An Unsupervised Joint Model for Claim Detection

Claim detection is one of the most important tasks in argument mining. Most existing work employs supervised methods that rely on not only good-quality and large-scale annotated corpora, but also highly engineered and sophisticated features. Unsupervised methods are a possible solution to the above problems but few work has been done from unsupervised perspective. In this paper, we propose an unsupervised joint model including position model, indicator model and TextRank model. Position information is important for argument components detection, and our position model not only considers the sentences at the beginning and the end of the whole text but also at the beginning and the end of each paragraph. Considering the discourse makers’ good indication of claims, we also introduce indicator model into our joint model. Experiments on three English argumentation corpora show that our model outperforms the state-of-the-art unsupervised methods for claim detection.