On Accurate Evaluation of GANs for Language Generation
This addresses the problem of unreliable evaluation in language generation research, which is crucial for researchers and practitioners in NLP, but it is incremental as it critiques and refines existing methods rather than introducing a new paradigm.
The paper argues that current evaluation practices for GANs in language generation, which rely on n-gram metrics like BLEU and report only best-run scores, are misleading due to sensitivity to initialization and hyperparameters, and it proposes alternative metrics to better assess quality and diversity, finding in experiments that GAN models do not convincingly outperform conventional language models.
Generative Adversarial Networks (GANs) are a promising approach to language generation. The latest works introducing novel GAN models for language generation use n-gram based metrics for evaluation and only report single scores of the best run. In this paper, we argue that this often misrepresents the true picture and does not tell the full story, as GAN models can be extremely sensitive to the random initialization and small deviations from the best hyperparameter choice. In particular, we demonstrate that the previously used BLEU score is not sensitive to semantic deterioration of generated texts and propose alternative metrics that better capture the quality and diversity of the generated samples. We also conduct a set of experiments comparing a number of GAN models for text with a conventional Language Model (LM) and find that neither of the considered models performs convincingly better than the LM.