Exploring Continual Learning of Compositional Generalization in NLI

TL;DR

The paper introduces C2Gen NLI, a continual-learning framework to study how models acquire primitive inferences over time and how this sequential learning affects compositional generalization in natural language inference. It shows that continual learning induces forgetting, hindering unseen compositional inferences, but that memorization-based strategies and carefully ordered learning (e.g., easy-before-hard, dependency-aware primitive ordering) can mitigate forgetting and improve generalization. Through offline CGen and continual C2Gen experiments, plus analyses of primitive recognition versus compositional inference, the work demonstrates when and how continual learning helps, and where it still falls short compared to non-continual training. The findings have practical implications for dynamic, knowledge-updating applications (e.g., Persona Dialogue) and point to future directions in learning-order optimization and curriculum-inspired approaches to enhance continual compositional generalization in NLP.

Abstract

Compositional Natural Language Inference has been explored to assess the true abilities of neural models to perform NLI. Yet, current evaluations assume models to have full access to all primitive inferences in advance, in contrast to humans that continuously acquire inference knowledge. In this paper, we introduce the Continual Compositional Generalization in Inference (C2Gen NLI) challenge, where a model continuously acquires knowledge of constituting primitive inference tasks as a basis for compositional inferences. We explore how continual learning affects compositional generalization in NLI, by designing a continual learning setup for compositional NLI inference tasks. Our experiments demonstrate that models fail to compositionally generalize in a continual scenario. To address this problem, we first benchmark various continual learning algorithms and verify their efficacy. We then further analyze C2Gen, focusing on how to order primitives and compositional inference types and examining correlations between subtasks. Our analyses show that by learning subtasks continuously while observing their dependencies and increasing degrees of difficulty, continual learning can enhance composition generalization ability.

Figures (4)

• Figure 1: Training and testing setup for compositional inference for continual stages $\mathcal{S}_i$, in C$^{2}$Gen. In $\mathcal{S}_1$ we feed various veridicality samples and a few primitive NLI samples. $\mathcal{S}_2$ works analogously.
• Figure 2: Multi-task architecture for compositional generalization evaluation in CGen & C$^{2}$Gen NLI. Task$_{CI}$ and Task$_{P}$ are jointly optimized.
• Figure 3: Changes of learned primitive representations from $\mathcal{S}_1$ to $\mathcal{S}_2$ with different learning orders.
• Figure 4: Persona Dialogue application for C$^2$Gen: NLI verifies the consistency of dialogue turns generated from dynamically updated persona information. We show a profile with new information and compositional inferences using it (in ).