Let's be explicit about that: Distant supervision for implicit discourse relation classification via connective prediction

In implicit discourse relation classification, we want to predict the\nrelation between adjacent sentences in the absence of any overt discourse\nconnectives. This is challenging even for humans, leading to shortage of\nannotated data, a fact that makes the task even more difficult for supervised\nmachine learning approaches. In the current study, we perform implicit\ndiscourse relation classification without relying on any labeled implicit\nrelation. We sidestep the lack of data through explicitation of implicit\nrelations to reduce the task to two sub-problems: language modeling and\nexplicit discourse relation classification, a much easier problem. Our\nexperimental results show that this method can even marginally outperform the\nstate-of-the-art, in spite of being much simpler than alternative models of\ncomparable performance. Moreover, we show that the achieved performance is\nrobust across domains as suggested by the zero-shot experiments on a completely\ndifferent domain. This indicates that recent advances in language modeling have\nmade language models sufficiently good at capturing inter-sentence relations\nwithout the help of explicit discourse markers.\n