A Collocation Method to Extract Biological Keywords and Its Application to Protein Name Recognition

Mining biological relationships from scientific text, which facilitates automatic construction of knowledge base and finding of new information, becomes one of the emerging applications. Named entity recognition is a fundamental task in relationship mining. Traditional methods listed a pre-defined set of words to indicate protein or gene interactions by intuition. The argument is that we cannot assure if the keyword set is complete. A collocation approach not only mines biological keywords for relationship establishment, but also employs the keywords to improve performance of protein name recognition. This paper proposes a collocation approach to extract biological keywords or key phrases. Frequency, mean-and-variance, and t-test statistics are computed. Compound collocates are also resolved in this study. The experiments show that the performances of the t-test model are 76.67% and 95.45% without and with considerations of text domains, respectively. This paper suggests many useful terms that the previous literatures do not touch on. The results are valuable for the investigation of relationships between proteins, and can be integrated into genome analysis systems. In the application of protein name recognition, it enhances the precision from 70.90% to 81.94% on Yapex system.