CHisIEC: An Information Extraction Corpus for Ancient Chinese History

This repository provides the datasets and code used in the paper “CHisIEC: An Information Extraction Corpus for Ancient Chinese History”. It includes Named Entity Recognition (NER) and Relation Extraction (RE) resources for pre-modern Chinese historical texts.

---

📦 Data

🔤 NER

Path: ./data/ner/

• Data is processed in CoNLL format.

🔗 RE

Path: ./data/re/

• Data is stored in JSON format.

---

🧪 Code

1. 🏷️ NER Code

You can refer to our separate repository: 👉 https://github.com/tangxuemei1995/AnChineseNERE

---

2. 🔍 Relation Extraction Code

2.1 📘 BERT / RoBERTa Models (SikuBERT & SikuRoBERTa)

Directory: code/bert_roberta_re/

• Corresponds to Table 5 models in the paper
• Based on the method from “Enriching Pre-trained Language Model with Entity Information for Relation Classification” https://arxiv.org/abs/1905.08284
• Training settings located in: code/bert_roberta_re/config.ini

You may obtain SikuBERT and SikuRoBERTa from HuggingFace:

• Either download to a local directory
• Or load directly during training (if your environment allows Internet access)

---

2.2 🤖 ChatGLM2 (6B, P-tuning)

Steps:

1. Download the alpaca2 instruction-tuned model into the glm2/ directory
2. Ensure all data is placed in data/
3. Modify training settings inside train_coling.sh

Run:

bash train_coling.sh          # fine-tune the model
bash evaluate_coling.sh       # generate generated_predictions.txt
python evaluate_re_coling.py  # evaluate

---

2.3 🐪 Alpaca2 (7B, LoRA)

Run:

bash run_sft_chapter.sh    # train model
bash merge_new_model.sh    # merge LoRA
bash run_test.sh           # inference
python evaluate.py         # evaluation

---

⚠️ Important

Before running any models, please ensure:

• Model paths are correct
• Data paths are set correctly
• Necessary dependencies and GPU environments are available

---

🖥️ Annotation Platform

The two datasets were annotated using: 👉 https://wyd.pkudh.net/ This platform was developed by the Digital Humanities Research Centre, Peking University. It supports deep-learning–based annotation for arbitrary corpora.

---

📚 Citation

If you use CHisIEC, please cite:

@article{Tang_Deng_Su_Yang_Wang_2024,
  title={CHisIEC: An Information Extraction Corpus for Ancient Chinese History},
  url={http://arxiv.org/abs/2403.15088},
  note={arXiv:2403.15088 [cs]},
  number={arXiv:2403.15088},
  publisher={arXiv},
  author={Tang, Xuemei and Deng, Zekun and Su, Qi and Yang, Hao and Wang, Jun},
  year={2024},
  month=mar
}