DisSent: Sentence Representation Learning from Explicit Discourse Relations

TL;DR

DisSent introduces a method to learn high-quality sentence representations using explicitly annotated discourse relations derived from dependency parsing, improving transfer task performance and state-of-the-art results on implicit relation prediction.

Contribution

The paper presents a novel dataset and approach leveraging explicit discourse relations for sentence embedding learning, reducing reliance on large-scale text or manual annotations.

Findings

High-quality sentence embeddings from discourse relations

State-of-the-art performance on Penn Discourse Treebank implicit task

Effective transfer learning on various NLP tasks

Abstract

Learning effective representations of sentences is one of the core missions of natural language understanding. Existing models either train on a vast amount of text, or require costly, manually curated sentence relation datasets. We show that with dependency parsing and rule-based rubrics, we can curate a high quality sentence relation task by leveraging explicit discourse relations. We show that our curated dataset provides an excellent signal for learning vector representations of sentence meaning, representing relations that can only be determined when the meanings of two sentences are combined. We demonstrate that the automatically curated corpus allows a bidirectional LSTM sentence encoder to yield high quality sentence embeddings and can serve as a supervised fine-tuning dataset for larger models such as BERT. Our fixed sentence embeddings achieve high performance on a variety of transfer tasks, including SentEval, and we achieve state-of-the-art results on Penn Discourse Treebank's implicit relation prediction task.