Modelling document-query interaction in a hierarchical neural model for IR.

Recent deep approaches to information retrieval are either representation-oriented or interaction-oriented, depending on how they view the modelling of document and query representations and their interactions. We explore a hierarchical approach to document encoding that enables modelling the query-document interaction at different levels of granularity. The proposed model splits the input documents into blocks that are individually matched to a given query through a series of self-attention modules, along with pooling and projection layers. We test our method on the LETOR 4.0 MQ2007 standard IR collection. The approach shows promising preliminary results, albeit a more in-depth exploration of the modelling choices could provide further gains.