Harnessing Large Language Models for Rule Generation in Specialist Expert Systems: A Human-Verified Approach

 Abstract:

Expert systems have proven invaluable in fields that require precision and transparency, but their expansion is often limited by the sheer scale of rule creation required to cover specialized domains. This paper posits that large language models (LLMs) are not suited to replace expert systems but can serve as powerful tools for generating the vast number of rules needed to scale them. We propose a hybrid approach where LLMs generate potential rules for expert systems, which are then verified by human experts before integration. This methodology maintains the verifiability and trustworthiness of expert systems while leveraging LLMs for efficient rule creation.

Introduction:

Expert systems have long been the backbone of decision support in highly specialized fields such as healthcare, law, and finance. These systems rely on rule-based frameworks that mirror expert decision-making processes, ensuring transparency, consistency, and accountability. And because their outputs are based on rules, an expert system can fully explain why it is making a specific recommendation. This is crucial in any situation where lives are on the line and we need verifiable human safety rated systems.  However, expanding these systems to accommodate new and more complex domains is an immense challenge due to the exponential growth of rules and exceptions.

In contrast, large language models (LLMs) are designed to process and generate human-like text based on patterns in vast amounts of data. They excel at identifying patterns and generating insights but suffer from issues like hallucination, where the model confidently produces false or misleading information. While LLMs are powerful, they lack the precision, verifiability, and determinism required in expert systems.

This paper argues that while LLMs should not replace expert systems or be trusted to give outputs where there is a risk of harm, they can play a crucial supporting role in generating the massive rule sets needed for specialist expert systems. By relying on LLMs for rule generation and humans for verification, we create a robust and scalable approach to building expert systems that can evolve to meet the demands of complex fields while maintaining their reliability and verifiability. 

Background:

Expert systems operate by applying rules to specific data inputs to deliver consistent and explainable decisions. These rules are traditionally created by human experts, often through a time-consuming process of knowledge acquisition, encoding, and verification. However, as domains expand in complexity, the need for new rules grows exponentially. Scaling expert systems to handle these large rule sets has proven challenging due to the combinatorial explosion of rules and exceptions.

Large language models like GPT have demonstrated an ability to process and generate natural language from diverse sources of data. While these models lack the structure and determinism of expert systems, they are highly capable of identifying potential rules by analyzing vast amounts of unstructured text. However, their outputs often require human verification due to the risk of generating false or nonsensical information.

In this paper, we explore a novel approach where LLMs generate the bulk of the rules required for an expert system, and human experts verify and curate those rules before they are integrated into the expert system’s knowledge base.

LLMs as Rule Generation Engines:

LLMs are well-suited for generating rules because they can quickly process and synthesize large amounts of data across various domains. In specialized fields such as medical diagnostics or legal reasoning, there are millions of potential rules and exceptions that could be encoded into an expert system. Relying solely on humans to develop these rules would be prohibitively time-consuming and costly.

By leveraging LLMs for rule generation, we can automate the initial phase of rule creation. LLMs can scan domain-specific literature, guidelines, and case studies to propose candidate rules, capturing nuances and exceptions that might be missed by human experts. These rules, though not immediately ready for implementation, form the basis of a scalable knowledge base that can then be refined by human experts.

Human Verification and Integration:

While LLMs can generate potential rules at scale, their outputs are not inherently trustworthy. This is where the role of human verification becomes critical. Expert systems are used in fields where precision is paramount, and any error could lead to significant consequences. Thus, before any LLM-generated rule is incorporated into an expert system, it must undergo a process of human validation. Domain experts review, modify, and approve the rules to ensure they are accurate and applicable to real-world scenarios.

The human-in-the-loop approach ensures that the expert system remains verifiable and trustworthy while benefiting from the scale that LLMs offer in rule generation. Experts can evaluate each proposed rule, rejecting those that are incorrect or irrelevant and refining those that are valid. By maintaining this level of human oversight, we can ensure that the final rule set upholds the rigor required in high-stakes domains.

Mitigating Hallucination Risks:

One of the key challenges in using LLMs for rule generation is the risk of hallucinations—where the model generates information that appears valid but is factually incorrect or nonsensical. To mitigate this risk, several strategies can be employed:

1. Source Verification: LLMs should generate rules alongside references to their sources, enabling experts to trace the origin of each rule. This traceability ensures that the rule is based on verifiable knowledge, not speculative or erroneous data.
2. Human Review: No rule generated by the LLM is integrated into the expert system without human review. Domain experts must cross-check each rule against authoritative sources and their own expertise before it is approved for use.
3. Rule-Based Constraints: LLM-generated rules can be filtered through the existing rule framework of the expert system to ensure consistency with previously verified rules. This prevents contradictory or illogical rules from entering the system.

By combining LLM-generated rule proposals with rigorous human verification, we can mitigate hallucination risks while benefiting from the LLM’s ability to process and synthesize information at scale.

Case Studies:

1. Medical Diagnosis System: A medical expert system for rare diseases requires thousands of rules to handle various symptoms, tests, and treatment options. LLMs are tasked with generating preliminary rules by analyzing medical journals, guidelines, and case studies. Human medical experts then review these rules, refining and validating them before they are integrated into the system. The result is an expanded expert system capable of diagnosing a wider range of conditions while maintaining its accuracy and reliability.
2. Legal Reasoning System: In legal expert systems, the complexity of case law, statutes, and legal precedents creates a need for thousands of rules to cover every possible scenario. LLMs generate potential rules by analyzing court cases and legal texts, proposing new rules for handling complex legal reasoning. Legal experts review these rules, ensuring they align with current legal standards before integrating them into the system. This approach enables the system to evolve in tandem with legal developments while preserving its credibility.

Conclusion:

While large language models offer remarkable capabilities for generating knowledge, they lack the precision and verifiability required to operate as expert systems in specialized domains when lives are on the line. However, their ability to process and generate vast rule sets makes them invaluable in the development and expansion of expert systems. By combining the rule-generation capacity of LLMs with human verification, we can scale expert systems to handle increasingly complex domains while maintaining the trust and reliability that make them essential in fields like medicine, law, and finance.

The future of expert systems lies not in replacing them with LLMs but in using LLMs as tools to assist in their growth. With human oversight, LLM-generated rules can transform expert systems into more powerful, scalable, and adaptable decision-making engines, capable of keeping pace with the growing complexity of specialized domains.