Is BERT a Cross-Disciplinary Knowledge Learner? A Surprising Finding of Pre-trained Models’ Transferability

Wei-Tsung Kao, Hung-yi Lee


Abstract
This paper investigates whether the power of the models pre-trained on text data, such as BERT, can be transferred to general token sequence classification applications. To verify pre-trained models’ transferability, we test the pre-trained models on text classification tasks with meanings of tokens mismatches, and real-world non-text token sequence classification data, including amino acid, DNA, and music. We find that even on non-text data, the models pre-trained on text converge faster, perform better than the randomly initialized models, and only slightly worse than the models using task-specific knowledge. We also find that the representations of the text and non-text pre-trained models share non-trivial similarities.
Anthology ID:
2021.findings-emnlp.189
Volume:
Findings of the Association for Computational Linguistics: EMNLP 2021
Month:
November
Year:
2021
Address:
Punta Cana, Dominican Republic
Editors:
Marie-Francine Moens, Xuanjing Huang, Lucia Specia, Scott Wen-tau Yih
Venue:
Findings
SIG:
SIGDAT
Publisher:
Association for Computational Linguistics
Note:
Pages:
2195–2208
Language:
URL:
https://aclanthology.org/2021.findings-emnlp.189
DOI:
10.18653/v1/2021.findings-emnlp.189
Bibkey:
Cite (ACL):
Wei-Tsung Kao and Hung-yi Lee. 2021. Is BERT a Cross-Disciplinary Knowledge Learner? A Surprising Finding of Pre-trained Models’ Transferability. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2195–2208, Punta Cana, Dominican Republic. Association for Computational Linguistics.
Cite (Informal):
Is BERT a Cross-Disciplinary Knowledge Learner? A Surprising Finding of Pre-trained Models’ Transferability (Kao & Lee, Findings 2021)
Copy Citation:
PDF:
https://aclanthology.org/2021.findings-emnlp.189.pdf
Software:
 2021.findings-emnlp.189.Software.zip
Video:
 https://aclanthology.org/2021.findings-emnlp.189.mp4
Data
GLUE