Third Workshop on Building and Evaluating Resources for Biomedical Text Mining Workshop Programme

Previous studies have shown that various biomedical subdomains have lexical, syntactic, semantic and discourse structure variations. It is essential to recognise such differences to understand that biomedical natural language processing tools, such as named entity recognisers, that work well on some subdomains may not work as well on others. In this paper, we investigate the pairwise similarity (or dissimilarity) amongst twenty selected biomedical subdomains, at the level of named entity types. We evaluate the contribution of these types in the classification task by computing the chi-squared statistic over their distributions. We then build a binary classifier for each possible pair of subdomains, the results of which indicate the subdomains that are highly different or similar to others. The findings can be of potential use to those building or using named entity recognisers in determining which types of named entities need to be taken into consideration or in adapting already existing tools.