A Simple and Strong Baseline for End-to-End Neural RST-style Discourse Parsing

TL;DR

This paper establishes a strong baseline for RST-style discourse parsing by integrating simple strategies with transformer-based pre-trained models, showing that the choice of language model significantly impacts performance.

Contribution

It demonstrates that leveraging pre-trained language models, especially DeBERTa with span-masking, greatly improves discourse parsing accuracy over existing methods.

Findings

Pre-trained language models are more influential than parsing strategies.

Bottom-up parser with DeBERTa outperforms previous best parsers.

Span-masking in language models enhances intra- and multi-sentential parsing.

Abstract

To promote and further develop RST-style discourse parsing models, we need a strong baseline that can be regarded as a reference for reporting reliable experimental results. This paper explores a strong baseline by integrating existing simple parsing strategies, top-down and bottom-up, with various transformer-based pre-trained language models. The experimental results obtained from two benchmark datasets demonstrate that the parsing performance strongly relies on the pretrained language models rather than the parsing strategies. In particular, the bottom-up parser achieves large performance gains compared to the current best parser when employing DeBERTa. We further reveal that language models with a span-masking scheme especially boost the parsing performance through our analysis within intra- and multi-sentential parsing, and nuclearity prediction.