[Paper Review] Neural-Symbolic Recursive Machine for Systematic Generalization

Qing Li, Yixin Zhu+4 more|arXiv (Cornell University)|Oct 4, 2022

Topic Modeling4 citations

TL;DR

This paper proposes the Neural-Symbolic Recursive Machine (NSR), a modular framework that learns a Grounded Symbol System (GSS) from data to enable systematic generalization in sequence-to-sequence tasks. By integrating neural perception, syntactic parsing, and semantic reasoning through a novel deduction-abduction training algorithm, NSR achieves 100% generalization accuracy on SCAN and PCFG and surpasses prior models by 23% on HINT, demonstrating state-of-the-art performance with strong transferability and reduced reliance on domain-specific knowledge.

ABSTRACT

Current learning models often struggle with human-like systematic generalization, particularly in learning compositional rules from limited data and extrapolating them to novel combinations. We introduce the Neural-Symbolic Recursive Machine (NSR), whose core is a Grounded Symbol System (GSS), allowing for the emergence of combinatorial syntax and semantics directly from training data. The NSR employs a modular design that integrates neural perception, syntactic parsing, and semantic reasoning. These components are synergistically trained through a novel deduction-abduction algorithm. Our findings demonstrate that NSR's design, imbued with the inductive biases of equivariance and compositionality, grants it the expressiveness to adeptly handle diverse sequence-to-sequence tasks and achieve unparalleled systematic generalization. We evaluate NSR's efficacy across four challenging benchmarks designed to probe systematic generalization capabilities: SCAN for semantic parsing, PCFG for string manipulation, HINT for arithmetic reasoning, and a compositional machine translation task. The results affirm NSR's superiority over contemporary neural and hybrid models in terms of generalization and transferability.

Motivation & Objective

• To address the challenge of systematic generalization in neural models, particularly in extrapolating from limited data to novel combinations.
• To develop a unified framework that jointly learns perception, syntax, and semantics with strong inductive biases for compositionality and equivariance.
• To eliminate reliance on handcrafted, domain-specific knowledge by enabling the Grounded Symbol System (GSS) to emerge directly from training data.
• To improve transferability and robustness across diverse sequence-to-sequence tasks, including semantic parsing, string manipulation, arithmetic reasoning, and machine translation.
• To design a differentiable training procedure for non-differentiable symbolic components using a novel deduction-abduction algorithm.

Proposed method

• NSR employs a modular architecture integrating a neural perception module, a transition-based neural dependency parser, and a functional program-based semantic reasoner.
• The Grounded Symbol System (GSS) is learned end-to-end from data, representing symbols as grounded triplets of (input, symbol, output) across tasks.
• A novel deduction-abduction algorithm enables joint training: first, greedy deduction forms an initial GSS; then, top-down abduction refines it by exploring neighboring GSS structures.
• The refined GSS acts as pseudo-supervision to independently train each NSR component, overcoming the challenge of non-differentiable symbolic components.
• The framework embeds inductive biases of compositionality and equivariance, enabling decomposition and recombination of inputs for systematic generalization.
• A probabilistic learning framework supports end-to-end optimization despite the lack of explicit GSS annotations.

Experimental results

Research questions

• RQ1 Can a neural-symbolic model learn a compositional, interpretable symbol system directly from data without domain-specific priors?
• RQ2 How effectively can a model with inductive biases of compositionality and equivariance generalize to unseen combinations in sequence-to-sequence tasks?
• RQ3 Can a joint training procedure overcome the non-differentiability of symbolic components and enable end-to-end learning?
• RQ4 How does NSR’s performance compare to state-of-the-art neural and hybrid models on systematic generalization benchmarks?
• RQ5 To what extent can NSR generalize in real-world, ambiguous tasks such as compositional machine translation?

Key findings

• NSR achieves 100% generalization accuracy on the SCAN benchmark, outperforming all prior models.
• On the PCFG benchmark, NSR attains 100% generalization accuracy, demonstrating perfect extrapolation to novel string manipulation rules.
• On the HINT benchmark, NSR achieves 90.1% average test accuracy, surpassing the previous best model by 23 percentage points.
• NSR achieves 100% generalization accuracy on a compositional machine translation task, matching the performance of NeSS but with a more principled, data-driven symbol system.
• The model shows strong transferability across diverse domains, including semantic parsing, string manipulation, arithmetic reasoning, and real-world translation, with minimal reliance on domain-specific knowledge.
• Ablation studies confirm that the inductive biases of compositionality and equivariance are critical for NSR’s superior generalization performance.