Accurate identification of the domain of user commands is a crucial first step in digital voice assistant systems. Consider domain identification as a multi-class classification problem where the inputs are transcribed utterances. Supervised learning is limited by the availability of annotated datasets, an issue that is exacerbated for non-English languages.
In this project, I tackle the problem of domain classification for French, a relatively lower resource language than English. With access to a parallel annotated dataset in French, I set out to compare the performance gap between fully-supervised training in the target language and cross-lingual zero-shot transfer from the source language using massively pre-trained Transformer-based masked language models (PLMs).
The results reinforce that given the recent advancements in PLMs, domain identification is a somewhat trivial task with fully-supervised fine-tuning in the target language achieving near-perfect results (~0.98 F1) and even zero-shot transfer from the source not lagging far behind (~0.95 F1).
See Technical Report for more details on methodology and results.
MTOP: A Comprehensive Multilingual Task-Oriented Semantic Parsing Benchmark (Li et al., EACL 2021) is a multi-lingual, multi-domain task-oriented dialog dataset consisting of synthetic utterances in 6 languages across 11 domains with fine-grained intent labels. Different from the authors' work, this project will focus on the more coarse-grained domain prediction. Dataset with domain labels is available on HuggingFace mteb/mtop_domain.