PEML: Prototype-enhanced Meta-learning for Multi-lingual Text Classification

Longxuan Ma, Lin Chen, HAOWEI MA, Shengxiang Gao, Zhengtao Yu

Keywords: Multi-lingual text classification, Few-shot learning, Meta-learning, Prototypical Network

Abstract: Multi-lingual text classification is a challenging task in natural language processing, which not only faces language differences between multiple languages but also faces the challenge of scarce annotated data. This paper proposes a prototype-enhanced meta-learning (PEML) method to address the challenges in the multi-lingual text classification task. The PEML method consists of two steps: firstly, to enhance the model's ability to understand multi-lingual samples, we design a multi-lingual label-fusion technique to better map labels from different languages into a unified semantical space; secondly, in response to the problem that class prototypes for support sets are difficult to apply to query sets in meta-learning, we use a query-enhanced technique to associates the prototype vectors of the support set with samples in the query set. After training with our method, the classification model can quickly update the class prototypes to the data distribution of the query set, thereby expanding the model's multi-lingual classification ability from the support set to the unseen query set. Extensive experiments demonstrate that the proposed method significantly outperforms state-of-the-art methods in multi-lingual text classification tasks. The code and data of this paper will be released on GitHub.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 15705