Abstract

Previous data-driven work investigating the types and distributions of discourse relation signals, including discourse markers such as ’however’ or phrases such as ’as a result’ has focused on the relative frequencies of signal words within and outside text from each discourse relation. Such approaches do not allow us to quantify the signaling strength of individual instances of a signal on a scale (e.g. more or less discourse-relevant instances of ’and’), to assess the distribution of ambiguity for signals, or to identify words that hinder discourse relation identification in context (’anti-signals’ or ’distractors’). In this paper we present a data-driven approach to signal detection using a distantly supervised neural network and develop a metric, Δs (or ’delta-softmax’), to quantify signaling strength. Ranging between -1 and 1 and relying on recent advances in contextualized words embeddings, the metric represents each word’s positive or negative contribution to the identifiability of a relation in specific instances in context. Based on an English corpus annotated for discourse relations using Rhetorical Structure Theory and signal type annotations anchored to specific tokens, our analysis examines the reliability of the metric, the places where it overlaps with and differs from human judgments, and the implications for identifying features that neural models may need in order to perform better on automatic discourse relation classification.