A Mutual Information Maximization Perspective of Language Representation Learning

Abstract

We show state-of-the-art word representation learning methods maximize an objective function that is a lower bound on the mutual information between different parts of a word sequence (i.e., a sentence). Our formulation provides an alternative perspective that unifies classical word embedding models (e.g., Skip-gram) and modern contextual embeddings (e.g., BERT, XLNet). In addition to enhancing our theoretical understanding of these methods, our derivation leads to a principled framework that can be used to construct new self-supervised tasks. We provide an example by drawing inspirations from related methods based on mutual information maximization that have been successful in computer vision, and introduce a simple self-supervised objective that maximizes the mutual information between a global sentence representation and n-grams in the sentence. Our analysis offers a holistic view of representation learning methods to transfer knowledge and translate progress across multiple domains (e.g., natural language processing, computer vision, audio processing).

12 pages, 3 figures

References in corpus (5)

• Cross-lingual Language Model Pretraining
• Learning Representations by Maximizing Mutual Information Across Views
• MASS: Masked Sequence to Sequence Pre-training for Language Generation
• A Theoretical Analysis of Contrastive Unsupervised Representation Learning
• Learning and Evaluating General Linguistic Intelligence

Cited by in corpus (12)

• Contrastive Representation Learning: A Framework and Review
• Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing
• On Mutual Information in Contrastive Learning for Visual Representations
• The Effect of Natural Distribution Shift on Question Answering Models
• A Survey on Self-supervised Pre-training for Sequential Transfer Learning in Neural Networks
• Understand and Improve Contrastive Learning Methods for Visual Representation: A Review
• Evolution Is All You Need: Phylogenetic Augmentation for Contrastive Learning
• Improving Robustness and Generality of NLP Models Using Disentangled Representations
• Hybrid Generative-Contrastive Representation Learning
• Self-supervised Representation Learning with Relative Predictive Coding
• Self-Supervised Graph Learning with Proximity-based Views and Channel Contrast
• Cross-modal Image Retrieval with Deep Mutual Information Maximization