Semi-supervised Pattern Learning for Extracting Relations from Bioscience Texts

A variety of pattern-based methods have been exploited to extract biological relations from literatures. Many of them require significant domain-specific knowledge to build the patterns by hand, or a large amount of labeled data to learn the patterns automatically. In this paper, a semisupervised model is presented to combine both unlabeled and labeled data for the pattern learning procedure. First, a large amount of unlabeled data is used to generate a raw pattern set. Then it is refined in the evaluating phase by incorporating the domain knowledge provided by a relatively small labeled data. Comparative results show that labeled data, when used in conjunction with the inexpensive unlabeled data, can considerably improve the learning accuracy.