A Side-by-side Comparison of Transformers for English Implicit Discourse Relation Classification

TL;DR

This study compares seven pre-trained language models for implicit discourse relation classification using the PDTB-3 dataset, achieving state-of-the-art accuracy and revealing that models with MLM objectives and full attention perform best, contrary to previous reports.

Contribution

It provides the first comprehensive comparison of multiple language models for implicit discourse relation classification and establishes new state-of-the-art results.

Findings

MLM-based models outperform sentence-level pre-training objectives.

Full attention models yield better performance.

Achieved SOTA accuracy of 0.671 on PDTB-3.

Abstract

Though discourse parsing can help multiple NLP fields, there has been no wide language model search done on implicit discourse relation classification. This hinders researchers from fully utilizing public-available models in discourse analysis. This work is a straightforward, fine-tuned discourse performance comparison of seven pre-trained language models. We use PDTB-3, a popular discourse relation annotated dataset. Through our model search, we raise SOTA to 0.671 ACC and obtain novel observations. Some are contrary to what has been reported before (Shi and Demberg, 2019b), that sentence-level pre-training objectives (NSP, SBO, SOP) generally fail to produce the best performing model for implicit discourse relation classification. Counterintuitively, similar-sized PLMs with MLM and full attention led to better performance.