Analysis and Enhancement of Conditional Random Fields Gene Mention Taggers in BioCreative II Challenge Evaluation

Background: Tagging gene and gene product mentions in scientific text is an important initial step of literature mining. In BioCreative 2 challenge, the conditional random fields model (CRF) was the most prevailing method in the gene mention task. In this paper, we analyze two best performing CRF-based systems in BioCreative 2. We examine their key claims and propose enhancement based on the analysis results. Results: We implemented their systems in MALLET as specified in their report and in CRF++, a different CRF package, to empirically analyze their claims. We found that their feature set is effective for models trained by MALLET, but a smaller set works better for those by CRF++. We confirmed the effectiveness of pairing parentheses as a post processing step. We found that backward parsing is not always superior to forward parsing. The benefit of applying bidirectional parsing is the creation of a wider variety of complementary models. We elaborated the notion of divergent models by relating it to the difference of the increments of ture positives and false positives of the union model. Conclusions: To further enhance the performance, we can integrate more models based on the elaborated notion of divergent models that we derived to minimize the number of models required.