Learning to Compare for Better Training and Evaluation of Open Domain Natural Language Generation Models
Wangchunshu Zhou, Ke Xu
TL;DR
<3-5 sentence high-level summary>Open-domain NLG evaluation is challenged by the poor alignment of traditional metrics with human judgments. The authors introduce a self-supervised, BERT-based pairwise evaluator that compares two generated samples for quality and a Glicko2-based skill-rating system for model-level assessment, with optional human preference fine-tuning. Across story generation and open-domain dialogue, the method shows higher correlation with human judgments than BLEU, perplexity, ADEM, or adversarial evaluators, and improves hyperparameter tuning and early stopping decisions. This framework enables scalable, human-aligned evaluation without heavy reliance on gold references, enhancing both evaluation fidelity and training efficiency.
Abstract
Automated evaluation of open domain natural language generation (NLG) models remains a challenge and widely used metrics such as BLEU and Perplexity can be misleading in some cases. In our paper, we propose to evaluate natural language generation models by learning to compare a pair of generated sentences by fine-tuning BERT, which has been shown to have good natural language understanding ability. We also propose to evaluate the model-level quality of NLG models with sample-level comparison results with skill rating system. While able to be trained in a fully self-supervised fashion, our model can be further fine-tuned with a little amount of human preference annotation to better imitate human judgment. In addition to evaluating trained models, we propose to apply our model as a performance indicator during training for better hyperparameter tuning and early-stopping. We evaluate our approach on both story generation and chit-chat dialogue response generation. Experimental results show that our model correlates better with human preference compared with previous automated evaluation approaches. Training with the proposed metric yields better performance in human evaluation, which further demonstrates the effectiveness of the proposed model.