Multi-lingual Transfer Learning for Intent Classification
Yutong Zhou · 2021
Intent Classification (IC) is critical in Natural Language Understanding (NLU) in dialogue systems in which separate models are usually required for different languages. However, the maintenance and deployment of multiple models for a number of languages are costly and inefficient. Moreover, there are issues in the data collection of different languages for training these models, such as imbalanced feature distribution. In this paper, we develop and experiment with monolingual and language-agnostic joint models in IC. In these experiments, we compare the performances of the two schemes and conclude with the advantages of the language-agnostic model for three languages. In addition, we study language similarities through transfer learning and visualizing word piece embedding projections. The results show that the language-agnostic model performs close to monolingual models with the advantages of being easy to maintain. From transfer learning and embedding visualization, we also confirm that our analysis on language similarity aligns with language family partition.