Pretraining with Contrastive Sentence Objectives Improves Discourse Performance of Language Models
This work addresses the lack of focus on discourse-level representations in unsupervised text learning, offering a method that enhances performance across various NLP tasks, though it is incremental in building upon existing pretraining techniques.
The authors tackled the problem of improving discourse-level representations in language models by proposing CONPONO, an inter-sentence pretraining objective that models discourse coherence and sentence distance, resulting in up to 13% absolute improvement on the DiscoEval benchmark and 2%-6% gains on non-discourse tasks.
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).