Identifying intention posts in discussion forums using multi-instance learning and multiple sources transfer learning

This paper proposes a novel method for identifying intention posts in discussion forums. The main problem of identifying intention posts in discussion forums is that there exist a few intention sentences even in a post expressing an intention. That is, an intention post consists of a few intention sentences and a number of non-intention sentences, while non-intention posts have only non-intention sentences. Therefore, multi-instance learning which regards a post as a bag and the sentences in the post as instances of the bag is adopted as a solution to this problem. One distinct characteristic of the posts is that the ways of expressing an intention are similar across domains. Thus, we incorporate a multiple sources transfer learning into the multi-instance learning. As a result, the multi-instance learning is enhanced by leveraging knowledge of expressing intentions from multiple source domains. Through a set of experiments, it is proven that the proposed method is effective at identifying intention posts in discussion forums.