Synonym-Based Expansion and Boosting-Based Re-Ranking: A Two-phase Approach for Genomic Information Retrieval.
Zhongmin Shi, Baohua Gu, Fred P. Popowich, Anoop Sarkar · Text REtrieval Conference · 2005
We describe in this paper the design and evaluation of the system built at Simon Fraser University for the TREC 2005 adhoc retrieval task in the Genomics track. The main approach taken in our system was to expand synonyms by exploiting a fusion of a set of biomedical and general ontology sources, and apply machine learning and natural language processing techniques to re-rank retrieved documents. In our system, we integrated EntrezGene, HUGO, Eugenes, ARGH, GO, MeSH, UMLSKS and WordNet into a large reference database and then used a conventional Information Retrieval (IR) toolkit, the Lemur toolkit (Lemur, 2005), to build an IR system. In the postprocessing phase, we applied a boosting algorithm (Kudo and Matsumoto, 2004) that captured natural language substructures embedded in texts to re-rank the retrieved documents. Experimental results show that the boosting algorithm worked well in cases where a conventional IR system performs poorly, but this re-ranking approach was not robust enough when applied to broad coverage task typically associated with IR.