Mutual Exclusivity Training and Primitive Augmentation to Induce Compositionality

TL;DR

This paper introduces mutual exclusivity training and primitive augmentation techniques to improve the compositional generalization of seq2seq models, demonstrating significant empirical gains on standard datasets.

Contribution

It proposes two novel methods addressing memorization and bias issues in seq2seq models, enhancing their systematic generalization capabilities.

Findings

Significant performance improvements on SCAN and COGS datasets.

Mutual exclusivity training reduces incorrect generalizations.

Primitive augmentation diversifies argument structures effectively.

Abstract

Recent datasets expose the lack of the systematic generalization ability in standard sequence-to-sequence models. In this work, we analyze this behavior of seq2seq models and identify two contributing factors: a lack of mutual exclusivity bias (i.e., a source sequence already mapped to a target sequence is less likely to be mapped to other target sequences), and the tendency to memorize whole examples rather than separating structures from contents. We propose two techniques to address these two issues respectively: Mutual Exclusivity Training that prevents the model from producing seen generations when facing novel, unseen examples via an unlikelihood-based loss; and prim2primX data augmentation that automatically diversifies the arguments of every syntactic function to prevent memorizing and provide a compositional inductive bias without exposing test-set data. Combining these two techniques, we show substantial empirical improvements using standard sequence-to-sequence models (LSTMs and Transformers) on two widely-used compositionality datasets: SCAN and COGS. Finally, we provide analysis characterizing the improvements as well as the remaining challenges, and provide detailed ablations of our method. Our code is available at https://github.com/owenzx/met-primaug