Repetition and diversification in multi-session task oriented search

As the number of documents and the availability of information online grows, so to can the difficulty in sifting through documents to find what we're searching for. Traditional Information Retrieval (IR) systems consider the query as the representation of the user's needs, and as such are limited to the user's ability to describe the information he or she is seeking. Several challenges can arise within this single query-results paradigm. In this model the user is required to know enough about the domain to express his or her information need. Constructing a query can be difficult in it's own right. Without being well versed in the jargon of a new domain, it may be difficult to formulate a query that adequately expresses one's information need. It may also be challenging to understand the contents of a document, let alone how a given document relates to other documents in the field. This fact can make it more challenging for the user to identify the relevant documents out of the search results. Thus learning about new ideas and new topics that one is unfamiliar with can be challenging. In many situations it may not be possible for all the information a user is seeking to be present in a single source, or to be returned within a single query or small set of queries. The knowledge sought by the user may be too broad or too deep. Someone researching medical treatments may not find one single source with enough information to make an adequate decision regarding his or her health. In a literature review, the absence of information on a given idea is itself information. An idea could be so novel that it's never been attempted or discussed. To know that an idea of is novel, however, would require a good understanding of many documents, not just any single source. The user would require a solid grasp on the entire search space, not just the search results from a single query. Finally, the information itself may be changing. As the medical community evaluates new drugs and treatments that might outperform current methods, society's understanding of bioscience grows as does the literature and information available to the searcher. Similarly, new research continues to be published at conferences and in journals, by academics and industry alike. A literature review from a few years ago may be considered incomplete and obsolete today. New information is being built upon past information. Thus a user may need to constantly go back and revisit past documents and information already acquired, even as they continue to seek out new information. In this dissertation we present explore multi-session task oriented search. We begin this dissertation with a study of on-going multi-session search, search that cannot be satisfied within a single query or session. We focus primary on repetitive queries and repeat search result clicks to find patterns in the way individuals search. These actions can be viewed as a simple form of on-going search. When studying repetitive queries we discover that as users continue to issue the same query, they show a desire to return to past content as well as find new information. We find users are selective in their new-findings, and even though users may revisit the same URLs frequently, these new-findings might offer more insight into the user's underlying interest. We then study repeat findings, when a user revisits a webpage following a search result. We find a distinction between intra and cross session re-findings, where the former suggested a possibly difficulty understanding documents and the latter was indicative of picking up a task. We find that a user's understanding of a URL may evolve over time, and that the query used when finding this URL becomes ``better'', in that it is shorter, more commonly used, and ranks the revisited URL higher. From these findings about query reissuing and page revisitation we propose a relationship based framework called ARTEMIS. ARTEMIS stands for Assisted Relationship Tracing for Exploratory Multi-Session Informational Search. ARTEMIS is designed to help users better understand the search space by promoting connections between new URLs that the user may be unfamiliar with to URLs the user has previously visited and shown a strong interest in. By showing how a new document may relate to prior ones the user gains a better understanding of both the content of the new document, and the potential relevance of the document to the user's underlying task. Finally, we demonstrate the potential of ARTEMIS with a proof of concept implementation accompanied by a task oriented user study. The study is conducted using lay people preforming a skilled task, which is an important and challenging subtask of the general search behaviors explored in the previous sections. The results show novice searches have an increased ability to find documents, and to understand them while using ARTEMIS.