Comparison of different strategies for utilizing two CHEMDNER corpora

To identify chemical entities and drug names in patent according to CHEMDNER patent task-CEMP subtask, we use machine learning technique to construct a chemical named entity recognition (CNER) system. It is desirable for machine-based CNER system to have large training examples. Two CHEMDNER corpora have been developed. One is the corpus for the patent task and the other is the CHEMDNER corpus for PubMed abstract constructed for CHEMDNER task in BioCreative IV. Both corpora were constructed based on very similar guidelines. However, the style of writing is different. In this paper, we are discussing different strategies to utilize these two corpora to identify chemical entities in patent. Our basic system uses conditional random field (CRF) as a machine learning technique that uses linguistic features in addition to domain knowledge feature produced by ChemSpot. We compare the results of these strategies using simple system performance measures (e.g., recall, precision, and F-score) and analysis on the unique findings of each system.