SHIRO: Semantic Hierarchical Intent Recursive Operator

SHIRO is a lightweight, recursive NLU engine designed to split complex multi-intent sentences into atomic actions without relying on heavy Transformer-based attention. It uses a recursive MLP head and GRU-based context tracking to achieve high precision with minimal overhead.

Key Stats

• F1-Score: 0.95+
• Precision: 99.3%
• Latency: ~60ms (CPU/GPU)
• Idiom Protection: 94%+ (Native semantic bond handling)

Quick Start

1. Requirements

• Python 3.10+
• pip install -r requirements.txt

2. Interactive Testing

Run the CLI to test splits manually:

python cli.py

3. Verification

Execute the pro benchmark suite:

python bench_pro.py

Structure

• src/: Core logic (MLP Splitter, GRU History, Recursive Parser).
• data/production/: Active training corpus.
• data/benchmark/: Gold-standard test set.
• research/: Archived tuning scripts and diagnostics.
• tools/: Data generators and management utilities.

Core Logic

Unlike standard "bag-of-words" or linear classifiers, SHIRO treats sentence parsing as a binary tree construction. It evaluates the "Split Energy" at every word junction, using previous intent history as a dynamic bias. Resulting trees allow for complex execution pipelines in agentic systems.

---

MIT License