An Analysis of Sentential Neighbors in Implicit Discourse Relation Prediction

TL;DR

This work tackles implicit discourse relation prediction without explicit markers, testing whether broader sentential context improves accuracy. It introduces three context-expansion strategies—Direct Neighbors (DNs), Expanded Window Neighbors (EWNs), and Part-Smart Random Neighbors (PSRNs)—and evaluates them against a DistilBERT-based baseline on the Penn Discourse TreeBank 2.0. The main finding is that additional context beyond a single discourse unit generally harms performance, with the Baseline model outperforming the context-window approaches, though PSRN and EWNs can be competitive in certain settings. These results highlight the importance of careful context design and dataset considerations for implicit relation classification and suggest directions for future research and dataset construction in this area.

Abstract

Discourse relation classification is an especially difficult task without explicit context markers (Prasad et al., 2008). Current approaches to implicit relation prediction solely rely on two neighboring sentences being targeted, ignoring the broader context of their surrounding environments (Atwell et al., 2021). In this research, we propose three new methods in which to incorporate context in the task of sentence relation prediction: (1) Direct Neighbors (DNs), (2) Expanded Window Neighbors (EWNs), and (3) Part-Smart Random Neighbors (PSRNs). Our findings indicate that the inclusion of context beyond one discourse unit is harmful in the task of discourse relation classification.

Figures (3)

• Figure 1: Expanded Context Windows.
• Figure 2: PDTB-2.0 Sense Tag Hierarchy.
• Figure 3: F1 Metrics for Baseline and All Models.