[Paper Review] Compositional Generalization by Learning Analytical Expressions

Qian Liu, Shengnan An+7 more|arXiv (Cornell University)|Jun 18, 2020

Topic Modeling19 citations

TL;DR

This paper proposes a memory-augmented neural model, Composer-Solver, that achieves compositional generalization by learning analytical expressions through hierarchical reinforcement learning. By combining symbolic function slots with neural modules, it attains 100% accuracy on all SCAN benchmark tasks, significantly outperforming prior methods in zero-shot compositionality.

ABSTRACT

Compositional generalization is a basic and essential intellective capability of human beings, which allows us to recombine known parts readily. However, existing neural network based models have been proven to be extremely deficient in such a capability. Inspired by work in cognition which argues compositionality can be captured by variable slots with symbolic functions, we present a refreshing view that connects a memory-augmented neural model with analytical expressions, to achieve compositional generalization. Our model consists of two cooperative neural modules, Composer and Solver, fitting well with the cognitive argument while being able to be trained in an end-to-end manner via a hierarchical reinforcement learning algorithm. Experiments on the well-known benchmark SCAN demonstrate that our model seizes a great ability of compositional generalization, solving all challenges addressed by previous works with 100% accuracies.

Motivation & Objective

• To address the critical limitation of neural networks in compositional generalization—recombining known components in novel ways.
• To bridge symbolic cognition principles with neural network training by modeling compositionality through variable slots and analytical functions.
• To develop a trainable, end-to-end framework that mimics human-like recombination of known parts.

Proposed method

• The model employs two cooperative neural modules: Composer, which generates analytical expressions using symbolic function slots, and Solver, which evaluates these expressions.
• The system is trained via a hierarchical reinforcement learning algorithm that optimizes both expression generation and execution.
• Variable slots in the analytical expressions represent placeholders for known components, enabling recombination across unseen combinations.
• The framework integrates memory augmentation to store and retrieve intermediate symbolic representations during expression construction.
• The training process uses reward shaping to guide the model toward correct compositional generalization.
• The method is end-to-end differentiable and scalable to complex, compositional tasks.

Experimental results

Research questions

• RQ1 Can a neural model learn to generalize compositionally by representing tasks as analytical expressions with symbolic slots?
• RQ2 How well does the Composer-Solver framework perform on zero-shot compositional generalization compared to prior approaches?
• RQ3 To what extent can hierarchical reinforcement learning enable the model to discover and execute correct analytical expressions?

Key findings

• The model achieves 100% accuracy on all compositional generalization challenges in the SCAN benchmark, outperforming previous methods.
• The model generalizes effectively to unseen combinations of known components, demonstrating strong zero-shot generalization.
• The hierarchical reinforcement learning approach enables stable and effective training of the Composer-Solver system.
• The integration of symbolic function slots with neural modules allows for interpretable and generalizable reasoning.
• The model successfully captures the cognitive principle of compositionality through analytical expression learning.