DuNST: Dual Noisy Self Training for Semi-Supervised Controllable Text Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8760–8785 July 9-14, 2023 ©2023 Association for Computational Linguistics DuNST: Dual Noisy Self Training for Semi-Supervised Controllable Text Generation Yuxi Feng1, Xiaoyuan Yi2∗, Xiting Wang2, Laks V.S. Lakshmanan1, Xing Xie2 1The University of British Columbia, Vancouver, Canada 2Microsoft Research Asia, Beijing, China {fyx14, laks}@cs.ubc.ca, {xiaoyuanyi, xitwan, xing.xie}@microsoft.com Abstract Self-training (ST) has prospered again in lan- guage understanding by augmenting the fine- tuning of big pre-trained models when la- beled data is insufficient. However, it remains challenging to incorporate ST into attribute- controllable language generation. Augmented only by self-generated pseudo text, generation models over-exploit the previously learned text space and fail to explore a larger one, suffering from a restricted generalization boundary and limited controllability. In this work, we pro- pose DuNST, a novel ST framework to tackle these problems. DuNST jointly models text generation and classification as a dual process and further perturbs and escapes from the col- lapsed space by adding two kinds of flexible noise. In this way, our model could construct and utilize both pseudo text generated from given labels and pseudo labels predicted from available unlabeled text, which are gradually refined during the ST phase. Theoretically, we show that DuNST can be viewed as enhancing the exploration of the potentially larger real text space while maintaining exploitation, guaran- teeing improved performance. Experiments on three controllable generation tasks show that DuNST significantly boosts control accuracy with comparable generation fluency and diver- sity against several strong baselines. 1 Introduction Recently, Pretrained Language Models (PLM) (Liu et al., 2019; Dong et al., 2019; Radford et al., 2019; Raffel et al., 2020) have shown superior- ity in Natural Language Processing (NLP). How- ever, the ever-growing size of these models de- mands more training data, which destabilizes the fine-tuning of PLMs when labeled data is highly insufficient (Zhang et al., 2021). In this case, Self- training (ST) (Scudder, 1965; Yarowsky, 1995; ∗Work done during Yuxi Feng’s internship at Microsoft Research Asia mentored by Xiaoyuan Yi. Labeled Data Model Unlabeled Data Pseudo Data Step 1: Train a base model Step 2: Synthetic labeling Step 3: Select a subset Top-K Pseudo Data Step 4: Update Model Updated Model Figure 1: Classic Self-training (ST) procedure. ST trains a base classifier on a small labeled dataset, then itera- tively predicts pseudo labels for unlabeled data to aug- ment the original labeled training set, and finally fits the model to the augmented dataset. Grandvalet and Bengio, 2004), a classical semi- supervised paradigm, has come to the fore again. As depicted in Fig. 1, ST produces pseudo labels for text using a classifier and then retrains the clas- sifier with augmented data in an iterative process. By this means, ST utilizes massive unlabeled text to denoise the pseudo-annotated neighbors and im- prove the generalization on real data (Wei et al., 2021; Zhang et al., 2022), boosting various Natural Language Understanding (NLU) tasks (Mukherjee and Hassan Awadallah, 2020; Li et al., …

Content truncated to 500 words