An automated method to build a corpus of rhetorically-classified sentences in biomedical texts

The rhetorical classification of sentences in biomedical texts is an important task in the recognition of the components of a scientific argument. Generating supervised machine learned models to do this recognition requires corpora annotated for the rhetorical categories Introduction (or Background), Method, Result, Discussion (or Conclusion). Currently, a few, small annotated corpora exist. We use a straightforward feature of co-referring text using the word “this” to build a selfannotating corpus extracted from a large biomedical research paper dataset. The corpus is annotated for all of the rhetorical categories except Introduction without involving domain experts. In a 10-fold cross-validation, we report an overall Fscore of 97% with Naı̈ve Bayes and 98.7% with SVM, far above those previously reported.