Abstract
Due to exposure bias, most existing natural language generation (NLG) models trained by maximizing the likelihood objective predict poor text results during the inference stage. In this paper, to tackle this problem, we revisit the generate-then-rank framework and propose a joint generator-ranker (JGR) training algorithm for text generation tasks. In JGR, the generator model is trained by maximizing two objectives: the likelihood of the training corpus and the expected reward given by the ranker model. Meanwhile, the ranker model takes input samples from the generator model and learns to distinguish good samples from the generation pool. The generator and ranker models are alternately optimized till convergence. In the empirical study, the proposed JGR model achieves new state-of-the-art performance on five public benchmarks covering three popular generation tasks: summarization, question generation, and response generation. We will make code, data, and models available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2206.13974
