Multi-Granularity Retrieval Model for Bridging Gaps between Biomedical Concepts and Entities: THUIR at TREC 2007 Genomics Track

Abstract General concepts are always used to describe query requirement (In the example “What tumor types are associated with Rb1 mutations?”, “Tumor types” is a general concept, and its entity in a relevant documents can be “brain tumor”). To bridge the gaps between concepts in user queries and entities in relevant documents, we proposed a multi-granularity retrieval model in TREC 2007 Genomics task. The model consists of three components: (1) Paragraph retrieval is employed to retrieve candidate paragraph initially; (2) Dictionary-based NER is utilized to recognize named entities of given types; (3) Passage ranking is used to rank retrieved candidate passages. Our proposed model achieve promising result (Passage MAP=0.1023, with NER bottleneck eliminated).