TL;DR
This paper introduces GRoWE, a novel approach for Named Entity Recognition (NER) in ancient Chinese texts. By combining a specialized pre-trained language model (GujiRoBERTa) with a word-word relation classification module and an ensemble strategy, GRoWE significantly improves NER performance on challenging historical texts.
What Problem Does This Paper Address?
Ancient Chinese texts are a treasure trove of historical and cultural information, but extracting this information is difficult due to the nature of the language. Classical Chinese is concise, often omits subjects or objects, and words can have different meanings depending on context. These ambiguities make traditional Named Entity Recognition (NER) tasks, which aim to identify entities like names, titles, and places, particularly challenging. The EvaHan 2025 initiative aims to standardize and advance NER for these texts, providing a unified dataset and pre-trained models to spur progress.
Methods
GRoWE builds upon the GujiRoBERTa-jian-fan model, a pre-trained language model specifically designed for ancient Chinese texts. The core innovation lies in its architecture:
- GujiRoBERTa Encoder: This serves as the foundation, capturing the semantic nuances of ancient Chinese.
- Word-Word Relation Classification (W2NER): Inspired by the W2NER framework, this module explicitly models relationships between word pairs. This is crucial for handling complex entity structures and overlapping entities, which are common in classical texts.
- Ensemble Strategy: To further boost performance and reduce prediction bias, GRoWE employs a multi-model ensemble. Five models are trained using five-fold cross-validation, and their predictions (logits) are summed to produce a final, more robust result.
Datasets
The research utilizes datasets from the EvaHan 2025 initiative:
- Shiji (dataset A): Historical records.
- Twenty-Four Histories (dataset B): A collection of historical texts.
- Traditional Chinese Medicine Classics (dataset C): Historical medical texts.
The paper also describes a data enhancement process involving sentence segmentation, de-duplication, and a sliding window technique to expand the training data.
Key Insights and Ideas
- Specialized Pre-training is Key: Using a language model pre-trained on ancient Chinese texts (GujiRoBERTa) provides a significant advantage over general-purpose models.
- Word-Word Relations Matter: Explicitly modeling relationships between word pairs (via W2NER) is highly effective for capturing complex and overlapping entities in classical Chinese, overcoming limitations of simpler sequential models.
- Ensembling for Robustness: Combining multiple models through ensembling not only improves accuracy but also makes the predictions more stable and reliable, mitigating the risk of individual model errors.
Key Results
GRoWE achieved a remarkable F1-score of 86.79% on the EvaHan 2025 public test set. This represents a significant improvement of 6.18% over the mainstream BERT_LSTM_CRF baseline model. The paper also shows that the W2NER architecture, when used with the GujiRoBERTa encoder, outperforms other baseline architectures like GlobalPointer and MRC.
Limitations and Future Work
- Data Breach Risk in Stratified Sampling: The authors acknowledge that their initial stratified sampling for cross-validation might have introduced a potential risk of data leakage, although they argue it was acceptable for model selection. They later re-divided the data using sequential cutting for more reliable final results.
- Future Work: The paper implicitly suggests that further exploration of ensemble strategies and potentially more advanced pre-trained models for ancient Chinese could lead to even greater improvements.
Summary updated