Semi-supervised prediction of protein interaction sentences exploiting semantically encoded metrics

Protein-protein interaction (PPI) identification is an integral component of many biomedical research and database curation tools. Automation of this task through classification is one of the key goals of text mining (TM). However, labelled PPI corpora required to train classifiers are generally small. In order to overcome this sparsity in the training data, we propose a novel method of integrating corpora that do not contain relevance judgements. Our approach uses a semantic language model to gather word similarity from a large unlabelled corpus. This additional information is integrated into the sentence classification process using kernel transformations and has a re-weighting effect on the training features that leads to an 8% improvement in F-score over the baseline results. Furthermore, we discover that some words which are generally considered indicative of interactions are actually neutralised by this process