Investigate neural word embeddings as a source of evidence in document ranking.
Presented in [this Stanford course on IR](/doc/?uri=https%3A%2F%2Fweb.stanford.edu%2Fclass%2Fcs276%2Fhandouts%2Flecture20-distributed-representations.pdf) by Chris Manning (starting slide 44)
They train a word2vec model, but retain both the input and the output projections.
> During ranking we map the query words into the input space and the document words into the output space, and compute a query-document relevance score by aggregating the cosine similarities across all the query-document word pairs.
> However, when ranking a larger set of candidate documents, we find the embeddings-based approach is prone to false positives