[1Win4] Poster session 1

[1Win4-35] Investigating the Impact of Sequential Multilingual Training on Classification Accuracy in a Multilingual Model

〇Lingfang Zhang¹, Daisuke Kawahara¹ (1.Waseda University)

Keywords:Natural Language Processing, Multilingual Model, Sequential Learning, Knowledge Transfer, Cross-lingual

This study explores the impact of sequential multilingual training on a multilingual BERT model, focusing on Japanese and Korean news classification tasks. The training is conducted iteratively in two sequences: Japanese to Korean and Korean to Japanese, to examine how classification accuracy changes. Results showed that after initial training on one language, the classification accuracy for the untrained language improved, indicating a cross-lingual knowledge transfer. Furthermore, as the training continued, the performance for both languages consistently improved and ultimately converged to high accuracy, regardless of the training order. This demonstrates that multilingual models can effectively adapt to sequential training across multiple languages, holding previously learned knowledge while acquiring new information. These findings provide insights into the robustness and adaptability of multilingual models in sequential learning scenarios, offering potential strategies for optimizing cross-lingual training in classification tasks.