Computational User Intent Modeling Hongning Wang (wang296@illinois.edu) Advisor: ChengXiang Zhai (czhai@illinois.edu) Department of Computer Science, University of Illinois at Urbana-Champaign Urbana IL, 61801 USA Joint Relevance and Freshness Learning (WWW’ 2012) Content-Aware Click Modeling (WWW’2013) Cross-Session Search Task Extraction (WWW’2013) Unsupervised Discovery of Opposing Opinion Networks (CIKM’2012) In contrast to traditional Web search, where topical relevance is often the main ranking criterion, news search is characterized by the increased importance of freshness. However, the estimation of relevance and freshness, and especially the relative importance of these two aspects, are highly specific to the query and the time when the query was issued. In this work, we proposed a unified framework for modeling the topical relevance and freshness, as well as their relative importance, based on click logs. We explored click statistics and content analysis techniques to define a set of temporal features, which predict the right mix of freshness and relevance for a given query. In this work, we proposed a general Bayesian Sequential State (BSS) model for addressing two deficiencies of existing click modeling approaches, namely failing to utilize document content information for modeling clicks and not being optimized for distinguishing the relative order of relevance among the candidate documents. As our solution, a set of descriptive features and ranking-oriented pairwise preference are encoded via a probabilistic graphical model, where the dependency relations among a document's relevance quality, examine and click events under a given query are automatically captured from the data. Search tasks frequently span multiple sessions, and thus developing methods to extract these tasks from historic data is central to understanding longitudinal search behaviors and in developing search systems to support users' long-running tasks. In this work, we developed a semi-supervised clustering model based on the latent structural SVM framework, which is capable of learning inter-query dependencies from users' searching behaviors. A set of effective automatic annotation rules are proposed as weak supervision to release the burden of manual annotation. Our method paves the way for user modeling and long-term task based personalized applications. With more and more people freely express opinions as well as actively interact with each other in discussion threads, online forums are becoming a gold mine with rich information about people’s opinions and social behaviors. In this work, we study an interesting new problem of automatically discovering opposing opinion networks of users from forum discussions, which are subset of users who are strongly against each other on some topic. Signals from both textual content (e.g., who says what) and social interactions (e.g., who talks to whom) are explored in an unsupervised optimization framework. Relevance v.s. Freshness Modeling User Clicks Semi-supervised Structural Learning Identifying Opposing Opinion Networks • An atomic information need that may result in one or more queries • Relevance • Topical relatedness • Metric: tf*idf, BM25, Language Model • Freshness • Temporal closeness • Metric: age, elapsed time • Trade-off • Query specific • To meet user’s information need Match my query? An impression Thread, e.g. “health care reform” Reply To … Supporting Group Post It’s human right! User tѱ = 30 minutes Redundant doc? Shall I move on? Joint Relevance and Freshness Learning Query => trade-off Key: Freshness v.s. Relevance Chance to further examine the results: e.g., position, # clicks, distance to last click Chance to click on an examined and relevant document: e.g., clicked/skipped content similarity Click => overall impression URL => freshness URL => relevance Experimental Results 1. P@1 comparison between different click models over the random bucket click set and normal click set from Yahoo! news search log. Budget increase It is nonsense! 5/29/2012 S1 5/29/2012 5:26 bank of america 5/29/2012 S2 5/29/2012 11:11 macy's sale 5/29/2012 11:12 sas shoes 5/30/2012 S1 5/30/2012 10:19 credit union 5/30/2012 S2 5/30/2012 12:25 6pm.com 5/30/2012 12:49 coupon for 6pm shoes I insist my point. I agree with you! … Signal 1: ReplyTo Text (R: agree/disagree) Hot Topics & Current Events forum in Military.com: •43,483 threads •1,343,427 posts •34,332 users •7.7 reply-to relation/ thread Signal 2: Author Consistency (A) Signal 3: Topical Similarity (T: agree/disagree) Heuristic constraints • Identical queries • Sub-queries • Identical clicked URLs Structural knowledge • Same task => tasks sharing related queries • Latent Relevance quality of a document: e.g., ranking features Experimental Results Experimental Results Sentiment prior Opinions Agree Opinion of posts Text 1 Text 2 Text 3 … … v1 v2 v3 Opinions Disagree subject to 1. Task extraction performance on Bing web search log with increasing volume of weak supervision. Experimental Results 1. Model update trace in training process. 1. Accuracy of Agree/Disagree relation classification. 2. Identified latent search task structure. 2. Ranking performance comparison with baselines on Yahoo! news search log. (a) On normal bucket clicks Against Group (b) On random bucket clicks 2. Feature weights learned by BSS model. 2. Accuracy of user opinion prediction.