Learning to Generate Diversified Query Interpretations using Biconvex Optimization

International Joint Conference on Natural Language Processing, pages 733–739, Nagoya, Japan, 14-18 October 2013. Learning to Generate Diversified Query Interpretations using Biconvex Optimization Ramakrishna B Bairi IITB-Monash Research Academy IIT Bombay Mumbai, India, 400076 bairi@cse.iitb.ac.in Ambha A IIT Bombay Mumbai, India, 400076 ambha.career @gmail.com Ganesh Ramakrishnan IIT Bombay Mumbai, India, 400076 ganesh@cse.iitb.ac.in Abstract The wealth of information present in the World Wide Web has made internet search a de-facto medium for obtaining any re- quired information. Users typically spec- ify short and/or ambiguous queries and expect the answer to appear at the top. Hence, it can be extremely important to produce a diverse but relevant set of results in the precious top k positions. This calls for addressing two types of needs: (i) pro- ducing relevant results for queries that are often short and ambiguous and (ii) select- ing a set of k diverse results to satisfy dif- ferent classes of information needs. In this paper, we present a novel technique using a Biconvex optimization formulation as well as adaptations of existing techniques from other areas, for addressing these two problems simultaneously. We propose a graph based iterative method to choose di- versified results. We evaluate these ap- proaches on the QRU (Query Represen- tation and Understanding) dataset used in SIGIR 2011 workshop as well as on the AMBIENT (Ambiguous Entities) dataset and present results on generating diversi- fied query interpretations. We also com- pare these approaches against other on- line systems such as Surf Canyon, Car- rot2, Exalead and DBpedia and empir- ically demonstrate that our system pro- duces competitive results. 1 Introduction The growth of internet has resulted in the prolifer- ation of electronic documents on the World Wide Web. Every search engine, be it generic or applica- tion and domain specific, serves as a portal to ac- cess these documents. User queries, in general, are short and often tend to be ambiguous and/or under- specified. In addition, a query can have multi- ple concealed interpretations. For example, Sun could be interpreted as “The sun as a star”, “Com- position of Sun”, “Sun Micro systems company”, “Sun news paper”, “Sun Record music company”, and so on. We believe that, in addition to these concealed interpretations, related interpretations are also equally important. As examples, “Solar Cells” and “Photosynthesis”, could be interpreta- tions related to this query. To improve user in- teraction and to guide him/her in further refining the query, it could help if the search engine gener- ated these relevant interpretations as well. Due to the sheer size of online information and its diver- sity, the possible interpretations to a short query are enormous. In addition, users expect their in- tended answer to be present in the top few search results. This calls for presenting a diversified but relevant set of results in the top k positions. Note that, in this paper we consider the diverse search results produced by the search system as interpre- tations of the query in some sense. In addition, we consider each search …

Content truncated to 500 words