Full-texts Representations with Medical Subject Headings, and Co-citations Network Reranking Strategies for TREC 2014 Clinical Decision Support Track.

摘要

In TREC 2014 Clinical Decision Support Track, the task was to retrieve full-texts relevant for answering generic clinical questions about medical records. For this purpose, we investigated a large range of strategies in the five runs we officially submitted. Concerning Information Retrieval (IR), we tested two different indexing levels: documents or sections. Section indexing was clearly below (-40% in R-Precision). In the domain of Information Extraction, we enriched documents with Medical Subject Headings concepts that were collected from MEDLINE or extracted in the text with exact match strategies. We also investigated a target-specific semantic enrichment: MeSH terms representing diagnosis, treatments or tests (relying on UMLS semantic types) were used both in collection and in queries to guide the retrieval. Unfortunately, the MeSH representation was not as complementary with the text as we expected, and the results were disappointing. Concerning post-processing strategies, we tested the boosting of specific articles types (e.g. review articles, case reports), but the IR process already tended to favour these article types. Finally, we applied a reranking strategy relying on the cocitations network, thanks to normalized references provided in the corpus. This last strategy led to a slight improvement (+5%).