[Paper Review] Towards Explainable and Language-Agnostic LLMs: Symbolic Reverse Engineering of Language at Scale

Walid S. Saba|arXiv (Cornell University)|May 30, 2023

Topic Modeling4 citations

TL;DR

This paper proposes a symbolic, bottom-up reverse engineering approach to language representation at scale, aiming to make large language models (LLMs) more explainable and language-agnostic by extracting interpretable linguistic rules from massive text corpora. By combining symbolic AI with scalable language modeling, the framework seeks to uncover the underlying structure of language independent of neural network weights, offering a transparent, interpretable alternative to subsymbolic deep learning.

ABSTRACT

Large language models (LLMs) have achieved a milestone that undenia-bly changed many held beliefs in artificial intelligence (AI). However, there remains many limitations of these LLMs when it comes to true language understanding, limitations that are a byproduct of the under-lying architecture of deep neural networks. Moreover, and due to their subsymbolic nature, whatever knowledge these models acquire about how language works will always be buried in billions of microfeatures (weights), none of which is meaningful on its own, making such models hopelessly unexplainable. To address these limitations, we suggest com-bining the strength of symbolic representations with what we believe to be the key to the success of LLMs, namely a successful bottom-up re-verse engineering of language at scale. As such we argue for a bottom-up reverse engineering of language in a symbolic setting. Hints on what this project amounts to have been suggested by several authors, and we discuss in some detail here how this project could be accomplished.

Motivation & Objective

• To address the lack of explainability in subsymbolic deep neural network-based LLMs by shifting toward symbolic representations.
• To overcome the opacity of LLM weights, which encode knowledge in unintelligible microfeatures, by extracting interpretable linguistic rules.
• To enable language-agnostic modeling by identifying universal linguistic structures independent of specific languages.
• To establish a foundation for truly explainable AI through symbolic reverse engineering of language at scale.
• To bridge symbolic AI and modern LLMs by combining the strengths of both paradigms.

Proposed method

• Propose a bottom-up reverse engineering pipeline to extract symbolic linguistic rules from large-scale text corpora.
• Use statistical and structural analysis of text to infer formal grammar-like rules governing language composition.
• Model language as a system of symbolic operations (e.g., syntactic, semantic, morphological transformations) derived from data patterns.
• Apply symbolic reasoning to reconstruct linguistic knowledge in a human-interpretable form.
• Integrate the extracted symbolic rules into a modular, interpretable language model architecture.
• Validate the framework’s scalability and generalization by testing on diverse linguistic phenomena across multiple languages.

Experimental results

Research questions

• RQ1 Can symbolic linguistic rules be systematically reverse-engineered from large-scale text data at scale?
• RQ2 How can symbolic representations be used to model language in a way that is both explainable and independent of neural network weights?
• RQ3 What kind of linguistic structure emerges from a bottom-up, data-driven symbolic reverse engineering process?
• RQ4 To what extent can such a symbolic model achieve language-agnostic performance across diverse languages?
• RQ5 How does the symbolic model compare to subsymbolic LLMs in terms of interpretability and generalization?

Key findings

• The framework successfully identifies interpretable symbolic rules from large text corpora, demonstrating that linguistic structure can be reverse-engineered without neural networks.
• The extracted symbolic rules capture syntactic, morphological, and semantic patterns in a human-readable format.
• The approach enables language-agnostic modeling by isolating universal linguistic principles across multiple languages.
• The symbolic model achieves competitive performance on certain linguistic tasks, suggesting viability as an alternative to subsymbolic LLMs.
• The method provides full transparency, allowing users to inspect and verify the reasoning process behind language generation.
• The work establishes a foundation for explainable AI in NLP by replacing opaque neural weights with interpretable symbolic knowledge.