Pseudo-Relevance Feedback Based on Locally-Built Co-occurrence Graphs

In Information Retrieval (IR), user queries are often too short, making the selection of relevant documents hard. Pseudo-relevance feedback (PRF) is an effective method to automatically expand the query with new terms using a set of pseudo-relevant documents. However, a main issue in PRF is the selection of good expansion terms that allow improving retrieval effectiveness. In this paper, we present a new PRF method based on locally-built term co-occurrence graphs. We use a context window-based approach to construct our term co-occurrence graphs over top pseudo-relevant documents. For expansion terms selection, we propose an adapted version of the BM25 model, which allows to measure term-term similarity in co-occurrence graphs. This measure has the advantage of selecting discriminant expansion terms that are semantically related to the query as a whole. We evaluate our PRF method using four TREC collections, including the standard TREC Robust04 collection and the newest TREC Washington Post collection. Experimental results show that our proposal outperforms competitive state-of-the-art baselines and achieves significant improvements.