💻 computer science

Neuro-Symbolic AI for Korean Criminal Law: Sentencing Prediction and Document Drafting

This paper proposes a Neuro-Symbolic AI framework that combines Large Language Models for semantic extraction with Satisfiability Modulo Theories solvers for deterministic statutory calculations to enhance the accuracy and reliability of sentencing predictions and document drafting for Korean traffic offenses, thereby mitigating the hallucination risks inherent in purely neural legal systems.

Original authors: Yeonseok Lee

Published 2026-07-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yeonseok Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where computers are like incredibly talented but slightly scatterbrained chefs. They can read a thousand recipes in a second, understand the flavor of a dish, and even write a beautiful poem about a tomato. This is the realm of Artificial Intelligence, specifically a type called Large Language Models (LLMs). These digital chefs are amazing at understanding language, but they have a tricky habit: sometimes, when asked to do precise math or follow strict rules, they "hallucinate." They might confidently tell you that a tomato costs $500 because it sounds like a good number, even though the recipe says $5. In the real world, this is fine for writing stories, but in a courtroom, where a mistake could mean someone loses their freedom or pays the wrong fine, being "almost right" isn't good enough.

This is where Neuro-Symbolic AI comes in. Think of it as a team-up between that creative, scatterbrained chef and a super-strict, rule-following accountant. The chef reads the messy notes and understands the story, while the accountant does the math and checks the rules to make sure everything adds up perfectly. This paper explores how to build this dream team to help prosecutors in South Korea handle thousands of minor traffic cases, like drunk driving, without getting overwhelmed or making dangerous mistakes.

The Problem: Too Many Tickets, Not Enough Humans

In South Korea, the legal system has a special fast lane for minor crimes called summary proceedings (or guyaksik). Imagine a highway where the police catch hundreds of drivers for small offenses like driving without a license, causing a tiny fender-bender, or having a little too much to drink. Instead of holding a full-blown trial for every single one, prosecutors just review the paperwork and issue a fine. It's efficient, but it creates a mountain of paperwork. Prosecutors are drowning in stacks of police reports, breathalyzer logs, and witness statements. They need a way to turn these messy, handwritten, or scanned documents into a clear, mathematically perfect legal decision.

The Trap: Why Smart Computers Get Stuck

Recently, people tried using those "scatterbrained chef" computers (LLMs) to solve this. The computers are great at reading a police report and saying, "Ah, this guy had 0.12% alcohol in his blood and crashed into a barrier." But when asked to calculate the exact fine based on complex laws, the computer starts to guess. It might add up the numbers wrong or invent a rule that doesn't exist. In criminal law, you can't have a computer guessing. If the math is wrong, the justice system breaks. The paper argues that we cannot rely on these computers to do the math themselves; they are too prone to making up facts.

The Solution: A Team of Chef and Accountant

The authors propose a Neuro-Symbolic framework, which is a fancy way of saying they split the job into two distinct parts to get the best of both worlds.

Step 1: The Chef (The LLM)
First, the system takes the messy, unstructured documents—like a breathalyzer result or a police officer's notes—and feeds them to the LLM. The LLM's only job is to be a "semantic extractor." It reads the story and pulls out the important facts: "The driver's blood alcohol was 0.12," "They crashed into a median barrier," and "They have no prior record." It puts these facts into a neat, structured list (like a JSON file). Crucially, the LLM is not allowed to calculate the fine yet. It just gathers the ingredients.

Step 2: The Human Inspector (The Safety Check)
Before the math happens, a real human prosecutor looks at the list. The system highlights exactly where in the original document the computer found each fact. The human clicks a button to say, "Yes, that is correct," or "No, that's wrong." This is called Human-in-the-Loop (HITL). It ensures that the computer doesn't make up a fact that the human didn't verify. Only the facts the human approves get to move to the next step.

Step 3: The Accountant (The Z3 Solver)
Now, the verified facts go to the "Accountant," which is a piece of software called a Z3 Satisfiability Modulo Theories (SMT) solver. This isn't a guessing machine; it's a mathematical proof engine. It takes the human-approved facts and runs them through the strict rules of the 2026 Sentencing Guidelines for Traffic Offenses.

If the blood alcohol is between 0.08 and 0.20, the computer must categorize it as "Type 3."
If the driver crashed into a barrier, the computer must add an "aggravating factor."
It then does the actual math to figure out the fine range.

Because this part is pure logic and math, it cannot hallucinate. If the rules say the fine is between 5,000,000 and 8,000,000 Korean Won, that is exactly what it outputs. If the math doesn't add up (for example, if the human accidentally approved two facts that contradict each other), the Accountant stops and says, "Error! Something is wrong," and sends the case back to the human to fix it.

The Result: A Perfectly Drafted Document

Once the Accountant has the final number, the system goes back to the "Chef" (the LLM) one last time. This time, the LLM is given the exact math result and asked to write the final legal document, like a Summary Indictment Request. The LLM writes the document in the proper legal format, but it copies the fine amount directly from the Accountant's result. It doesn't calculate anything; it just writes the story.

What This Means for the Future

The paper shows that this method works in theory and creates a pipeline that is both fast and safe. It suggests that by keeping the "creative" AI away from the "math" and letting a human check the work, we can automate the boring parts of the legal system without risking errors.

The authors are careful to note that this is currently a framework they have designed and tested logically. They haven't rolled it out to every police station in Korea yet. However, they plan to build the full software and test it with real data. If successful, this system could eventually handle not just drunk driving, but other minor crimes like simple assault, minor theft, or online insults, helping prosecutors clear their desks while ensuring that every fine is calculated with mathematical precision. It's a way to let computers do the heavy lifting, while keeping the human in the driver's seat.

Technical Summary: Neuro-Symbolic AI for Korean Criminal Law

1. Problem Statement

The Korean criminal justice system relies heavily on summary proceedings (guyaksik) to expedite high-volume minor infractions, particularly traffic offenses such as simple Driving Under the Influence (DUI), unlicensed driving, and minor traffic casualties. While this mechanism improves judicial throughput, it imposes a significant administrative burden on prosecutors who must manually process unstructured legal documents (e.g., breathalyzer logs, police reports, interrogation records) into deterministic statutory outcomes.

Recent advances in Large Language Models (LLMs) have demonstrated strong capabilities in semantic extraction from dense legal texts. However, deploying pure LLMs for Legal Judgment Prediction (LJP) presents fundamental limitations:

Probabilistic Nature: LLMs generate text based on probability, which is ill-suited for the strict arithmetic and rule-based reasoning required for statutory fine calculations.
Hallucination Risk: In legal contexts, where accountability demands complete determinism, LLMs frequently produce "hallucinations" when confronted with complex mathematical constraints or specific statutory thresholds.
Lack of Formal Verification: Pure neural architectures cannot mathematically prove the validity of a sentencing outcome, rendering them insufficient for direct judicial application where stochastic errors are intolerable.

2. Methodology

The authors propose a Neuro-Symbolic framework that bridges unstructured legal facts with formal verification. The architecture strictly partitions tasks between neural semantic understanding and symbolic logical reasoning, organized into a five-module pipeline:

Module 1: Raw Data Ingestion

The system ingests heterogeneous, layout-heavy evidentiary documents (DUI circumstantial statements, breathalyzer results, interrogation records, and accident reports). It employs a hybrid layout-aware ingestion protocol using custom rule-based Optical Character Recognition (OCR) and bidirectional encoder text segmentation to isolate narrative sections from tabular indices and metadata, converting them into clean text streams.

Module 2: LLM Extraction

An LLM acts as a semantic parser to extract core case elements (crime types, continuous variables like Blood Alcohol Concentration (BAC), and discrete categorical variables).

Constraint: The LLM is structurally constrained to map every extracted factor directly to its exact textual source ("Reference Text") to enforce legal syllogism and grounding.
Output: Extracted entities are populated into an intermediate JSON schema, serving as the bridge between probabilistic extraction and deterministic logic.

Module 3: Human-in-the-Loop (HITL) Verification

To ensure legal compliance and due process, a mandatory verification phase is introduced before any calculation occurs.

Interface: A split-screen UI displays the structured JSON output alongside the original document with highlighted reference text.
Mechanism: All extracted variables default to is_valid_by_human = false. A legal practitioner must explicitly verify each factor to toggle it to true.
Data Provenance: Only parameters explicitly validated by a human are forwarded to the solver, ensuring the system acts as a decision-support tool rather than an autonomous decision-maker.

Module 4: Logic Translation and Z3 SMT Solver

This is the core neuro-symbolic bridge. Validated JSON fields are mapped to formal state variables (Boolean flags and Integer constraints) within the Z3 Satisfiability Modulo Theories (SMT) solver.

Axiom Modeling: The 2026 Sentencing Guidelines for Traffic Offenses are formalized as mathematical axioms. For example, BAC thresholds are modeled as conditional constraints (e.g., $0.08 \le BAC < 0.20 \implies Crime\_Category = 3$ ).
Zone Derivation: The solver evaluates Boolean arrays of mitigating and aggravating factors, summing them to determine the sentencing zone (Basic, Mitigated, or Aggravated).
Mandatory Imprisonment Logic: The solver checks composite logical constraints (e.g., if prior convictions $\ge 3$ or if aggravating factors are dominant) to override fine calculations with mandatory imprisonment recommendations.
Satisfiability Check: Before outputting a result, Z3 performs a SAT check to ensure no logical contradictions exist (e.g., a case cannot be both "Mitigated" and "Aggravated"). If an UNSAT state is detected, the pipeline loops back to the HITL interface for conflict resolution.

Module 5: Visualization and Document Generation

The deterministic outputs from the Z3 solver are translated into actionable artifacts:

Sentencing Tree Visualization: The system renders the logical decision path (from BAC categorization to final fine) into a visual tree, providing an interpretable audit trail.
Document Drafting: Using rigidly structured prompts, an LLM generates legally compliant documents (Police Investigation Reports, Forwarding Opinions, and Summary Indictment Requests). Crucially, the fine amounts are injected verbatim from the Z3 solver output, preventing the LLM from recalculating or hallucinating numbers.

3. Key Contributions

The paper outlines three primary contributions:

Neuro-Symbolic Pipeline Design: A framework that transforms unstructured police and prosecutorial records into deterministic legal outcomes, specifically isolating the LLM to semantic extraction to mitigate hallucination risks during fine computation.
Human-in-the-Loop (HITL) Verification: A transparent scheme that anchors extracted legal variables to source reference texts, ensuring human oversight and interpretability while maintaining the final legal authority with the prosecutor.
Formal Legal Logic Framework: The construction of a logic framework using the Z3 SMT solver to compute precise sentencing ranges aligned with the 2026 Sentencing Guidelines, replacing probabilistic generation with mathematical proof.

4. Results and Claims

The paper does not present empirical performance metrics (e.g., accuracy percentages on a test set) as the study focuses on the theoretical formulation and logical mapping of the architecture. Instead, the authors claim the following based on the design:

Determinism: By offloading statutory calculations to a theorem prover, the system guarantees that sentencing outcomes are free from the stochastic errors inherent in LLMs.
Robustness: The SAT check mechanism ensures that logical contradictions are detected and resolved before a verdict is generated, preventing invalid legal outcomes.
Accountability: The HITL integration ensures that the system remains a verifiable assistive tool, addressing transparency concerns inherent in "black box" AI models.

5. Significance

The authors position this work as a necessary evolution in Legal Judgment Prediction for high-stakes environments. By acknowledging that LLMs are insufficient for arithmetic and strict rule adherence, the paper proposes a hybrid approach that leverages the semantic strengths of neural networks while relying on symbolic logic for the "hard" constraints of law.

The significance lies in its potential to alleviate the administrative bottleneck of minor prosecutorial caseloads in Korea without compromising judicial integrity. The framework offers a pathway to automate the drafting of summary indictments (yaksik-gongsojang) while ensuring that every mathematical outcome is grounded in formally verified, human-validated facts. The authors note that while the current focus is on traffic offenses, the architecture is designed to be scalable to other minor crimes (e.g., simple assault, defamation, minor theft) that conclude with summary proceedings.