Information extraction (IE) is recently gaining much attention due to the explosive growth of electronically available documents. In order to turn preliminary results from IE into useful resources for knowledge discovery, however, we need to augment the performance of an available IE system further, since the state-of-the-art performance of IE systems in the field is not yet perfect. In this paper, we present novel methods for enhancing the correctness and coverage of an IE system. In particular, a method for identifying incorrectly extracted results due to coordination in natural language is proposed to deal with the issue of correctness, and a method visualizing both the extracted results and the corresponding natural language texts is proposed to deal with the issue of coverage. The preference over the kind of information varies greatly among users, and thus the information that is output from an IE system needs further customization. For this purpose, we present novel methods for visualization that can be customized to the individual user`s varying expectations and expertise.
We have applied our proposal for knowledge discovery in bioinformatics. Through a case study for molecular interactions in the yeast, we show a meaningful (6.5%) increase in precision of a relatively mature IE system. We also provide a way to give visual clues and insights to extractable information from source texts for coverage. Finally, we demonstrate sample pathway maps on three kinds of MAPK, that are successfully constructed from molecular interactions in the mitogen-activated protein kinase (MAPK) pathways.
