Efficient Pre-training of Masked Language Model via Concept-based Curriculum Masking

Mingyu Lee, Jun-Hyung Park, Junho Kim, Kang-Min Kim, SangKeun Lee


Abstract
Self-supervised pre-training has achieved remarkable success in extensive natural language processing tasks. Masked language modeling (MLM) has been widely used for pre-training effective bidirectional representations but comes at a substantial training cost. In this paper, we propose a novel concept-based curriculum masking (CCM) method to efficiently pre-train a language model. CCM has two key differences from existing curriculum learning approaches to effectively reflect the nature of MLM. First, we introduce a novel curriculum that evaluates the MLM difficulty of each token based on a carefully-designed linguistic difficulty criterion. Second, we construct a curriculum that masks easy words and phrases first and gradually masks related ones to the previously masked ones based on a knowledge graph. Experimental results show that CCM significantly improves pre-training efficiency. Specifically, the model trained with CCM shows comparative performance with the original BERT on the General Language Understanding Evaluation benchmark at half of the training cost.
Anthology ID:
2022.emnlp-main.502
Volume:
Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing
Month:
December
Year:
2022
Address:
Abu Dhabi, United Arab Emirates
Editors:
Yoav Goldberg, Zornitsa Kozareva, Yue Zhang
Venue:
EMNLP
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
7417–7427
Language:
URL:
https://aclanthology.org/2022.emnlp-main.502
DOI:
10.18653/v1/2022.emnlp-main.502
Bibkey:
Cite (ACL):
Mingyu Lee, Jun-Hyung Park, Junho Kim, Kang-Min Kim, and SangKeun Lee. 2022. Efficient Pre-training of Masked Language Model via Concept-based Curriculum Masking. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7417–7427, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
Cite (Informal):
Efficient Pre-training of Masked Language Model via Concept-based Curriculum Masking (Lee et al., EMNLP 2022)
Copy Citation:
PDF:
https://aclanthology.org/2022.emnlp-main.502.pdf