Efficient semantic indexing via neural networks with dynamic supervised feedback

We describe a portable system for e cient semantic indexing of documents via neural networks with dynamic supervised feedback. We initially represent each document as a modified TF-IDF sparse vector and then apply a learned mapping to a compact embedding space. This mapping is produced by a shallow neural network which learns a latent representation for the textual graph linking words to nearby contexts. The resulting document embeddings provide significantly better semantic representation, partly because they incorporate information about synonyms. Query topics are uniformly represented in the same manner as documents. For each query, we dynamically train an additional hidden layer which modifies the embedding space in response to relevance judgements. The system was tested using the documents and topics provided in the Total Recall track.