The Effectiveness of Masked Language Modeling and Adapters for Factual Knowledge Injection

Sondre Wold


Abstract
This paper studies the problem of injecting factual knowledge into large pre-trained language models. We train adapter modules on parts of the ConceptNet knowledge graph using the masked language modeling objective and evaluate the success of the method by a series of probing experiments on the LAMA probe. Mean P@K curves for different configurations indicate that the technique is effective, increasing the performance on sub-sets of the LAMA probe for large values of k by adding as little as 2.1% additional parameters to the original models.
Anthology ID:
2022.textgraphs-1.6
Volume:
Proceedings of TextGraphs-16: Graph-based Methods for Natural Language Processing
Month:
October
Year:
2022
Address:
Gyeongju, Republic of Korea
Editors:
Dmitry Ustalov, Yanjun Gao, Alexander Panchenko, Marco Valentino, Mokanarangan Thayaparan, Thien Huu Nguyen, Gerald Penn, Arti Ramesh, Abhik Jana
Venue:
TextGraphs
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
54–59
Language:
URL:
https://aclanthology.org/2022.textgraphs-1.6
DOI:
Bibkey:
Cite (ACL):
Sondre Wold. 2022. The Effectiveness of Masked Language Modeling and Adapters for Factual Knowledge Injection. In Proceedings of TextGraphs-16: Graph-based Methods for Natural Language Processing, pages 54–59, Gyeongju, Republic of Korea. Association for Computational Linguistics.
Cite (Informal):
The Effectiveness of Masked Language Modeling and Adapters for Factual Knowledge Injection (Wold, TextGraphs 2022)
Copy Citation:
PDF:
https://preview.aclanthology.org/nschneid-patch-4/2022.textgraphs-1.6.pdf
Code
 sondrewold/adapters-mlm-injection
Data
ConceptNetGLUELAMA