Consistency of Compositional Generalization Across Multiple Levels

• DOI: 10.1609/aaai.v39i5.32492
• Publication Date: 2025-04-11T11:05:40Z
• Authors: Chuanhao Li; Zhen Li; Chenchen Jing; Xiaomeng Fan; Wenbo Ye; Yuwei Wu; Yunde Jia

Compositional generalization is the capability of a model to understand novel compositions composed seen concepts. There are multiple levels including phrase-phrase level, phrase-word and word-word level. Existing methods achieve promising compositional generalization, but consistency across remains unexplored. The refers that should generalize level composition, phrase-word/word-word can be derived from it simultaneously. In this paper, we propose meta-learning based framework, for achieving consistent levels. basic idea progressively learn simple complex consistency. Specifically, divide original training set into validation sets on complexity, introduce meta-weight-nets generate sample weights samples in different sets. To fit order increasing optimize parameters each meta-weight-net independently sequentially multilevel optimization manner. We build GQA-CCG dataset quantitatively evaluate Experimental results visual question answering temporal video grounding, demonstrate effectiveness proposed framework.