Improved Natural Language Generation via Loss Truncation
arXiv:2004.14589
Abstract
Neural language models are usually trained to match the distributional properties of a large-scale corpus by minimizing the log loss. While straightforward to optimize, this approach forces the model to reproduce all variations in the dataset, including noisy and invalid references (e.g., misannotation and hallucinated facts). Worse, the commonly used log loss is overly sensitive to such phenomena and even a small fraction of noisy data can degrade performance. In this work, we show that the distinguishability of the models and reference serves as a principled and robust alternative for handling invalid references. To optimize distinguishability, we propose loss truncation, which adaptively removes high loss examples during training. We show this is as easy to optimize as log loss and tightly bounds distinguishability under noise. Empirically, we demonstrate that loss truncation outperforms existing baselines on distinguishability on a summarization task, and show that samples generated by the loss truncation model have factual accuracy ratings that exceed those of baselines and match human references.
ACL 2020 Camera Ready Submission
References in corpus (19)
- Sequence to Sequence Learning with Neural Networks
- Large Scale GAN Training for High Fidelity Natural Image Synthesis
- The Curious Case of Neural Text Degeneration
- CTRL: A Conditional Transformer Language Model for Controllable Generation
- Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyond
- Neural Text Generation with Unlikelihood Training
- Adversarial Learning for Neural Dialogue Generation
- Evaluating the State-of-the-Art of End-to-End Natural Language Generation: The E2E NLG Challenge
- Sever: A Robust Meta-Algorithm for Stochastic Optimization
- Get To The Point: Summarization with Pointer-Generator Networks
- Adversarial Generation of Natural Language
- HYPE: A Benchmark for Human eYe Perceptual Evaluation of Generative Models
- Defending Against Neural Fake News
- Language GANs Falling Short
- Challenges in Data-to-Document Generation
- Sticking to the Facts: Confident Decoding for Faithful Data-to-Text Generation
- Minimum Risk Training for Neural Machine Translation
- Shaping the Narrative Arc: An Information-Theoretic Approach to Collaborative Dialogue
- Constrained Decoding for Neural NLG from Compositional Representations in Task-Oriented Dialogue