SemTree: Judge-Aware Semantic Decision Trees for Inspectable and Repairable Rule Decisions
Rule decisions such as content moderation, advertising compliance, and eligibility checks combine several conditions, exemptions, and missing-information cases, and an error can be fixed only once the failing condition is known.
This paper introduces SemTree, a framework for creating interpretable and repairable rule decision systems using semantic decision trees. It addresses challenges in error correction for complex rule-based systems by enabling transparent reasoning and iterative improvements.
What happened
The article presents SemTree, a framework that structures rule decisions as semantic decision trees. Each tree node contains natural language questions with explicit criteria, answered by a pluggable semantic judge. The system enables error repair by correcting specific nodes, with results validated through experiments on 8,276 test cases and real-world applications.
Why it matters
The framework addresses critical challenges in AI transparency and maintainability for rule-based systems. Its ability to isolate and repair errors directly improves reliability in applications like content moderation and compliance checks, which require explainable decision-making processes.
- Semantic decision trees enable transparent rule-based decision systems
- Node-level repairs allow targeted error correction without retraining entire models
- LLM editors can automate subtree refinement through execution feedback
- The framework balances interpretability with performance in real-world applications
- Error localization improves maintainability of complex rule systems
The bigger picture
The approach provides a practical method for creating audit-able rule systems, reducing error correction costs and improving trust in automated decision-making. The demonstrated accuracy improvements suggest potential for adoption in regulated industries requiring explainable AI.
SemTree offers a novel approach to building interpretable rule decision systems through semantic decision trees. By combining natural language reasoning with iterative repair mechanisms, it addresses key challenges in AI transparency and maintainability while achieving competitive performance metrics.
Technical details
- Semantic decision trees with natural language questions and Yes/No/Unknown criteria
- Pluggable semantic judges for node evaluation
- Iterative node repair with training label generation
- LLM-based editor for subtree revisions
- Execution path tracking for decision transparency
FAQ
What is SemTree's main innovation?
SemTree introduces semantic decision trees that combine natural language questions with pluggable judges, enabling transparent reasoning and targeted error repair in rule-based systems.
How does the framework improve error correction?
By isolating failing conditions through tree node analysis, reviewers can correct specific nodes. These corrections become training data for judges, and an LLM editor refines tree structures based on execution results.
What performance metrics were achieved?
The system reached 67.70% accuracy on 8,276 test cases with 2.99 questions per case, outperforming alternative approaches and demonstrating practical effectiveness.
What real-world applications were tested?
The framework was validated on banking data, Chinese e-commerce moderation, and general rule decision tasks, showing improvements in accuracy and false positive reduction.
How does this differ from traditional rule systems?
Unlike opaque rule systems, SemTree provides explicit reasoning paths, allows iterative repairs, and integrates machine learning for continuous improvement through node-level feedback.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv cs.LG on Oct 14, 2026.
Comments
Sign in to join the discussion.