How is BERT surprised? Layerwise detection of linguistic anomalies

Bai Li, Zining Zhu, Guillaume Thomas, Yang Xu, Frank Rudzicz


Abstract
Transformer language models have shown remarkable ability in detecting when a word is anomalous in context, but likelihood scores offer no information about the cause of the anomaly. In this work, we use Gaussian models for density estimation at intermediate layers of three language models (BERT, RoBERTa, and XLNet), and evaluate our method on BLiMP, a grammaticality judgement benchmark. In lower layers, surprisal is highly correlated to low token frequency, but this correlation diminishes in upper layers. Next, we gather datasets of morphosyntactic, semantic, and commonsense anomalies from psycholinguistic studies; we find that the best performing model RoBERTa exhibits surprisal in earlier layers when the anomaly is morphosyntactic than when it is semantic, while commonsense anomalies do not exhibit surprisal at any intermediate layer. These results suggest that language models employ separate mechanisms to detect different types of linguistic anomalies.
Anthology ID:
2021.acl-long.325
Volume:
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)
Month:
August
Year:
2021
Address:
Online
Venues:
ACL | IJCNLP
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
4215–4228
Language:
URL:
https://aclanthology.org/2021.acl-long.325
DOI:
10.18653/v1/2021.acl-long.325
Bibkey:
Cite (ACL):
Bai Li, Zining Zhu, Guillaume Thomas, Yang Xu, and Frank Rudzicz. 2021. How is BERT surprised? Layerwise detection of linguistic anomalies. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4215–4228, Online. Association for Computational Linguistics.
Cite (Informal):
How is BERT surprised? Layerwise detection of linguistic anomalies (Li et al., ACL 2021)
Copy Citation:
PDF:
https://preview.aclanthology.org/update-css-js/2021.acl-long.325.pdf
Code
 SPOClab-ca/layerwise-anomaly
Data
BLiMP