Legal Clause Risk Highlighter for Rental Agreements

Rental agreements are long, dense, and almost never read fully — yet they contain clauses that can cost a tenant months of rent in penalties or lock them into one-sided terms. Professional contract review is expensive, and generic chatbots give unreliable legal-sounding answers with no grounding. This project takes the grounded route: a Legal-BERT classifier fine-tuned on CUAD (Contract Understanding Atticus Dataset, 510 real contracts with 41 labeled clause categories) identifies clause types in a rental agreement, and a risk-mapping layer scores each category as high, medium or low risk for the signing party. Every highlight is traceable — the demo shows the clause text, its CUAD category, the risk tier and a plain-language explanation of why it matters. The report is explicit that this is an educational aid, not legal advice.

1. The rental agreement text is segmented into clauses using heading and numbering structure, then tokenized for the model.
2. Legal-BERT, fine-tuned on CUAD, assigns each clause its most likely clause categories with confidence scores (multi-label classification).
3. A risk-mapping table converts categories to risk tiers for the tenant — e.g. unilateral termination rights map to high risk, governing-law clauses to low.
4. The demo highlights each clause in the document with its risk color; hovering or clicking opens the detail panel with category, confidence and explanation.
5. A summary generator aggregates the findings into a risk report: high-risk clause count, the most severe items first, and suggested negotiation points.
6. Evaluation runs on CUAD's held-out contracts with per-category precision/recall and AUPR, reported honestly in the report.

Tech stack:

• Python 3.11, Hugging Face Transformers
• Legal-BERT (nlpaueb/legal-bert-base-uncased)
• CUAD dataset (The Atticus Project)
• PyTorch (fine-tuning)
• scikit-learn (metrics: precision/recall/AUPR)
• Flask + HTML/CSS/JS (demo)
• pdfminer (agreement text extraction)

Dataset & model details

• Dataset: CUAD — Contract Understanding Atticus Dataset (Hendrycks et al., 2021): 510 real contracts, 13,000+ expert-labeled clause spans across 41 categories. Source: Hugging Face datasets (cuad). The demo's sample rental agreement is written for the project and labeled in CUAD style.
• Task: Multi-label clause classification — input = contract clause text, output = CUAD clause categories with confidence, mapped to tenant risk tiers.
• Model: Legal-BERT base (110M parameters) fine-tuned on CUAD clause spans; risk tiers from a documented category→risk mapping table.
• Metrics: Per-category precision/recall and area under the precision-recall curve (AUPR) on CUAD's test split — the published CUAD baseline reaches ~44% AUPR, which is cited as the literature reference; this project's achieved figures are design targets reported after training.

• [CUAD-trained clause classifier] Legal-BERT fine-tuned on the CUAD dataset recognizes 41 contract clause categories (termination, indemnity, governing law, renewal terms and more).
• [Risk-tier mapping] Each clause category is mapped to high/medium/low risk for the tenant, with a documented rationale for every mapping.
• [In-document highlighting] The demo renders a realistic rental agreement with risky clauses highlighted in place and the exact risky phrasing underlined.
• [Plain-language explanations] Every flagged clause carries a non-technical explanation of what it means and what to negotiate — written for students, not lawyers.
• [Clause detail panel] Click any highlight to see its CUAD category, model confidence, risk tier and the reference definition from the dataset's annotation guide.
• [Risk summary report] A one-page summary lists all high-risk clauses with counts, so the whole agreement's risk profile is visible at a glance.
• [Educational-use framing] The interface and report state clearly that this is a study aid, not legal advice — an honest scope boundary.

• Fine-tuned Legal-BERT clause classifier
• Clause segmentation + risk-mapping pipeline source
• Web demo with a realistic preloaded rental agreement
• CUAD evaluation notebook (per-category precision/recall/AUPR)
• Plain-language explanation bank for all risk-flagged categories
• Project report PDF (background, legal NLP, CUAD, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (Legal-BERT, CUAD, multi-label classification, AUPR)