Self-training Improves Pre-training for Few-shot Learning in Task-oriented Dialog Systems

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1887–1898 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 1887 Self-training Improves Pre-training for Few-shot Learning in Task-oriented Dialog Systems Fei Mi1, Wanhao Zhou2, Fengyu Cai2, Lingjing Kong2, Minlie Huang3 and Boi Faltings2 1Huawei Noah’s Ark Lab 2LIA, EPFL 3CoAI, DCST, Tsinghua University mifei2@huawei.com, aihuang@tsinghua.edu.cn {wanhao.zhou,fengyu.cai,lingjing.kong,boi.faltings}@epfl.ch Abstract As the labeling cost for different modules in task-oriented dialog (ToD) systems is expen- sive, a major challenge is to train different mod- ules with the least amount of labeled data. Re- cently, large-scale pre-trained language mod- els, have shown promising results for few-shot learning in ToD. In this paper, we devise a self- training approach to utilize the abundant unla- beled dialog data to further improve state-of- the-art pre-trained models in few-shot learning scenarios for ToD systems. Specifically, we propose a self-training approach that iteratively labels the most confident unlabeled data to train a stronger Student model. Moreover, a new text augmentation technique (GradAug) is pro- posed to better train the Student by replacing non-crucial tokens using a masked language model. We conduct extensive experiments and present analyses on four downstream tasks in ToD, including intent classification, dialog state tracking, dialog act prediction, and response se- lection. Empirical results demonstrate that the proposed self-training approach consistently improves state-of-the-art pre-trained models (BERT, ToD-BERT) when only a small number of labeled data are available. 1 Introduction Large-scale pre-trained language models, such as BERT (Devlin et al., 2019), UniLM (Dong et al., 2019), GPT (Radford et al., 2018), GPT-2 (Radford et al., 2019), and GPT-3 (Brown et al., 2020), have shown great few-shot or zero-shot learning abilities in various NLP tasks with the help of task-agnostic language knowledge learned via pre-training tasks. In task-oriented dialog (ToD) systems, the labeling cost is very high such that the size of well-labeled data is often small. Therefore, few-shot learning in ToD is important and valuable in many practical applications. Many attempts (Peng et al., 2020b,a; Wu et al., 2020) have been proposed to leverage large-scale pre-trained language models to improve few-shot learning in ToD. Specifically, a model pre- trained on general text corpora is further trained on public ToD datasets. Although the size of labeled data is often small, a practical ToD system de facto has many unlabeled dialog data. Therefore, utilizing unlabeled data to improve a ToD system is practically important. In this paper, we take a semi-supervised self-training (ST) perspective to iteratively train a better Stu- dent model using unlabeled data (Scudder, 1965; Yarowsky, 1995). ST has been successfully applied to a variety of tasks, including image classification (Yalniz et al., 2019; Xie et al., 2020; Zoph et al., 2020), automatic speech classification (Synnaeve et al., 2019; Kahn et al., 2020; Park et al., 2020; Likhomanenko et al., 2020), sequence generation (He et al., 2020), and natural language understand- ing (Du et al., 2020). We are going to study this research question: can self-training provide complementary benefits on top of the strong pre-training …

Content truncated to 500 words