Automatic Recognition of Topic-Classified Relations between Prostate Cancer and Genes from Medline Abstracts

To recognize instances of medical information concerning prostate cancer and its relevant genes, we developed a machine learning-based relation recognizer using rich contextual features. We collected prostate cancer-related abstracts from Medline. We then constructed an annotated corpus of prostate cancer and gene relations, which consisted of six topic − classified categories, with more detailed information describing the type of prostate cancer and gene relation. The corpus was made with the help of biologists and a disease and gene dictionarybased name recognition technique. The process of dictionary-based name recognition generates disease-gene pairs that become candidates for biomedically related pairs. Since dictionary matching tends to over-generate candidates, we used a machine learning-based named entity recognition method (1) to provide a feature for each candidate, and (2) to filter out overgenerated candidates. Experimental results showed that using a maximum entropy-based relation recognition method and a maximum entropybased named entity recognition method together greatly improves precision at the cost of a small reduction in recall for the topic-classified relations.