Development of bespoke machine learning and biocuration workflows in a BioC-supporting text mining workbench

As part of our participation in the Collaborative Biocurator Assistant Task of BioCreative V, we developed methods and tools for recognising and normalising mentions denoting genes/proteins and organisms. A combination of different approaches were used in addressing these tasks. The recognition of gene/protein and organism names was cast as a sequence labelling problem to which the conditional random fields algorithm was applied. In training our models, various lexical and orthographic features were extracted over the CHEMDNER GPRO and S800 corpora which were leveraged as gold standard data. Our feature set was further enriched with semantic attributes drawn from matches between mentions in text and entries in relevant dictionaries. In normalising recognised names, i.e., assigning resource identifiers to recognised names in text, the Jaro-Winkler and Levenshtein distance measures were used to estimate string similarity and -rank candidate matches. Integration of the various techniques and resources was facilitated by the Web-based Argo text mining workbench which allows for the straightforward construction of automatic text processing workflows. Upon using our training workflow to produce gene/protein and organism name recognition models and subsequently evaluating them, micro-averaged F-scores of 70% and 72.87% were obtained, respectively. Curation workflows applying our models on the provided BioC corpus of 120 full-text PubMed Central documents generated normalised named entity annotations which were serialised in the required BioC format.