Exploiting User Search Sessions for the Semantic Categorization of Question-like Informational Search Queries

International Joint Conference on Natural Language Processing, pages 902–906, Nagoya, Japan, 14-18 October 2013. Exploiting User Search Sessions for the Semantic Categorization of Question-like Informational Search Queries Alejandro Figueroa Yahoo! Research Latin America Av. Blanco Encalada 2120, Santiago, Chile afiguero@yahoo-inc.com G¨unter Neumann DFKI GmbH Stuhlsatzenhausweg 3, D-66123 Saarbr¨ucken, Germany neumann@dfki.de Abstract This work proposes to semantically clas- sify question-like search queries (e.g., “oil based heel creams”) based on the con- text yielded by preceding search queries in the same user session. Our novel ap- proach is promising as our initial results show that the classification accuracy im- proved in congruence with the number of previous queries used to model the ques- tion context. 1 Introduction Open question answering (QA), i.e., fully auto- matic systems that find best answers to natural language questions of any type and domain, is still a challenging research problem. On the other hand side, search engines are getting smarter and smarter in order to fulfill users’ information re- quests. This motivates users to enter more sophis- ticated search queries (e.g., more complete ques- tions) rather than few keywords, when they are looking for precise information needs (e.g., an- swers related to precise problems). This is also ex- perienced by the fact that through search engines, it is likely to exploit answer databases of commu- nity based question answering (cQA) systems in- cluding Yahoo! Answers, if the search query is close to a QA-system like question. Then match- ing such a question with those in the cQA database is more likely to recognize plausible cQA para- phrases because of close textual relatedness. Fur- thermore, as the analysis of our data sources sug- gests, users often express semantically related se- ries of questions in order to guide the search for better answers, and as such, are already perform- ing interactions with search engines. In a general sense, searching is a sequence of queries in the same user session aimed at satisfying an underly- ing goal that the user is trying to achieve (Rose and Levinson, 2004). Thus, we believe that it will be inevitable to further automatize a semantic analysis of search queries within user sessions, i.e., to analyze the se- mantic relatedness of a series of questions whether they constitute actually a session of semantically related questions entered by the same user. Our contribution into these directions is the ex- ploration of automatic methods to semantically classify question-like search queries, based on the context provided by preceding search queries in the same user session. An important aspect, tack- led in this paper, is whether and how much con- textual information extracted from user-specific search query sessions helps to effectively train and apply a model to predict the semantic category of a question-like informational search query (cf. (Broder, 2002; Rose and Levinson, 2004)). Our method recognizes question-like queries by inspecting their associations with Yahoo! Answers pages via user clicks, providing the additional ben- efit of linking each query with an entry in the Ya- hoo! Answers category system. …

Content truncated to 500 words