phrases, implicit relations within documents can be considered too. Thus, distributional and knowledge-based representations of complex semantics (i.e. words, sentences and documents) identify complementary semantic aspects of the underlying documents.

We propose to integrate documents' knowledge-based representations in case-based retrieval [2], as a form of complementary refinement for distributional representations. Recent approaches exploit semantic relations to enhance the quality of learned word or concept representations [1,4], we propose to explicitly leverage semantic relations to model document representations. Therefore, our approach extracts concepts from documents (and queries) and connect them using the semantic relations contained within a reference knowledge source -creating a knowledge graph representation for the document (or query). The intuition is that semantic relations carry high informative power that can boost precision.

The knowledge graph representation can reduce the contextual dependency of distributional representations and help discriminating more effectively semantically similar from non-semantically similar texts. Besides, since the concepts considered are only those extracted from a document or a query, the problem of topic driftoccurring when the query is expanded with concepts that are not pertinent to the information need -is reduced.

We propose to combine implicit and explicit representations for case-based retrieval in two different ways: (i) considering documentlevel knowledge graphs as additional inputs for end-to-end neural scoring models that learn the relevance of document-query pairs via semantic features; (ii) considering document-level knowledge graphs with pseudo relevance feedback to boost documents in top positions that present a more similar graph compared with the query graph. Both approaches aim at reducing the semantic gap between queries and documents.