Towards Low-Resource Semi-Supervised Dialogue Generation with Meta-Learning

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4123–4128 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4123 Towards Low-Resource Semi-Supervised Dialogue Generation with Meta-Learning Yi Huang, Junlan Feng, Shuo Ma, Xiaoyu Du, Xiaoting Wu JIUTIAN Team, China Mobile Research {huangyi, fengjunlan}@chinamobile.com, {mashuo, duxiaoyu, wuxiaoting}@chinamobile.com Abstract In this paper, we propose a meta-learning based semi-supervised explicit dialogue state tracker (SEDST) for neural dialogue genera- tion, denoted as MEDST. Our main motivation is to further bridge the chasm between the need for high accuracy dialogue state tracker and the common reality that only scarce annotat- ed data is available for most real-life dialogue tasks. Specifically, MEDST has two core step- s: meta-training with adequate unlabelled data in an automatic way and meta-testing with a few annotated data by supervised learning. In particular, we enhance SEDST via entropy reg- ularization, and investigate semi-supervised learning frameworks based on model-agnostic meta-learning (MAML) that are able to re- duce the amount of required intermediate s- tate labelling. We find that by leveraging un-annotated data in meta-way instead, the amount of dialogue state annotations can be reduced below 10% while maintaining equiva- lent system performance. Experimental result- s show MEDST outperforms SEDST substan- tially by 18.7% joint goal accuracy and 14.3% entity match rate on the KVRET corpus with 2% labelled data in semi-supervision. 1 Introduction Task-oriented dialogue systems (Young et al., 2013) are designed to help users to achieve specific goals such as restaurant reservation or navigation inquiry. In recent years, fully neural end-to-end architec- tures usually take the sequence-to-sequence (Se- q2Seq) (Sutskever et al., 2014) model to generate dialogue responses from the user inputs and contex- t history (Eric and Manning, 2017; Madotto et al., 2018; Wen et al., 2018; Qin et al., 2019; Wu et al., 2019). Neural generative models for task-oriented dialogue systems have achieved promising perfor- mance on generation tasks if given a huge training dataset and detailed annotations (Zhao et al., 2017; R , , t -1 R t U ‘t ‘ ‘ t R t S t S Response Decoder Copyflow Copyflow Copyflow Auto-Encoder unsupervised Posterior Prior Posterior data Prior data Internal module R t -1, t U R , , t -1 R t t U Belief Span Decoder Encoder Figure 1: Network architecture of SEDST model. Lei et al., 2018; Zhang et al., 2019). Arguably, high-quality intermediate labels play a key role in terms of obtaining satisfactory results in this line of tasks (Lei et al., 2018). However, collecting these labels is often the bottleneck in dataset creation, as the process is expensive and time-consuming, requiring domain and expert knowledge (Asri et al., 2017), holding back development in the area of dia- logue systems and greatly limiting their application in real-world settings. Various approaches (Kannan et al., 2018; Tseng et al., 2019; Yin et al., 2019; Chang et al., 2019; Peng et al., 2020) have been proposed recently to conquer this challenge, semi-supervised explicit dialogue state tracker (SEDST) (Jin et al., 2018)1 is …

Content truncated to 500 words