NovaSearch at TREC 2014 Clinical Decision Support Track
Abstract
This paper describes the participation of the NovaSearch group at TREC Clinical Decision Support 2014. As this is the first edition of the track, we decided to assess the performance of multiple information retrieval techniques: retrieval functions, re-ranking, query expansion and classification of medical articles into question categories. The best performing run was based on an ensemble of state-of-the-art retrieval algorithms combined with unsupervised fusion. Our best run was based on the late fusion of runs using MeSH query expansion, pseudo-relevance feedback with terms from top retrieved results and multiple retrieval functions (BM25L, BM25+, TF-IDF and Language Models) combined with RRF fusion algorithm. We also tested an algorithm to measure article relevance to the target medical questions (diagnosis, test and treatment articles), based on the frequency of words to some categories. An additional experiment was based on pseudo relevance feedback based on each article's journal reputation. Although some techniques did not increase our baseline performance, we are satisfied with our global performance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 2014
- Accession Number
- ADA618803
Entities
People
- Andre Mourao
- Flavio Martins
- Joao Magalhaes
Organizations
- NOVA University Lisbon