README.md

metadata

title: Structured Intelligence & Grounded Meaning Analyzer
emoji: 👀
colorFrom: yellow
colorTo: gray
sdk: gradio
app_file: app.py
pinned: false

Structured Intelligence & Grounded Meaning Analyzer

Try the interactive SIGMA reasoning tool via Hugging Face Space Link | UI snapshot

This project was developed using prompt engineering techniques with ChatGPT GPT-5.5.

Manual code review - notes:

roberta-large-mnli uses 0 for contradiction and 2 for entailment, while 
GLUE MNLI uses 0 for entailment and 2 for contradiction.
Generated code (for evaluate_mnli.py) applied 0 for contradiction and 2 for entailment. 

>>> Upon intervention, AI regenerated the code to apply label re-mapping to handle the differences.

---

Model Stack

Fact Extraction

• spaCy en_core_web_sm
• Explicit fact extraction, extracts structured factual triples: Subject → Action → Object

Inference Validation

• roberta-large-mnli (Multi-Genre Natural Language Inference)

• Transformer model to evaluate candidate hypotheses against the original passage, classifying each hypothesis as:

  • ENTAILMENT
  • NEUTRAL
  • CONTRADICTION

• Confidence scores are derived from the model's softmax probability distribution over the three inference classes.

Summarization

• facebook/bart-large-cnn
• Generates concise analytic summaries

Technical Notes

• Uses conservative syntactic extraction to avoid speculative relation generation
• Separates confirmed facts from analytical assessments
• Built entirely with pretrained transformer models
• Optimized for CPU deployment on Hugging Face Spaces

---

Architecture

Input Text
   ↓
[NER + Relation Extraction]
   ↓
Explicit Fact List
   ↓
[NLI Model - Entailment Testing]
   ↓
Validated Implicit Inferences
   ↓
[Summarization Model]
   ↓
Intelligence Brief Output

Architecture of Inference Layer

Structured Facts
       ↓
Embedded Clause Extraction
       ↓
Standalone Hypothesis Generation
       ↓
MNLI Entailment Test
       ↓
Filter (confidence threshold)
       ↓
Implicit Inference Output

---

Validation

Validated pretrained reasoning model against standardized benchmark with tracked metrics and reproducible experiment logging.

Evaluated roberta-large-mnli on a GLUE MNLI validation split and track results with MLflow.

Label re-mapping is required to handle:

• Model roberta-large-mnli: 0 = contradiction | 1 = neutral | 2 = entailment
• Dataset GLUE MNLI: 0 = entailment | 1 = neutral | 2 = contradiction

Results

---

Project Structure

/
├── src/                     # Core NLP logic
│   ├── __init__.py
│   ├── fact_extractor.py
│   ├── inference_engine.py
│   ├── summarizer.py
│   ├── pipeline.py
│
├── app.py                   # UI / deployment entrypoint
├── requirements.txt
└── README.md