Compositional Structured Explanation Generation with Dynamic Modularized Reasoning

Xiyan Fu, Anette Frank


Abstract
In this work, we propose a new task, compositional structured explanation generation (CSEG), to facilitate research on compositional generalization in reasoning. Despite the success of language models in solving reasoning tasks, their compositional generalization capabilities are under-researched. Our new CSEG task tests a model’s ability to generalize from generating entailment trees with a limited number of inference steps – to more steps, focusing on the length and shapes of entailment trees. CSEG is challenging in requiring both reasoning and compositional generalization abilities, and by being framed as a generation task. Besides the CSEG task, we propose a new dynamic modularized reasoning model, MORSE, that factorizes the inference process into modules, where each module represents a functional unit. We adopt modularized self-attention to dynamically select and route inputs to dedicated heads, which specializes them to specific functions. Using CSEG, we compare MORSE to models from prior work. Our analyses show that the task is challenging, but that the dynamic reasoning modules of MORSE are effective, showing competitive compositional generalization abilities in a generation setting.
Anthology ID:
2024.starsem-1.31
Volume:
Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (*SEM 2024)
Month:
June
Year:
2024
Address:
Mexico City, Mexico
Editors:
Danushka Bollegala, Vered Shwartz
Venue:
*SEM
SIG:
SIGLEX
Publisher:
Association for Computational Linguistics
Note:
Pages:
385–401
Language:
URL:
https://aclanthology.org/2024.starsem-1.31
DOI:
10.18653/v1/2024.starsem-1.31
Bibkey:
Cite (ACL):
Xiyan Fu and Anette Frank. 2024. Compositional Structured Explanation Generation with Dynamic Modularized Reasoning. In Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (*SEM 2024), pages 385–401, Mexico City, Mexico. Association for Computational Linguistics.
Cite (Informal):
Compositional Structured Explanation Generation with Dynamic Modularized Reasoning (Fu & Frank, *SEM 2024)
Copy Citation:
PDF:
https://aclanthology.org/2024.starsem-1.31.pdf