Information Value Theory in Query Augmentation

Information retrieval (IR) systems interact with users by returning a ranked list of relevant documents in response to a query. Through feedback mechanisms such as relevance feedback and automated keyword expansion, IR systems attempt to guide users in constructing search queries which better represent their information needs. These mechanisms, however, do not offer the user more insight into the content of the documents in the IR database nor do they provide direction as to which search terms might yield better search results in terms of relevance and certainty that the retrieved document contains the information the user intended to retrieve. This paper presents a methodology based on the decision-analytic concept of expected value of perfect information for controlling query augmentation in information retrieval. The system dynamically learns the content of the documents in the database to compute the utility (measured in terms of relevance) of retrieving certain documents in response to queries, where the words in the queries represent the random variables. By computing the expected value of perfect information for each query term, the system either suggests new search terms or suggests that the user terminate the search.