Determining causal and non-causal relationships in biomedical text by classifying verbs using a Naive Bayesian Classifier

Since scientific journals are still the most important means of documenting biological findings, biomedical articles are the best source of information we have on protein-protein interactions. The mining of this information will provide us with specific knowledge of the presence and types of interactions, and the circumstances in which they occur. There are various linguistic constructions that can describe a protein-protein interaction, but in this paper we will focus on subject-verb-object constructions. If a certain protein is mentioned in the subject of a sentence, and another protein in the object, we assume in this paper that some interaction is described between those proteins. The verb phrase that links the subject and object together plays an important role in this. However, there are a great many different verbs in the English language that can be used in a description of a protein-protein interaction. Since it is practically impossible to manually determine the specific biomedical meanings for all of these verbs, we try to determine these meanings automatically. We define two classes of proteinprotein interactions, causal and non-causal, and using a Naive Bayesian Classifier, we predict for a given verb in which class it belongs. This process is a first step in automatically creating a useful network of interacting proteins out of information from biomedical journals.