Pretraining with Contrastive Sentence Objectives Improves Discourse Performance of Language Models

TL;DR

This paper introduces CONPONO, a pretraining method with contrastive sentence objectives that enhances discourse understanding in language models, leading to significant improvements on discourse and general NLP tasks.

Contribution

The paper presents a novel inter-sentence pretraining objective, CONPONO, that models discourse coherence and improves language model performance across multiple tasks.

Findings

Up to 13% improvement on DiscoEval benchmark

Outperforms larger BERT-Large model despite same size

Gains of 2%-6% on non-discourse tasks like RTE, COPA, ReCoRD

Abstract

Recent models for unsupervised representation learning of text have employed a number of techniques to improve contextual word representations but have put little focus on discourse-level representations. We propose CONPONO, an inter-sentence objective for pretraining language models that models discourse coherence and the distance between sentences. Given an anchor sentence, our model is trained to predict the text k sentences away using a sampled-softmax objective where the candidates consist of neighboring sentences and sentences randomly sampled from the corpus. On the discourse representation benchmark DiscoEval, our model improves over the previous state-of-the-art by up to 13% and on average 4% absolute across 7 tasks. Our model is the same size as BERT-Base, but outperforms the much larger BERT- Large model and other more recent approaches that incorporate discourse. We also show that CONPONO yields gains of 2%-6% absolute even for tasks that do not explicitly evaluate discourse: textual entailment (RTE), common sense reasoning (COPA) and reading comprehension (ReCoRD).