Visualization and user-interface techniques for interactive information retrieval systems

In this dissertation, we have designed and evaluated a graphical display for Information Retrieval (IR) systems which improves the accuracy of identifying relevant information. By utilizing the excellent capability of humans to recognize visual patterns, it also increases the speed of identifying relevant documents. The display design is motivated by an examination of the cognitive processes that influence users in their information retrieval activity. This examination points to the need for certain properties in the display that could help users in different aspects of their IR activity. We focus on the following three aspects in the context of ranked-output IR systems. First, users often try to modify their free-form query in order to obtain a better ranking of documents relevant to their query. To enable users to construct effective queries, we examine why systems should explicitly display the causal relationships between the query constructs and the ranking of documents. We design a graphical display tool by examining what information could be displayed and how it could be displayed in order to explicitly show the causal relationship. Second, users inspect the titles of documents retrieved by the system before requesting their full texts. In addition to titles, examining what information could be displayed and how it could be displayed to enable users to quickly and accurately identify relevant documents shows that the same graphical display tool is useful in this situation. Third, users employ a variety of information sources to identify relevant information. We examine simple user interface techniques provide seamless access to various information sources including a thesaurus. The graphical interface design thus allows two-way feedback: from system to user and from user to system. Our choice of graphical display is validated by user experiments. The user experiments empirically show the superiority of the proposed display in increasing the speed and accuracy of identifying relevant documents. These experiments show that improvements in interactive precision and interactive recall are statistically significant. The reduction in time taken to judge document relevance is also statistically significant. Further, the differences are large enough to be considered practically significant. We also present the results of structured interviews with subjects and analysis of their think-aloud sessions that suggest possible improvements to the system.