Leveraging Wikipedia's Article Structure to Build Search Agents: TUW at CLEF 2017 Dynamic Search

Often, single query search sessions are not enough to solve complex problems or to gather sufficient information to take an informed decision. Such complex search tasks include many ordinary tasks such as planning a vacation trip, studying for a school or college exam or gathering information on a symptom or condition. Nevertheless, complex search tasks can be broken into multiple smaller specific subtasks. In order to assist users in dealing with complex searches, a search agent could be employed to automatically break a complex search task into smaller tasks, to issue multiple queries for those subtasks, and to report the results back to the user in a meaningful way. A key problem that the Information Retrieval community aims to solve in order to create such agents is the understanding of complex search tasks, which includes the identification of smaller subtasks. To foster research in such interesting problem a number of challenges have been recently proposed (e.g., [5,4,2]) and this paper describes the efforts of Vienna University of Technology (TUW) in one of such challenges, the first CLEF Dynamic Search [2]. We propose the creation of a search agent that specifically leverage the structure of Wikipedia articles to understand search tasks. Our assumption is that human editors carefully choose meaningful section titles to cover the various aspects of an article. Our proposed search agent explores this fact, being responsible for two tasks: (1) identifying the key Wikipedia articles related to a complex search task, and (2) selecting section titles from those articles. For instance, consider a user seeking information on how to quit smoking. Some of the relevant subtasks, in this case, are the description of different ways to quit smoking, the benefits of quitting smoking and second effects of quitting smoking. A possible query that expresses this information need is simply “quit smoking”. The Wikipedia article Smoking Cessation1 is the top hit for such query