Implicit Contextual Modelling for Information Seeking

In this position paper we present an overview of our current work in utilising contextual models to help web searchers find relevant web documents. The traditional approach to such information retrieval (IR) assumes that a user’s information need is static and does not change as they peruse the documents the search system presents to them. However, this approach is simplistic, and does not consider the dynamic nature of an information need, something that has been well documented. In our approach we use implicit evidence, captured unobtrusively from searcher interaction to model such change. We are particularly interested in the intentionality behind this interaction, and which document representations (i.e. summaries, sentences) users view. This evidence is firstly used to enhance the user’s query, then automatically update the display and if required, re-search the web. 1. CONTEXT IN INFORMATION SEEKING Web search systems operate under a simple retrieval paradigm, where a user, with a need for information, motivated by some gap in their current knowledge, seeks the information required to close this gap and hence satisfy their need. Typically, users are expected to express this need via a set of query terms submitted to the search system. This query is then compared to each document in the collection, and a set of potentially relevant documents is returned. However, web queries are typically short, ambiguous, and are often only an approximation to the searcher’s real information need. The retrieval operation is a process of query-document inference, where the query infers relevant documents. In essence, an IR system operates on a ‘quality-in, quality-out principle’, where a query that closely represents the user’s real need increases the likelihood that more of the documents suggested by the system will be relevant. Even a query that is a good approximation to a user’s real need may still lack some of the terms necessary to adequately discern relevant documents from non-relevant documents. The influence of context on this is unquestionable. Novice, or inexperienced, users can have difficulty in formulating queries that are sufficiently indicative of their real information need. The problem is amplified where the user’s need is vague or ‘ill-defined’. In such cases, the user’s inability to even outline the nature of their need can lead to problems in both the retrieval and document assessment processes. For such users, perusing even some of the retrieved document set can lead to marked changes in their understanding and/or knowledge of the topic. Typically, the influence of the user’s contextual predicament and/or awareness when interacting with IR systems is ignored. Instead, these systems model user need based only on the query terms they submit, relying heavily on the user’s ability to formulate such queries. Current IR systems play only a passive role in the development of better defined information needs. In our work we develop systems and use approaches that help users more actively. The ‘interactive revolution’ of the early 1980’s pointed toward the need to address IR systems from an end-user perspective. Interactive Information Retrieval (IIR) systems are defined as those where the user dynamically conducts searching tasks and correspondingly reacts to system responses over session time. Despite this shift of focus toward the user, and on increasing the quality of the IR system ‘input’, it is the traditional approach to IR that many web search systems adopt. The user represents their need in a one-off textual query, submitted to the system, which then processes the query and returns a list of results. This is more commonly known as the ‘black-box’ approach, shown in Figure 1. Fig. 1. The ‘black-box’ approach to information retrieval Such systems are static, and not aware of the context in which they operate; an abstraction necessitated by the mathematical models that underlie them. The query (often combined with statistical information on the document collection) is the sole factor used in determining output, something that is problematic for those who have difficulty expressing their needs, either because they are illdefined or they are unsure of what the system expects from them. Capturing implicit evidence from searcher interaction can provide additional information useful in devising a more complete query. In IIR, context effectively constitutes all factors that influence both the user’s and the system’s role in the information seeking process, but are not explicitly defined. Contextual influence plays a vital role in relevance assessment, but is not considered by the systems intended to assess relevance on the user’s behalf. IR systems use ‘algorithmic relevance’ to determine whether a document pertains to a user’s information need. Such algorithms assume the form of a query-document matching function, usually producing a ranked output of documents, ranked in descending order depending on their ‘relevance’. It has been suggested that neither the system nor the query-document matching function are relevant to the context from which the user directs their query. We can therefore postulate that because their notion of relevance lacks some of the qualities intrinsic for users in determining relevance then the results presented by the system are only an approximation to relevance. Through developing IR systems that take account of the user’s context, the actual relevance of the recommendations presented by these systems can be improved. In general, context-aware computing systems are sensor-based, sometimes wearable devices, working through contextual feedback from the environment in which they operate. In IIR, context is traditionally represented through the content of active applications and their relation to user’s long term goals and objectives. Typically, no attention is paid to what contextual information user interaction can yield and the use of current goals and search intentions. There is also no guarantee that the information currently being sought will coincide with long term aspirations. Users’ behaviour when searching for information can yield much about the contextual influence (i.e. time constraints, user mood) being imposed upon them. Contextually-aware IIR systems take such influence into account, tailoring their responses to suit the current search context. It is important that such systems, as well as reacting to contextual change, are also proactive, recommending future search alternatives to users. Developing the model that underlies such a system is outlined in what follows in this paper. information retrieval system query results Department of Computer and Information Sciences University of Strathclyde Glasgow, Scotland. G1 1XH Department of Computing Science University of Glasgow Glasgow, Scotland. G12 8QQ 2. IMPLICIT CONTEXTUAL MODELLING In our current work we develop a contextual model for information seeking based around user interaction and information seeking behaviour. This model, once complete could be applied to operational retrieval environments, gathering implicitly, information on the context of the search, analysing and applying this to the benefit of the user.