Abstract
In this paper, we prove that separable negative log-likelihood losses for structured prediction are not necessarily Bayes consistent, that is minimizing these losses may not result in a model that predicts the most probable structure in the data distribution for a given input. This fact opens the question of whether these losses are well-adapted for structured prediction and, if so, why.- Anthology ID:
- 2023.eacl-main.109
- Volume:
- Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics
- Month:
- May
- Year:
- 2023
- Address:
- Dubrovnik, Croatia
- Editors:
- Andreas Vlachos, Isabelle Augenstein
- Venue:
- EACL
- SIG:
- Publisher:
- Association for Computational Linguistics
- Note:
- Pages:
- 1491–1498
- Language:
- URL:
- https://aclanthology.org/2023.eacl-main.109
- DOI:
- 10.18653/v1/2023.eacl-main.109
- Cite (ACL):
- Caio Corro. 2023. On the inconsistency of separable losses for structured prediction. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1491–1498, Dubrovnik, Croatia. Association for Computational Linguistics.
- Cite (Informal):
- On the inconsistency of separable losses for structured prediction (Corro, EACL 2023)
- PDF:
- https://preview.aclanthology.org/finnlp-2volume-ingestion/2023.eacl-main.109.pdf