ContractGuard classifies contract clauses into 12 business categories and feeds a deterministic 14-rule risk engine. The shipped classifier is a hybrid ensemble:

Component	Detail
ML model	scikit-learn TF-IDF (word 1-2 + char_wb 2-4) → Logistic Regression (C=4)
Keyword prior	Specificity-weighted bilingual legal lexicon per category
Blend	`0.65 × P(ml) + 0.35 × P(keyword)`
Artifact	`clause_classifier.joblib` (pipeline + labels bundle)
Fallback	Keyword-only heuristics when the artifact is absent

Categories: payment, delivery, warranty, confidentiality, termination, liability, penalty, intellectual_property, dispute_resolution, force_majeure, compliance, general.

Evaluation (template-disjoint — unseen phrasings)

Whole templates are held out of training, so the test set contains only phrasings the model has never seen. This avoids the template leakage that would otherwise report a misleading 100%.

Metric	Raw ML	Hybrid (shipped)
Accuracy	70.0%	92.1%
Macro F1	0.697	0.918
EN accuracy	—	95.8%
FA accuracy	—	88.3%

Per-category F1 (hybrid): payment 1.00 · liability 1.00 · penalty 1.00 · IP 1.00 · force majeure 1.00 · dispute resolution 0.98 · warranty 0.91 · delivery 0.89 · compliance 0.86 · confidentiality 0.86 · termination 0.82 · general 0.71.

Risk engine (deterministic, not learned)

Metric	Result
Recall on 140 seeded risky clauses (14 rules, EN+FA)	100%
High-severity false-positive rate (735 clean clauses)	0.68%

Full reproducible reports: eval_results.json and benchmark_report.json in the dataset repo.

Usage

python

import joblib

bundle = joblib.load("clause_classifier.joblib")
pipeline, labels = bundle["pipeline"], bundle["labels"]
proba = pipeline.predict_proba(["The Client shall pay each invoice within 30 days."])[0]
print(labels[proba.argmax()], proba.max())   # payment 0.9…

For the full hybrid path (recommended) use ClauseClassifier from the project source (see the Space files).

Intended Uses

Pre-screening commercial contracts before legal review (procurement, sales, EPC, NDAs)
Clause inventory/triage for contract migration projects
Counsel-in-the-loop review workflows (the shipped product includes an RBAC review queue)

Out-of-Scope Uses

Autonomous contract approval without qualified human review
Legal advice of any kind
Scanned/handwritten contracts without an OCR + LayoutLM upgrade (documented roadmap)

Training Data

875 synthetic bilingual clauses (contractguard-clause-samples) — template-generated, no real contracts or PII.

Limitations & Bias

Trained on synthetic commercial-contract phrasing; niche domains (insurance, construction claims) may need lexicon extension.
Persian coverage is strong on formal contract register; colloquial drafting reduces accuracy.
general is the hardest category (F1 0.71) — boilerplate overlaps with every other class.

Demo

Interactive Space: contractguard-clause-analyzer

Citation

bibtex

@misc{contractguard_clause_analyzer_2026,
  title={ContractGuard: Bilingual Contract Clause Analysis and Risk Detection},
  author={Aria AI Engineering Team},
  year={2026},
  url={https://huggingface.co/alirezaaminzadeh/contractguard-clause-analyzer}
}

License

Apache 2.0