Dyna-bAbI: unlocking bAbI’s potential with dynamic synthetic benchmarking

Ronen Tamari, Kyle Richardson, Noam Kahlon, Aviad Sar-shalom, Nelson F. Liu, Reut Tsarfaty, Dafna Shahaf


Abstract
While neural language models often perform surprisingly well on natural language understanding (NLU) tasks, their strengths and limitations remain poorly understood. Controlled synthetic tasks are thus an increasingly important resource for diagnosing model behavior. In this work we focus on story understanding, a core competency for NLU systems. However, the main synthetic resource for story understanding, the bAbI benchmark, lacks such a systematic mechanism for controllable task generation. We develop Dyna-bAbI, a dynamic framework providing fine-grained control over task generation in bAbI. We demonstrate our ideas by constructing three new tasks requiring compositional generalization, an important evaluation setting absent from the original benchmark. We tested both special-purpose models developed for bAbI as well as state-of-the-art pre-trained methods, and found that while both approaches solve the original tasks (99% accuracy), neither approach succeeded in the compositional generalization setting, indicating the limitations of the original training data. We explored ways to augment the original data, and found that though diversifying training data was far more useful than simply increasing dataset size, it was still insufficient for driving robust compositional generalization (with 70% accuracy for complex compositions). Our results underscore the importance of highly controllable task generators for creating robust NLU systems through a virtuous cycle of model and data development.
Anthology ID:
2022.starsem-1.9
Volume:
Proceedings of the 11th Joint Conference on Lexical and Computational Semantics
Month:
July
Year:
2022
Address:
Seattle, Washington
Editors:
Vivi Nastase, Ellie Pavlick, Mohammad Taher Pilehvar, Jose Camacho-Collados, Alessandro Raganato
Venue:
*SEM
SIG:
SIGLEX
Publisher:
Association for Computational Linguistics
Note:
Pages:
101–122
Language:
URL:
https://aclanthology.org/2022.starsem-1.9
DOI:
10.18653/v1/2022.starsem-1.9
Bibkey:
Cite (ACL):
Ronen Tamari, Kyle Richardson, Noam Kahlon, Aviad Sar-shalom, Nelson F. Liu, Reut Tsarfaty, and Dafna Shahaf. 2022. Dyna-bAbI: unlocking bAbI’s potential with dynamic synthetic benchmarking. In Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 101–122, Seattle, Washington. Association for Computational Linguistics.
Cite (Informal):
Dyna-bAbI: unlocking bAbI’s potential with dynamic synthetic benchmarking (Tamari et al., *SEM 2022)
Copy Citation:
PDF:
https://preview.aclanthology.org/naacl-24-ws-corrections/2022.starsem-1.9.pdf
Data
SQuAD