the going rate for the tooth fairy? Professional 11% Which university is better for Masters? Cornell or Georgia Tech? Places 8% Planning a trip to Whistler in the off-season. Recommendation on sites to see? Restaurants 6% Hanging in Ballard tonight. Dinner recs? Current events 5% What is your opinion on the recent proposition that was passed in California? Shopping 5% What’s a good Mother’s Day gift? Philosophy 2% What would you do if you had a week to live? Questions: Who Asks What Type Recommendation Topic old men Opinion Twitter Entertainment women Factual Rhetorical Invitation Technology Home & Family Professional young Places Favor Restaurants Social connection Current events Offer Shopping Philosophy Facebook Questions: Motives for Asking Topic % Example Trust 24.8% I trust my friends more than I trust strangers. Subjective 21.5% Search engine can provide data but not an opinion. Thinks search would fail 15.2% I’m pretty search engine couldn’t answer a question of that nature. Audience 14.9% Friends with kids, first hand real experience. Connect 12.4% I wanted my friends to know I was asking the question. Speed 6.6% Quick response time, no formalities. Context 5.4% Friends know my tastes. Tried search 5.4% I tried searching and didn’t get good results. Easy 5.4% Didn’t want to look through multiple search results. Quality 4.1% Human-vetted responses. Questions About People’s Questions • What questions do people ask? – How are the questions phrased? – What are the question types and topics? – Who asks which questions and why? • Which questions get answered? – How is answer speed and utility perceived? – What are people’s motivations for answering? Answers: Speed and Utility • 94% of questions received an answer • Answer speed – A quarter in 30 minutes, almost all in a day – People expected faster, but satisfied with speed – Shorter questions got more useful responses • Answer utility – 69% of responses helpful Answers: Speed and Utility Type Topic Recommendation Opinion Technology Fast Factual Rhetorical Entertainment Home & Family Unhelpful Professional Invitation Places Favor Restaurants Social connection Current events Offer Shopping Philosophy Answers: Motives for Answering Motive % Example Altruism 37.0 Just trying to be helpful. Expertise 31.9 If I’m an expert in the area. Question 15.4 Interest in the topic. Relationship 13.7 If I know and like the person. Connect 13.5 Keeps my network alive. Free time 12.3 Boredome/procrastination. Social capital 10.5 I will get help when I need it myself. Obligation 5.4 A tit-for-tat. Humor 3.7 Thinking I might have a witty response. Ego 3.4 Wish to seem knowledgeable. Answers About People’s Questions • The questions people ask – Short, directed to “anyone” – Subjective questions on acceptable topics – Social relationships important motivators • The questions that get answered – Fast, helpful responses, related to length and type – Answers motivated by altruism and expertise QUESTIONS? Meredith Ringel Morris Jaime Teevan Katrina Panovich M. R. Morris, J. Teevan, and K. Panovich. What Do People Ask Their Social Networks, and Why? A Survey Study of Status Message Q&A Behavior. CHI 2010. M. R. Morris, J. Teevan, and K. Panovich. A Comparison of Information Seeking Using Search Engines and Social Networks. ICWSM 2010 (to appear).