Zero and Few-Shot Localization of Task-Oriented Dialogue Agents with a Distilled Representation

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 886–901 May 2-6, 2023 ©2023 Association for Computational Linguistics Zero and Few-Shot Localization of Task-Oriented Dialogue Agents with a Distilled Representation Mehrad Moradshahi Sina J. Semnani Computer Science DepartmentMonica S. Lam Stanford University Stanford, CA {mehrad,sinaj,lam}@cs.stanford.edu Abstract Task-oriented Dialogue (ToD) agents are mostly limited to a few widely-spoken lan- guages, mainly due to the high cost of acquir- ing training data for each language. Existing low-cost approaches that rely on cross-lingual embeddings or naive machine translation sac- rifice a lot of accuracy for data efficiency, and largely fail in creating a usable dialogue agent. We propose automatic methods that use ToD training data in a source language to build a high-quality functioning dialogue agent in an- other target language that has no training data (i.e. zero-shot) or a small training set (i.e. few- shot). Unlike most prior work in cross-lingual ToD that only focuses on Dialogue State Track- ing (DST), we build an end-to-end agent. We show that our approach closes the accu- racy gap between few-shot and existing full- shot methods for ToD agents. We achieve this by (1) improving the dialogue data repre- sentation, (2) improving entity-aware machine translation, and (3) automatic filtering of noisy translations. We evaluate our approach on the recent bilin- gual dialogue dataset BiToD. In Chinese to English transfer, in the zero-shot setting, our method achieves 46.7% and 22.0% in Task Success Rate (TSR) and Dialogue Success Rate (DSR) respectively. In the few-shot set- ting where 10% of the data in the target lan- guage is used, we improve the state-of-the-art by 15.2% and 14.0%, coming within 5% of full-shot training.1 1 Introduction While dialogue agents in various forms have be- come commonplace in parts of the world, their lack of support for most human languages has pre- vented access to the benefits they provide for much of the world. Commercial virtual assistants for 1Code can be accessed at https://github.com/ stanford-oval/dialogues example, only support a handful of languages, as extending their functionality to each new language is extremely costly, partially due to the need for collecting new annotated training data in that lan- guage. In recent years, several non-English task- oriented dialogue (ToD) datasets have been created; they are either collected from scratch such as Ri- SAWOZ (Quan et al., 2020) and CrossWOZ (Zhu et al., 2020), paraphrased from synthetic sentences by crowdworkers such as BiToD (Lin et al., 2021), or manually translated from another language (Li et al., 2021b). All of these approaches are labor- intensive, expensive, and time-consuming; such investment is unlikely to be made for less widely spoken languages. Cross-lingual transfer, i.e. using training data from other languages to build a dialogue agent for a specific language, seems especially appealing. An emerging line of work has employed machine trans- lation of training data, and multilingual pre-trained neural networks to tackle this task (Sherborne et al., 2020; Li et al., 2021a; Moradshahi et al., 2023). However, …

Content truncated to 500 words