Guidelines for the Annotation and Visualization of Legal Argumentation Structures in Chinese Judicial Decisions

This paper proposes a systematic framework for annotating and visualizing legal argumentation structures in Chinese judicial decisions by defining four proposition types and five relational types to enable consistent, reproducible computational analysis of judicial reasoning.

The Gist

Imagine you are trying to understand a complex recipe for a cake. You could just read the list of ingredients and the final result, but to truly understand how the baker made it, you need to see the steps: "Mix flour and eggs," "Add sugar," "Bake at 350 degrees."

This paper is essentially a instruction manual for turning a judge's written decision into a clear, step-by-step recipe.

In the legal world, judges write long documents explaining why they decided a case the way they did. This is called "legal reasoning." Currently, computers and even humans often struggle to see the logical structure hidden inside these long paragraphs. This paper proposes a new system to "annotate" (label) and "visualize" (draw) these arguments so they are easy to read, teach, and analyze by computers.

Here is a breakdown of the paper using simple analogies:

1. The Problem: The "Wall of Text"

Judges write decisions in natural language. It's like a long, winding story.

• The Issue: If you ask a computer, "What facts did the judge rely on to convict this person?" the computer might get lost in the story. It sees words, but not the logic.
• The Goal: The authors want to turn that "wall of text" into a flowchart or a blueprint. This makes it possible to teach law students, check if a judge's logic is sound, and build AI that can actually explain why it made a decision.

2. The Building Blocks: "Propositions" (The Bricks)

Before you can build a house, you need bricks. In this system, the "bricks" are called Propositions.
Instead of looking at whole sentences, the annotators break the text down into tiny, logical statements that can be proven true or false. They sort these bricks into four boxes:

• General Facts (GF): The "Rules of the World." (e.g., "Water freezes at 0°C.")
• General Rules (GM): The "Laws." (e.g., "The Civil Code says you must pay your debts.")
• Specific Facts (SF): The "Story of this Case." (e.g., "John borrowed $100 from Mary.")
• Specific Rules (SM): The "Verdict." (e.g., "Therefore, John must pay Mary $100.")

Analogy: Think of a courtroom like a courtroom drama.

• GF is the physics of the room (gravity exists).
• GM is the rulebook of the game (FIFA rules).
• SF is what actually happened in the match (Player A kicked the ball).
• SM is the referee's final call (Player A is offside).

3. The Glue: "Relations" (How the Bricks Connect)

Just having bricks isn't enough; you need mortar to hold them together. The paper defines five types of "glue" (relationships) that connect the bricks:

1. Support (The "Yes" Glue): One brick holds up another. (Fact A + Law B = Conclusion C).
2. Attack (The "No" Glue): One brick knocks another down. (Evidence X proves that Fact A is a lie).
3. Joint (The "Team" Glue): Two or more bricks must work together to hold up the roof. If you remove one, the whole thing collapses. (You need both the contract and the signature to prove a deal).
4. Match (The "Puzzle Piece" Glue): This is special. It connects a General Rule to a Specific Fact. It's like saying, "This specific puzzle piece (Fact) fits perfectly into this specific slot in the rulebook (Law)."
5. Identity (The "Twin" Glue): Two bricks say the exact same thing in different words. (We treat them as one).

4. The Blueprint: Visualizing the Argument

Once the text is broken into bricks and glued together, the paper says: "Draw it!"

They created a visual language so you can see the logic at a glance:

• Rectangles = The Bricks (Facts and Rules).
• Circles = The Glue (The relationships).
  • Solid Circle: Support (Good!).
  • Hollow Circle: Attack (Bad!).
  • Circle with a "+": Teamwork (Joint or Match).
  • Slash (/): Twins (Identity).

Why do this? Imagine looking at a tangled ball of yarn. It's impossible to see the pattern. Now, imagine that yarn has been untangled and laid out on a table with arrows showing exactly how one thread leads to the next. That is what this visualization does for a judge's decision.

5. The Recipe for Success: How to Do It

The paper doesn't just give the theory; it gives a strict recipe for humans to follow so they don't mess it up:

1. Training: Teach the annotators the rules.
2. Practice: Have them try on a few cases to find confusing spots.
3. Double-Check: Have two people label the same text independently. If they disagree, a "referee" (expert) decides who is right.
4. Consistency: Make sure everyone is using the same dictionary and rules.

Why Does This Matter? (The "So What?")

• For Law Students: Instead of memorizing a 50-page case, they can look at a diagram and instantly see the logic. It's like having a cheat sheet for how judges think.
• For Judges: It helps them check their own work. "Did I actually connect the facts to the law, or did I just skip a step?"
• For AI (Robots): This is the big one. Current AI can guess the outcome of a case, but it can't explain why. By feeding this structured "blueprint" data to AI, we can build Explainable AI. Imagine a robot lawyer that doesn't just say "You lose," but says, "You lose because Fact A didn't match Law B, and here is the diagram proving it."

Summary

This paper is a user manual for decoding the brain of the law. It turns messy, complex legal text into a clean, structured, and visual map. It bridges the gap between human logic and computer code, ensuring that when we use AI in the courtroom, we can actually understand and trust the reasoning behind the decisions.

Technical Summary

Here is a detailed technical summary of the paper "Guidelines for the Annotation and Visualization of Legal Argumentation Structures in Chinese Judicial Decisions (Working Draft v1.0)".

1. Problem Statement

Despite the rapid advancement of Judicial Artificial Intelligence (JAI), current research predominantly focuses on outcome prediction or document generation. These approaches typically treat legal reasoning as surface-level textual features or statistical patterns, failing to capture the internal logical structure, inferential relations, and normative constraints of judicial decisions.

• The Gap: Existing models lack fine-grained representation of argumentative structures, making it difficult to achieve explainability, auditability, and normative compliance in AI systems.
• The Challenge: Natural language in judicial decisions is implicit and unstructured. There is a lack of a standardized, reproducible framework to transform these texts into structured, computable argumentation objects that bridge legal theory and computational analysis.

2. Methodology

The paper proposes a systematic, three-stage annotation framework designed to extract and visualize legal argumentation structures from Chinese judicial decisions (specifically the "reasoning" or "court holds" sections).

A. Annotation Granularity: The Proposition

Instead of using sentences as the basic unit, the framework uses the "Proposition" as the minimal semantic unit. A proposition is defined as the smallest semantic unit capable of being evaluated as true or false within the reasoning section.

B. Proposition Type System (4 Types + 6 Sub-types)

The framework classifies propositions into four basic types based on two dimensions: Particular vs. General and Fact vs. Norm.

1. SF (Particular Factual Judgment): Facts specific to the case (e.g., "Defendant owes 11,000 RMB").
2. GF (General Factual Judgment): Background knowledge, common sense, or empirical rules.
3. SM (Particular Normative Judgment): The court's specific normative evaluation or conclusion (e.g., "Defendant shall pay").
4. GM (General Normative Judgment): The normative foundation (laws, contracts, customs).
   • Sub-types for GM: GM-L (Statutory), GM-I (Legal Interpretation), GM-C (Contracts), GM-U (Customs), GM-M (Morality/Values), GM-O (Other).
   • Principles: Priority matching, non-redundancy, and reliance on explicit textual expression.

C. Inter-Propositional Relation Types (5 Types)

The framework defines five logical relations to map the argument flow:

1. Support (S): One proposition (or group) provides reasons for another.
2. Attack (A): One proposition provides reasons against another or negates a support link.
3. Joint (J): Multiple propositions are jointly necessary (conjunctive) to support a conclusion.
4. Match (M): A specific relation between General Norms (GM) and Particular Facts (SF) to show how a rule applies to a case.
5. Identity (I): Semantic equivalence between different text segments.

D. Formal Representation & Nesting

The paper establishes rigorous rules for nested structures (e.g., a Joint relation embedded within a Match relation, which is then embedded in a Support relation).

• Formal Notation: Uses functional notation (e.g., $S(M(J(p1, p2), p3), p4)$) to represent complex hierarchies.
• Constraint: Hierarchical parentheses must strictly correspond to logical structure; cross-level merging is prohibited.

E. Visualization Standards

A unified diagrammatic language is introduced:

• Nodes: Rectangles for propositions; Circles for relations.
• Styles:
  • Solid Circle: Support.
  • Hollow Circle: Attack.
  • Circle with "+": Joint or Match.
  • Slash "/" in a rectangle: Identity.
• Direction: Arrows indicate the flow of support or attack. Nested structures are drawn layer-by-layer from the innermost relation outward.

F. Workflow & Consistency Control

A four-stage workflow is prescribed to ensure data quality:

1. Training: Annotator education on definitions and examples.
2. Pilot Annotation: Testing guidelines to identify ambiguities and calibrate rules.
3. Formal Annotation: Dual independent annotation (two annotators per text) to ensure objectivity.
4. Conflict Resolution: Automated comparison followed by expert adjudication for core conceptual disagreements.

3. Key Contributions

• Standardized Framework: The first comprehensive guideline specifically for Chinese judicial decisions, addressing the unique structure of "Court Holds" (Ben Yuan Ren Wei) sections.
• Novel Proposition Taxonomy: A distinct 2x2 matrix (Fact/Norm × Particular/General) that moves beyond simple "premise/conclusion" labeling to capture the nature of legal reasoning.
• The "Match" Relation: Explicitly defining the Match (M) relation to model the critical legal operation of subsuming facts under norms, distinguishing it from generic support.
• Handling Complexity: Detailed rules for nested argumentation structures, allowing the representation of complex legal reasoning where rules, facts, and conclusions are interwoven.
• Visual & Formal Duality: Providing both a formal logical notation and a standardized visual diagramming method, facilitating both human understanding and machine parsing.
• Quality Assurance Protocol: A robust, multi-stage annotation workflow with built-in consistency metrics and expert review mechanisms.

4. Results (Demonstrated via Case Studies)

The paper validates the framework through a full annotation workflow on a real-world labor contract dispute (Document I).

• Segmentation: Successfully broke down a paragraph of reasoning into 11 distinct propositions ($p1$ to $p11$).
• Classification: Correctly identified statutory laws (GM-L), factual findings (SF), and normative conclusions (SM).
• Structure Reconstruction: Mapped the complex logic:
  • Facts ($p4, p5$) + Norms ($p2$) $\rightarrow$ Match $\rightarrow$ Conclusion ($p6$).
  • Facts ($p6, p7$) + Norms ($p3$) $\rightarrow$ Match $\rightarrow$ Conclusion ($p8$).
  • Evidence insufficiency ($p10$) $\rightarrow$ Attack/Support $\rightarrow$ Rejection of claim ($p11$).
• Visualization: Generated clear argument diagrams that visually separate the successful claim (11,000 RMB) from the rejected claim (600 RMB), demonstrating the framework's ability to handle mixed outcomes in a single judgment.

5. Significance

• For Legal AI: Provides the essential ground truth data required to train Explainable AI (XAI) models. It shifts JAI from "black box" prediction to "white box" reasoning, enabling systems to justify decisions legally.
• For Legal Education: Offers a pedagogical tool to visualize the hidden logic of judgments, helping students understand how facts, laws, and conclusions interconnect.
• For Judicial Practice: Enhances the transparency and auditability of judicial reasoning, allowing for better quality control and supervision of court decisions.
• For Research: Enables large-scale empirical studies on judicial reasoning patterns, allowing researchers to compare argumentation structures across judges, courts, and case types, moving legal scholarship from qualitative case studies to quantitative structural analysis.