Implicit Discourse Relation Identification for Open-domain Dialogues.

摘要: Discourse relation identification has been an active area of research for many years, and the challenge identifying implicit relations remains largely unsolved task, especially in context open-domain dialogue system. Previous work primarily relies on a corpora formal text which is inherently non-dialogic, i.e., news journals. This data however not suitable to handle nuances informal nor it capable navigating plethora valid topics present dialogue. In this paper, we designed novel discourse pipeline specifically tuned systems. We firstly propose method automatically extract argument pairs labels from dataset dialogic turns, resulting corpus pairs; first its kind attempt identify connecting turns discourse. Moreover, have taken steps leverage features unique our task further improve such by performing feature ablation incorporating enhance state-of-the-art model.