Measuring Compositional Generalization: A Comprehensive Method on Realistic Data
This addresses the lack of comprehensive benchmarks for compositional generalization, which is crucial for evaluating and improving machine learning models in natural language processing and other domains, though it is incremental as it builds on existing methods like SCAN.
The paper tackles the problem of limited compositional generalization in machine learning by introducing a method to construct realistic benchmarks that maximize compound divergence while minimizing atom divergence, and it finds that three tested architectures fail to generalize compositionally with a strong negative correlation between compound divergence and accuracy.
State-of-the-art machine learning methods exhibit limited compositional generalization. At the same time, there is a lack of realistic benchmarks that comprehensively measure this ability, which makes it challenging to find and evaluate improvements. We introduce a novel method to systematically construct such benchmarks by maximizing compound divergence while guaranteeing a small atom divergence between train and test sets, and we quantitatively compare this method to other approaches for creating compositional generalization benchmarks. We present a large and realistic natural language question answering dataset that is constructed according to this method, and we use it to analyze the compositional generalization ability of three machine learning architectures. We find that they fail to generalize compositionally and that there is a surprisingly strong negative correlation between compound divergence and accuracy. We also demonstrate how our method can be used to create new compositionality benchmarks on top of the existing SCAN dataset, which confirms these findings.