A Good Sample is Hard to Find: Noise Injection Sampling and Self-Training for Neural Language Generation Models
This addresses the challenge of improving semantic fidelity in natural language generation for tasks like dialogue systems, though it is incremental as it builds on existing self-training and data augmentation techniques.
The paper tackles the problem of neural language generation models failing to produce utterances faithful to input meaning representations by proposing a self-training method with noise injection sampling to augment training data. The result shows that even simple models achieve state-of-the-art performance in generating semantically correct utterances, as validated by automatic and human evaluations.
Deep neural networks (DNN) are quickly becoming the de facto standard modeling method for many natural language generation (NLG) tasks. In order for such models to truly be useful, they must be capable of correctly generating utterances for novel meaning representations (MRs) at test time. In practice, even sophisticated DNNs with various forms of semantic control frequently fail to generate utterances faithful to the input MR. In this paper, we propose an architecture agnostic self-training method to sample novel MR/text utterance pairs to augment the original training data. Remarkably, after training on the augmented data, even simple encoder-decoder models with greedy decoding are capable of generating semantically correct utterances that are as good as state-of-the-art outputs in both automatic and human evaluations of quality.