The Irrationality of Neural Rationale Models

Yiming Zheng, Serena Booth, Julie Shah, Yilun Zhou


Abstract
Neural rationale models are popular for interpretable predictions of NLP tasks. In these, a selector extracts segments of the input text, called rationales, and passes these segments to a classifier for prediction. Since the rationale is the only information accessible to the classifier, it is plausibly defined as the explanation. Is such a characterization unconditionally correct? In this paper, we argue to the contrary, with both philosophical perspectives and empirical evidence suggesting that rationale models are, perhaps, less rational and interpretable than expected. We call for more rigorous evaluations of these models to ensure desired properties of interpretability are indeed achieved. The code for our experiments is at https://github.com/yimingz89/Neural-Rationale-Analysis.
Anthology ID:
2022.trustnlp-1.6
Volume:
Proceedings of the 2nd Workshop on Trustworthy Natural Language Processing (TrustNLP 2022)
Month:
July
Year:
2022
Address:
Seattle, U.S.A.
Editors:
Apurv Verma, Yada Pruksachatkun, Kai-Wei Chang, Aram Galstyan, Jwala Dhamala, Yang Trista Cao
Venue:
TrustNLP
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
64–73
Language:
URL:
https://aclanthology.org/2022.trustnlp-1.6
DOI:
10.18653/v1/2022.trustnlp-1.6
Bibkey:
Cite (ACL):
Yiming Zheng, Serena Booth, Julie Shah, and Yilun Zhou. 2022. The Irrationality of Neural Rationale Models. In Proceedings of the 2nd Workshop on Trustworthy Natural Language Processing (TrustNLP 2022), pages 64–73, Seattle, U.S.A.. Association for Computational Linguistics.
Cite (Informal):
The Irrationality of Neural Rationale Models (Zheng et al., TrustNLP 2022)
Copy Citation:
PDF:
https://aclanthology.org/2022.trustnlp-1.6.pdf
Code
 yimingz89/neural-rationale-analysis
Data
SST