PaErHaTi-DUTiR/BiLegalNER

Named Entity Recognition in Low-Resource Legal Documents via Semantics Aware Dual Translation # BiLegalNERD A Bilingual Legal NER Dataset and Semantics-Aware Dual Translation for Low-Resource Languages This repository contains the dataset construction tools and model training code for our research on **Named Entity Recognition (NER) in low-resource legal domains**, with a focus on cross-lingual transfer from Chinese to Uyghur. --- ## 🔍 Project Overview This work addresses the challenge of legal NER in low-resource languages by introducing a bilingual Uyghur–Chinese NER dataset and a semantics-aware annotation transfer method. The proposed framework includes: - **BiLegalNERD**: A bilingual legal NER dataset constructed from Chinese legal texts and aligned Uyghur translations, annotated with ten fine-grained legal entity types. - **CUTLM**: A semantics-aware data transfer approach based on **dual translation** and **Levenshtein distance alignment**, designed to preserve entity boundaries during translation. - **BiLegalNER**: A domain-adapted NER model enhanced by **vocabulary expansion** and **bilingual fine-tuning**. --- ## 📁 Repository Structure ```bash BiLegalNERD/ ├── README.md # Project introduction and usage instructions ├── cutlm.py # Dual translation and Levenshtein-based alignment (CUTLM method) ├── trainer.py # NER model training pipeline ├── ugnerd.py # Inference and evaluation on Uyghur legal texts 🚀 Usage 1. Annotation Transfer (CUTLM) python cutlm.py --input data/chinese_annotated.json --output data/uyghur_aligned.json 2. Model Training python trainer.py --train_file data/uyghur_aligned.json --model cino --output_dir ./models/bilegalner 3. Inference and Evaluation python ugnerd.py --model_dir ./models/bilegalner --test_file data/uyghur_test_manual.json 📊 Entity Types The dataset includes 10 fine-grained legal NER tags such as: NHCS: Suspect names NHVI: Victim names NS: Locations NT: Time expressions NO: Organizations NCGV / NCSM / NASI / NCSP / NATS: Various legal-specific case elements 🔗 Notes All da …