Named Entity Recognition for Legal Documents using BERT

Lawyers and compliance teams drown in documents where the key facts — who the parties are, the dates, the amounts, the cited sections, the jurisdiction — are buried in prose. Named Entity Recognition automates the first pass: finding and typing these spans. General NER trained on news (person/organization/location) misses legal specifics: "Section 11" is not a person, and "Rs. 45,00,000" in Indian digit grouping confuses amount patterns tuned on Western formats. This project fine-tunes bert-base-uncased for token classification with the BIO scheme on a five-type legal schema (PARTY, DATE, AMOUNT, SECTION, COURT), starting from CoNLL-2003 and adding ~2,000 annotated Indian legal sentences with shipped annotation guidelines. The demo highlights entities in uploaded text with per-entity confidence.

1. Legal text is tokenized with BERT's WordPiece tokenizer.
2. bert-base-uncased produces a contextual embedding for every token.
3. A dropout + linear token-classification head labels each token in the BIO scheme.
4. WordPiece sub-tokens merge to word-level tags; adjacent B-/I- spans become entities.
5. Rule-assisted post-processing normalizes Indian amount formats and citation shapes over the neural tags.
6. Fine-tuning runs 3-5 epochs (lr 3e-5, batch 16) on the annotated legal set, with validation F1 for early stopping.

Tech stack:

• Python 3.10, PyTorch
• Hugging Face transformers (bert-base-uncased)
• CoNLL-2003 (base corpus)
• Annotated Indian legal set (~2,000 sentences)
• FastAPI/Flask (inference API)
• HTML/CSS/JavaScript (demo UI)
• Git

Dataset & Model Details

• Dataset: CoNLL-2003 (Tjong Kim Sang & De Meulder; Reuters news, PER/ORG/LOC/MISC) for general entity sense + ~2,000 manually annotated Indian legal sentences (agreements, notices) in BIO format for the legal schema. Split: train / validation / held-out test.
• Model input: legal text tokens. Output: BIO labels per token → entity spans with types.
• Architecture: bert-base-uncased (110M) → dropout → linear over BIO label set; WordPiece sub-token merging; rule-assisted Indian-format post-processing.
• Metric: entity-level precision / recall / F1 (strict span+type match) on the held-out legal test set, reported per entity type after training. No F1 figure is claimed before the fine-tuning run.

• [Five legal entity types] PARTY, DATE, AMOUNT, SECTION, COURT — the schema that matters in Indian agreements and notices.
• [BIO tagging] Begin/Inside/Outside scheme handles long multi-token entities like "Arbitration and Conciliation Act, 1996".
• [Indian-format handling] Lakh/crore amount grouping and "Section N" / "Act, Year" citation shapes via rule-assisted post-processing.
• [Annotation guidelines] Two-pass labeling protocol ships with the project; inter-annotator agreement reported.
• [Entity-level evaluation] Strict precision/recall/F1 (exact span + type) on a held-out legal test set, broken down per type.
• [Highlighting demo] Upload legal text; get color-coded entities, a sortable table and per-entity confidence.
• [Inference API] REST endpoint returning entity spans for integration into review tools.

• Annotation guidelines + ~2,000-sentence labeled legal set
• Fine-tuning notebook (training, validation, per-type evaluation)
• Trained BERT weights
• Inference API (entity spans as JSON)
• Test-set evaluation: entity-level P/R/F1 per type from the actual run
• Web demo with color-coded highlighting
• Project report PDF (methodology, error analysis, measured results)
• PPT presentation for final review
• Viva Q&A preparation document

• Five entity types only; obligations, penalties and governing-law clauses are future scope.
• Long, convoluted sentences occasionally split entities — the per-type F1 table shows where.
• Trained on Indian legal English; other jurisdictions' formats need re-annotation.
• This is an extraction aid for review workflows, not legal advice.
• Fine-tuning benefits from a GPU; inference runs on CPU.