Integrating Two Linguistic Phenomena: Bi-directional Interplay between Discourse and Temporal Relations

Abstract

Discourse relations and temporal relations are interrelated. The former relation type describes the logical flow of a text, while the latter governs the timeline of events. For example, a discourse relation, Cause, inherently implies that the cause precedes the effect. However, there is currently a lack of empirical evidence in computational linguistics to support this notion. Thus, this work aims to explore the interplay between temporal relations and implicit discourse relations. We present a novel method that combines multi-task learning and iterative cross-learning to enable the model to transfer knowledge between the two tasks. In this process, discourse knowledge and temporal relations mutually assist the model in solving both tasks. Our experimental results demonstrate that the model’s performance benefits from knowledge exchange between these two tasks, providing empirical evidence for this interplay.