TL;DR
This paper introduces FÆRDXEL, an expert system designed to answer legal questions about Danish traffic law using Datalog and SLD-resolution. It aims to bridge the gap between AI's capabilities and the legal domain's need for explainability and traditional reasoning, offering a traceable inference process for legal professionals.
What Problem Does This Paper Address?
The paper tackles the challenge of integrating AI into the legal domain, which is often met with resistance due to the inherent requirement for explainable judgments and the legal profession's tradition of argumentation. While modern machine learning models (like deep neural networks) can provide answers, their reasoning is often opaque, making it difficult to satisfy legal requirements for clear explanations. Traditional expert systems, on the other hand, offer traceable reasoning but are notoriously time-consuming to build and can be difficult for non-experts to use. The authors propose an expert system as a more suitable AI approach for this domain, focusing on providing clear, step-by-step explanations for legal queries.
Methods
The core of FÆRDXEL is a Datalog encoding of Danish traffic laws, which serves as its knowledge base. Queries are processed using SLD-resolution, a standard inference algorithm for logic programming. The novelty lies in how the system leverages the trace of this resolution process to generate a tree-like structure of explanations for its answers. This allows users to follow the system's reasoning from the initial query down to the specific facts and rules used. The system also incorporates a user interface that translates Datalog predicates into natural language (NLT) to improve usability. Additionally, the authors employ a "reification" principle in their encoding, treating concepts like "driving a car" as objects to simplify the predicate set. Syntactic sugar is used to allow for more natural expression of disjunctions and conjunctions within the rules, which are then converted to standard Conjunctive Normal Form (CNF).
Datasets
Not discussed in the paper. The knowledge base is derived by hand from "Færdselsloven med Kommentarer," a book covering the practical interpretation of Danish traffic law.
Key Insights and Ideas
- Expert Systems as a Bridge: The paper argues that traditional expert systems, despite their development challenges, are better suited for domains like law that demand explainability and structured reasoning than current generative AI models.
- Traceable Reasoning is Key: The ability to trace the inference process of an expert system is crucial for providing legally acceptable explanations, allowing users to understand why a particular answer was reached.
- Usability is a Major Hurdle: Even with a functional system, legal experts identified that inputting facts in a formal language like Datalog is a significant barrier to adoption.
- Diverse End-User Needs: The system has potential applications for traffic police, citizens, prosecutors, and judges, each with specific requirements for speed, clarity, and functionality (e.g., punishment reasoning).
- Reification for Simplicity: Treating abstract concepts as objects in the knowledge base can lead to a more concise and manageable predicate set.
Key Results
The paper reports that FÆRDXEL was used in a court case, reaching the same conclusions as the judge regarding the laws broken by the defendant. However, it was unable to provide a suitable punishment range at that time, as this aspect of the knowledge base was not yet encoded. The initial encoding resulted in 2579 clauses.
Limitations and Future Work
The authors acknowledge several limitations and outline future directions:
- Usability Challenges: Inputting facts in Datalog is difficult for legal professionals.
- Punishment Reasoning: The system currently lacks the ability to reason about punishments, a critical aspect for judges.
- Empirical Evaluation: The next step involves an empirical evaluation of FÆRDXEL's soundness by applying it to ongoing court cases and comparing its verdicts with those of judges.
- Improving Punishment Reasoning: The team is exploring Fuzzy Logic to model "Harsh" and "Soft" punishment ranges, drawing inspiration from work on Polish punishment law.
- Natural Language Processing: To enhance usability, they plan to investigate using Large Language Models (LLMs) to convert natural language case descriptions into the Datalog format required by the system.
Summary updated