LAMDA at TREC CDS track 2015 - Clinical Decision Support Track

In TREC 2015 Clinical Decision Support Track, our goal is to retrieve the relevant medical articles for the questions about medical statement. We propose three main strategies of indexing, query expansion, and the ranking method. In the indexing stage, each medical article is indexed into 3 different fields: title, abstract, and body. Before querying, related words are appended to the query at the query expansion stage. Our system returns the score of each field corresponding to the query for all documents. The score of each field is calculated using Divergence-from-randomness (DFR) probabilistic model. With the 3 scores from each field, the total score is calculated as the weighted sum of each score. Finally, we pick up top 1000 documents and send the list of the articles for evaluation. To make it easier for building the IR system, Elasticsearch and MetaMap are adopted for general IR operations and query expansion, respectively. Elasticsearch supports the similarity module that defines how matching documents are scored. In our IR system, Divergence-from-randomness model is adopted for probabilistic term vector space model because it is figured out that DFR outperforms all the other vector space models supported by Elasticsearch. MetaMap is the online tool that maps biomedical text to the Metathesaurus, and its semantic type. Query expansion is executed by extracting the semantic type from the description of the question, and appending words in the same semantic types to the query. Keywords—TREC, Retrieve, DFR, MetaMap, Clinical Decision Support, Query expansion, Elasticsearch