BERT as a Teacher: Contextual Embeddings for Sequence-Level Reward

Abstract

Measuring the quality of a generated sequence against a set of references is a central problem in many learning frameworks, be it to compute a score, to assign a reward, or to perform discrimination. Despite great advances in model architectures, metrics that scale independently of the number of references are still based on n-gram estimates. We show that the underlying operations, counting words and comparing counts, can be lifted to embedding words and comparing embeddings. An in-depth analysis of BERT embeddings shows empirically that contextual embeddings can be employed to capture the required dependencies while maintaining the necessary scalability through appropriate pruning and smoothing techniques. We cast unconditional generation as a reinforcement learning problem and show that our reward function indeed provides a more effective learning signal than n-gram reward in this challenging setting.

References in corpus (14)

• Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
• BERTScore: Evaluating Text Generation with BERT
• Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks
• The Curious Case of Neural Text Degeneration
• MaskGAN: Better Text Generation via Filling in the______
• Professor Forcing: A New Algorithm for Training Recurrent Networks
• An Actor-Critic Algorithm for Sequence Prediction
• Hierarchical Neural Story Generation
• BERT has a Mouth, and It Must Speak: BERT as a Markov Random Field Language Model
• Language GANs Falling Short
• On Accurate Evaluation of GANs for Language Generation
• Latent Normalizing Flows for Discrete Sequences
• DiscoFuse: A Large-Scale Dataset for Discourse-Based Sentence Fusion
• On the Weaknesses of Reinforcement Learning for Neural Machine Translation

Cited by in corpus (1)

• Reward Optimization for Neural Machine Translation with Learned Metrics