Abduction, Explanation and Relevance Feedback

Selecting good query terms to represent an information need is difficult. The complexity of verbalising an information need can increase when the need is vague, when the document collection is unfamiliar or when the searcher is inexperienced with information retrieval (IR) systems. It is much easier, however, for a user to assess which documents contain relevant information. Relevance feedback (RF) techniques make use of this fact to automatically modify a query representation based on the documents a user considers relevant. RF has proved to be relatively successful at increasing the effectiveness of retrieval systems in certain types of search, and RF techniques have gradually appeared in operational systems and even some Web engines. However, the traditional approaches to RF do not consider the behavioural aspects of information seeking. The standard RF algorithms consider only what documents the user has marked as relevant; they do not consider how the user has assessed relevance. For RF to become an effective support to information seeking it is imperative to develop new models of RF that are capable of incorporating how users make relevance assessments. In this thesis I view RF as a process of explanation. A RF theory should provide an explanation of why a document is relevant to an information need. Such an explanation can be based on how information is used within documents. I use abductive inference to provide a framework for an explanation-based account of RF. Abductive inference is specifically designed as a technique for generating explanations of complex events, and has been widely used in a range of diagnostic systems. Such a framework is capable of producing a set of possible explanations for why a user marked a number of documents relevant at the current search iteration. The choice of which explanation to use is guided by information on how the user has interacted with the system---how many documents they have marked relevant, where in the document ranking the relevant documents occur and the relevance score given to a document by the user. This behavioural information is used to create explanations and to choose which type of explanation is required in the search. The explanation is then used as the basis of a modified query to be submitted to the system. I also investigate how the notion of explanation can be used at the interface to encourage more use of RF by searchers