Relevance Ranking for Web Search

Relevance ranking is a core problem of Information Retrieval which plays a fundamental role in various real world applications, such as search engines. Given a query and a set of candidate text documents, relevance ranking algorithms determine how relevant each text document is for the given query. This degree of relevance allows them to rank the text documents and perform actions such as returning the best matching documents for the query. As in other machine learning and computational intelligence disciplines, deep learning techniques have recently achieved state of the art results by successfully capturing relevance matching signals between query-textual document pairs. This paper focuses on the PositionAware Convolutional-Recurrent Relevance Matching approach. On a first phase, it reimplements the original work, reproduces the published results and performs a number of additional experiments that identify potential model limitations. On a second phase, it explores possible model improvements based on deep learning techniques such as soft self-attention and deep transfer learning. Experiments on the well-known TREC Web Track data show that it is possible to obtain small improvements over the original model and point to a number of limitations of the general approach due to the information bottlenecks involved.