Learning compositionally through attentive guidance

TL;DR

This paper introduces Attentive Guidance, a mechanism that helps sequence-to-sequence models learn more compositional solutions, improving their generalization and addressing the challenge of systematic compositionality in neural networks.

Contribution

The paper proposes Attentive Guidance to enhance the compositional capabilities of neural sequence models without additional components.

Findings

Guided models find more compositional solutions than vanilla models.

Guided models generalize well even with distribution shifts.

Vanilla models tend to overfit the training distribution.

Abstract

While neural network models have been successfully applied to domains that require substantial generalisation skills, recent studies have implied that they struggle when solving the task they are trained on requires inferring its underlying compositional structure. In this paper, we introduce Attentive Guidance, a mechanism to direct a sequence to sequence model equipped with attention to find more compositional solutions. We test it on two tasks, devised precisely to assess the compositional capabilities of neural models, and we show that vanilla sequence to sequence models with attention overfit the training distribution, while the guided versions come up with compositional solutions that fit the training and testing distributions almost equally well. Moreover, the learned solutions generalise even in cases where the training and testing distributions strongly diverge. In this way, we demonstrate that sequence to sequence models are capable of finding compositional solutions without requiring extra components. These results helps to disentangle the causes for the lack of systematic compositionality in neural networks, which can in turn fuel future work.