GlobalWoZ: Globalizing MultiWoZ to Develop Multilingual Task-Oriented Dialogue Systems

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1639 - 1657 May 22-27, 2022 c⃝2022 Association for Computational Linguistics GlobalWoZ: Globalizing MultiWoZ to Develop Multilingual Task-Oriented Dialogue Systems Bosheng Ding∗1,2 Junjie Hu3 Lidong Bing1 Sharifah Mahani Aljunied1 Shafiq Joty2 Luo Si1 Chunyan Miao2 1DAMO Academy, Alibaba Group 2Nanyang Technological University, Singapore 3University of Wisconsin-Madison {bosheng.ding, l.bing, mahani.aljunied, luo.si}@alibaba-inc.com junjie.hu@wisc.edu, {srjoty, ascymiao}@ntu.edu.sg Abstract Over the last few years, there has been a move towards data curation for multilingual task- oriented dialogue (ToD) systems that can serve people speaking different languages. However, existing multilingual ToD datasets either have a limited coverage of languages due to the high cost of data curation, or ignore the fact that dialogue entities barely exist in countries speaking these languages. To tackle these lim- itations, we introduce a novel data curation method that generates GlobalWoZ — a large- scale multilingual ToD dataset globalized from an English ToD dataset for three unexplored use cases of multilingual ToD systems. Our method is based on translating dialogue tem- plates and filling them with local entities in the target-language countries. Besides, we ex- tend the coverage of target languages to 20 lan- guages. We will release our dataset and a set of strong baselines to encourage research on multilingual ToD systems for real use cases.1 1 Introduction One of the fundamental objectives in pursuit of ar- tificial intelligence is to enable machines with the ability to intelligently communicate with human in natural languages, with one of the widely-heralded applications being the task-oriented dialogue (ToD) systems (Gupta et al., 2006; Bohus and Rudnicky, 2009). Recently, ToD systems have been success- fully deployed to assist users with accomplishing certain domain-specific tasks such as hotel book- ing, alarm setting or weather query (Eric et al., 2017; Wu et al., 2019; Lin et al., 2020; Zhang et al., 2020), thanks to the joint advent of neural networks and availability of domain-specific data. However, most existing ToD systems are predominately built for English, limiting their service for all of the ∗Bosheng Ding is under the Joint PhD Program between Alibaba and Nanyang Technological University. 1Our code is available at https://ntunlpsg. github.io/project/globalwoz/. world’s citizens. The reason of this limitation lies in the stark lack of high-quality multilingual ToD datasets due to the high expense and challenges of human annotation (Razumovskaia et al., 2021). One solution to this is annotating conversations in other languages from scratch, e.g., CrossWoZ (Zhu et al., 2020) and BiToD (Lin et al., 2021). However, these methods involve expensive human efforts for dialogue collection in the other lan- guages, resulting in a limited language/domain cov- erage. The other major line of work focused on translating an existing English ToD dataset into target languages by professional human transla- tors (Upadhyay et al., 2018; Schuster et al., 2019; van der Goot et al., 2021; Li et al., 2021). Despite the increasing language coverage, these methods simply translated English named entities (e.g., lo- cation, restaurant name) into the …

Content truncated to 500 words