Classification of Online Medical Discourse by Modified Co-Training

Online health communities have become an important resource that both patients and family members rely on to gain medical information and interact with contributors and experts outside the clinical setting. However, few studies have focused on analysing sentiment in medical forum discourse. Analysing sentiment expressed by members of a health community in medical forum discourse can be of significant value, such as by identifying characteristics of an information space, determining themes that predominate among a large dataset, and allowing people to summarize topics within a big data set. In this paper, we identify sentiment expressed in online medical forums discussing Lyme Disease and lupus. In many sentiment analysis applications, the labelled training datasets are fairly expensive to obtain, whereas unlabelled datasets are readily available. Therefore, we propose an adaptation of a well-known semi-supervised learning technique, in which co-training is implemented by combining labelled and unlabelled data. We use domain-independent and domain-dependent sentiment features as two distinct sets of views of a post and utilize them in order to identify the multi-class sentiment expressed using an adaptive co-training model. The results of our experiments demonstrate the effectiveness of the proposed approach.