Reward Augmented Maximum Likelihood for Neural Structured Prediction

TL;DR

This paper introduces Reward Augmented Maximum Likelihood (RAML), a method that directly incorporates task-specific reward functions into neural structured prediction models, improving performance over traditional maximum likelihood approaches.

Contribution

The paper proposes a novel framework linking log-likelihood with expected reward, enabling direct reward optimization in neural structured prediction models.

Findings

RAML improves speech recognition accuracy.

RAML enhances machine translation quality.

Reward-based training outperforms MLE baseline.

Abstract

A key problem in structured output prediction is direct optimization of the task reward function that matters for test evaluation. This paper presents a simple and computationally efficient approach to incorporate task reward into a maximum likelihood framework. By establishing a link between the log-likelihood and expected reward objectives, we show that an optimal regularized expected reward is achieved when the conditional distribution of the outputs given the inputs is proportional to their exponentiated scaled rewards. Accordingly, we present a framework to smooth the predictive probability of the outputs using their corresponding rewards. We optimize the conditional log-probability of augmented outputs that are sampled proportionally to their exponentiated scaled rewards. Experiments on neural sequence to sequence models for speech recognition and machine translation show notable improvements over a maximum likelihood baseline by using reward augmented maximum likelihood (RAML), where the rewards are defined as the negative edit distance between the outputs and the ground truth labels.