IIT TREC 2006: Genomics Track

For the TREC-2006 Genomics Track, we report on the effectiveness of composite information retrieval functions based on a dimensional data model for improving document, passage, and aspect search precision of genomics literature. We designed an approach, and developed a corresponding search engine, based on a novel dimensional data model capable of document, paragraph, sentence, and passage level retrieval of genomics literature. By constructing a data warehouse style index with the flexibility of aggregating term statistics at multiple levels of document granularity, and incorporating key biological entities through shallow parsing of individual sentences, composite retrieval models combining multiple levels of contextual evidence can be efficiently developed to improve retrieval performance. The genomics track for 2006 measured document, passage, and aspect retrieval using 27 topics created by active biological researchers. Each topic fit within one of four question-oriented topic templates: the role of a gene in a disease, the effect of a gene on a biological process, how genes interact in organ function, and how mutations in genes influence biological processes. Documents for this task come from a corpus of 162,048 full-text biomedical articles. Each form of retrieval was measured with a variant of mean average precision (MAP). We submitted automatically generated results from three composite models to the TREC forum. All three models delivered results that significantly exceed the median results reported for the 2006 TREC Genomics track. The results of our best performing TREC model had MAP of 0.426 for document retrieval (53% above median), 0.055 for passage retrieval (129% above median), and 0.262 for aspect retrieval (125% above median).