Exploring Multilingual Text Data Distillation

TL;DR

This paper introduces new data distillation techniques for multilingual text classification that improve model training efficiency, generalization across architectures, and fairness across languages.

Contribution

It presents novel multilingual text data distillation methods leveraging language models, addressing challenges of discrete data and cross-architecture generalization.

Findings

Enhanced cross-architecture generalization demonstrated

Improved classification performance on multilingual datasets

Insights into language-specific fairness of data summaries

Abstract

With the rise of deep learning, large datasets and complex models have become common, requiring significant computing power. To address this, data distillation has emerged as a technique to quickly train models with lower memory and time requirements. However, data distillation on text-based datasets hasn't been explored much because of the challenges rising due to its discrete nature. Additionally, existing dataset distillation methods often struggle to generalize to new architectures. In the paper, we propose several data distillation techniques for multilingual text classification datasets using language-model-based learning methods. We conduct experiments to analyze their performance in terms of classification strength, and cross-architecture generalization. Furthermore, we investigate the language-specific fairness of the data summaries generated by these methods. Our approach builds upon existing techniques, enhancing cross-architecture generalization in the text data distillation domain.