The effects of discourse connectives prediction on implicit discourse relation recognition

Implicit discourse relation recognition is difficult due to the absence of explicit discourse connectives between arbitrary spans of text. In this paper, we use language models to predict the discourse connectives between the arguments pair. We present two methods to apply the predicted connectives to implicit discourse relation recognition. One is to use the sense frequency of the specific connectives in a supervised framework. The other is to directly use the presence of the predicted connectives in an unsupervised way. Results on PDTB2 show that using language model to predict the connectives can achieve comparable F-scores to the previous state-of-art method. Our method is quite promising in that not only it has a very small number of features but also once a language model based on other resources is trained it can be more adaptive to other languages and domains.