Abstract

Learning sentence representations is an essential and challenging topic in the deep learning and natural language processing communities. Recent methods pre-train big models on a massive text corpus, focusing mainly on learning the representation of contextualized words. As a result, these models cannot generate informative sentence embeddings since they do not explicitly exploit the structure and discourse relationships existing in contiguous sentences. Drawing inspiration from human language processing, this work explores how to improve sentence-level representations of pre-trained models by borrowing ideas from predictive coding theory. Specifically, we extend BERT-style models with bottom-up and top-down computation to predict future sentences in latent space at each intermediate layer in the networks. We conduct extensive experimentation with various benchmarks for the English and Spanish languages, designed to assess sentence- and discourse-level representations and pragmatics-focused assessments. Our results show that our approach improves sentence representations consistently for both languages. Furthermore, the experiments also indicate that our models capture discourse and pragmatics knowledge. In addition, to validate the proposed method, we carried out an ablation study and a qualitative study with which we verified that the predictive mechanism helps to improve the quality of the representations.