Classifying Gene Sentences in Biomedical Literature by Combining High-Precision Gene Identifiers

摘要

Gene name identification is a fundamental step to solve more complicated text mining problems such as gene normalization and protein-protein interactions. However, state-of-the-art name identification methods are not yet sufficient for use in a fully automated system. In this regard, a relaxed task, gene/protein sentence identification, may serve more effectively for manually searching and browsing biomedical literature. In this paper, we set up a new task, gene/protein sentence classification and propose an ensemble approach for addressing this problem. Well-known named entity tools use similar gold-standard sets for training and testing, which results in relatively poor performance for unknown sets. We here explore how to combine diverse high-precision gene identifiers for more robust performance. The experimental results show that the proposed approach outperforms BANNER as a stand-alone classifier for newly annotated sets as well as previous gold-standard sets.