Effective reranking for extracting protein-protein interactions from biomedical literature

A semantic parser based on the hidden vector state (HVS) model has been proposed for extracting protein-protein interactions. The HVS model is an extension of the basic discrete hidden Markov model, in which context is encoded as a stack-oriented state vector and state transitions are factored into a stack shift operation followed by the push of a new preterminal category label. In this paper, we investigate three different models, log-linear regression (LLR), neural networks (NNs) and support vector machines (SVMs), to rerank parses generated by the HVS model for protein-protein interactions extraction. Features used for reranking are manually defined which include the parse information, the structure information, and the complexity information. The experimental results show that reranking can indeed improve the performance of protein-protein interactions extraction, and reranking based on SVM gives more stable performance than LLR and NN.