Improving Information Retrieval by Consideration of Temporal Evidence in a Unified Model

Time affects information retrieval in many ways. Collections of documents change as new items are indexed. The content of documents themselves may change. Users submit queries at particular moments in time. And perhaps most importantly, people’s assessment of a document’s relevance to a query is often time-dependent. For example, searchers of news archives might seek information on a past event where relevant documents cluster in a window of time. Users of social media services such as Twitter demand topically relevant information that is new. People who monitor particular topics in the news (for example, editors of Wikipedia) take action when they find information that is topically relevant and that changes current knowledge. The traces of information created by change in documents, collections, and language improve our ability to predict relevance, especially in cases where relevance itself has a temporal dimension.

To date there has been no concerted effort to identify, model, and capitalize on temporal matters in information retrieval. Previous work has identified strategies for handling particular aspects of temporality in information retrieval (IR). This project, however, offers a sustained analysis of time’s role in IR. The project approaches temporal factors under two lenses. First, we propose identifying novel sources of temporal evidence that can inform core IR operations in general. Second, the project will address problems where relevance itself is explicitly temporal. A core argument of this proposal is that these two lenses are best studied in conjunction. This project entails a sustained, organic analysis of temporality in IR, an analysis that will advance the realism of modern retrieval systems.