A Position-Aware Deep Model for Relevance Matching in Information Retrieval

In order to adopt deep learning for ad-hoc information retrieval, it is essential to establish suitable representations of query-document pairs and to design neural architectures that are able to digest such representations. In particular, they ought to capture all relevant information required to assess the relevance of a document for a given user query, including term overlap as well as positional information such as proximity and term dependencies. While previous work has successfully captured unigram term matches, none has successfully used position-dependent information on a standard benchmark test collection. In this work, we address this gap by encoding the relevance matching in terms of similarity matrices and using a deep model to digest such matrices. We present a novel model architecture consisting of convolutional layers to capture term dependencies and proximity among query term occurrences, followed by a recurrent layer to capture relevance over different query terms. Extensive experiments on TREC Web Track data confirm that the proposed model with similarity matrix representations yields improved search results.