Epitaphs are a literary form used to commemorate the deceased, typically consisting of two parts: the epitaph proper (
While maintaining overall stability, the epitaphs display rich diversity. Styles vary across different eras and social strata, reflecting the profound impact of temporal changes and class differences on literary forms. During the Tang and Song dynasties, epitaphs evolved from parallel prose to prose, marking a significant shift from mourning literature to biographical literature, enhancing their literary nature and realism.
During the Song dynasty, epitaphs became vital materials for historiography. In terms of content, their use expanded beyond the nobility to the common people.This trend continued and intensified in later periods, particularly during the Ming dynasty, leading to the proliferation of epitaph writing. Unlike noble epitaphs, commoner epitaphs focused more on daily life and the inner world, in contrast to earlier records that emphasised official positions and achievements. Understanding this diversity will aid the development of large language models that can provide customised services tailored to various subject areas.
From the Wei, Jin, Southern, and Northern dynasties to the Sui and Tang dynasties, epitaphs developed standardised structures and established writing conventions. By the time of the Southern dynasties’
Rough estimates suggest that more than 15,000 epitaphs from the Northern and Tang dynasties alone have been documented, with numerous new discoveries each year owing to ongoing construction and archaeological excavations. Major collections such as the
Moreover, the number of epitaphs from the Song, Yuan, Ming, and Qing dynasties far surpasses those from the Tang dynasty, although scholarly research on later era epitaphs remains in its early stages. This vast corpus provides a solid foundation for the development of large language models, enabling in-depth linguistic and historical analysis while supporting the structured digitisation of ancient textual resources.
Beyond their literary functions, epitaphs serve as valuable materials that supplement and corroborate historical records. Because of their unique genre and structured composition, the information they contain is often explicit and systematically organised, as reflected in Wang Xing’s thirteen items. This standardised structure facilitates the systematic extraction, integration, and categorisation of epitaphic data into relevant historical databases.
The formulaic nature of epitaphs significantly enhances information retrieval, as different types of information are typically indicated by specific markers. For example, official titles are often introduced by terms such as ‘appointed’ (
Using large language models to extract and analyse information from epitaphs can provide researchers with key data on the deceased and their families, helping to elucidate relationships and social connections, thereby presenting a clear and intuitive knowledge system.
Doubao, Zhipu Qingyan, DeepSeek, and Qianwen are the four major models that support Chinese language training. During data extraction stage, the commercial application programming interface of these models was used. To select the most suitable base model for domain adaptation to epitaphs, the models’ performance was evaluated under zero-shot conditions, that is, without any domain-specific fine-tuning.
Evaluation dataset. We selected three epitaphs from different dynasties: one from the Tang dynasty, one from the Northern Song dynasty, and one from the Ming dynasty. These epitaphs were chosen to assess the ability of the model to generalise across historical periods and text variations.
Evaluation tasks and scoring criteria. Each model was evaluated on four tasks: sentence segmentation, logical judgment, translation, and summary. All outputs were scored on a ten-point scale (1 = completely incorrect, 10 = fully accurate). The scoring criteria for each task is presented in
Evaluation tasks and scoring criteria for models
| Task | Score | Criteria |
|---|---|---|
| Sentence segmentation | 10 | all punctuation correct |
| 8—9 | 1–2 errors | |
| 6–7 | 3–5 errors | |
| ≤5 | more than 5 errors | |
| Logical judgment | 10 | all logical relationships correctly identified |
| 8–9 | minor errors | |
| 6–7 | 2–3 errors | |
| ≤5 | major errors | |
| Translation | 10 | fully accurate and fluent |
| 8—9 | 1–2 minor inaccuracies | |
| 6–7 | 3–5 errors | |
| ≤5 | significant errors | |
| Summary | 10 | all key information captured |
| 8—9 | missing minor details | |
| 6–7 | missing 2–3 key details | |
| ≤5 | missing most key information |
Sentence breaking ability of large language models (Full Score: 10 points)
| Large language model | Epitaph (Tang dynasty) | Epitaph (Northern Song dynasty) | Epitaph (Ming dynasty) | Mean score | Variance |
|---|---|---|---|---|---|
| Doubao | 9 | 9 | 9 | 9 | 0 |
| Zhipu Qingyan | 9 | 9 | 8 | 8.7 | 0.2 |
| DeepSeek | 9 | 9 | 8 | 8.7 | 0.2 |
| Qianwen | 9 | 9 | 8 | 8.7 | 0.2 |
Logical judgment ability of large language models (Full Score: 10 points)
| Large language model | Epitaph (Tang dynasty) | Epitaph (Northern Song dynasty) | Epitaph (Ming dynasty) | Mean score | Variance |
|---|---|---|---|---|---|
| Doubao | 9 | 9 | 8 | 8.7 | 0.2 |
| Zhipu Qingyan | 5 | 5 | 9 | 6.3 | 3.6 |
| DeepSeek | 9 | 9 | 8 | 8.7 | 0.2 |
| Qianwen | 5 | 9 | 9 | 7.7 | 3.6 |
Translation ability of large language models (Full Score: 10 points)
| Large language model | Epitaph (Tang dynasty) | Epitaph (Northern Song dynasty) | Epitaph (Ming dynasty) | Mean score | Variance |
|---|---|---|---|---|---|
| Doubao | 9 | 8 | 9 | 8.7 | 0.2 |
| Zhipu Qingyan | 6 | 8 | 9 | 7.7 | 1.6 |
| DeepSeek | 7 | 9 | 8 | 8 | 0.7 |
| Qianwen | 7 | 8 | 7 | 7.3 | 0.2 |
Summary and induction ability of large language models (Full Score: 10 points).
| Large language model | Epitaph (Tang dynasty) | Epitaph (Northern Song dynasty) | Epitaph (Ming dynasty) | Mean score | Variance |
|---|---|---|---|---|---|
| Doubao | 9 | 9 | 9 | 9 | 0 |
| Zhipu Qingyan | 8 | 9 | 9 | 8.7 | 0.2 |
| DeepSeek | 9 | 9 | 8 | 8.7 | 0.2 |
| Qianwen | 9 | 6 | 4 | 6.3 | 4.2 |
This study primarily collected data from the
To construct a high-quality dataset for domain-specific model development, this study organised the collected data into two complementary components:
The relationship between the two components is as follows: the 6,800-entry epitaph corpus provides the domain knowledge, while the 20,000 instruction–response pairs adapt this knowledge to specific task formats for supervised fine-tuning.
To enhance the model’s ability to comprehend and process epitaph texts, this study first established a knowledge structure based on the characteristics of epitaphs and research requirements, designing a reasonable framework that classifies the data by dynasty, region, and social status of individuals. For the epitaph texts, seven fields were defined and provided professional descriptions to support model understanding:
Epitaph text fields
| Title | Headings located at the beginning of the epitaph |
|---|---|
| Inscriber | The person who inscribed the epitaph, often indicated by terms such as |
| Author | The author of the epitaph text, commonly denoted by words such as |
| Dynasty | The dynasty during which the tomb owner was buried |
| Chinese era date | The traditional Chinese calendar date (year, month, and day) of the tomb owner’s burial is typically expressed using era names |
| Gregorian date | The Gregorian calendar date corresponding to the Chinese era date is numerically represented |
| Full text | The complete text of the epitaph, including title, main body, and colophon |
The data were structured in this manner to enhance the model’s ability to comprehend and process epitaph texts and improve its performance in tasks related to the analysis and interpretation of ancient Chinese inscriptions.
To characterise the epitaph corpus and demonstrate the complexity of the domain, a statistical analysis of the 6,800 texts was conducted.
Statistical characteristics of the epitaph corpus
| Characteristic | Value |
|---|---|
| Total number of epitaphs | 6,800 |
| Total characters | Approx. 4.2 million |
| Average length per epitaph | 618 characters (SD: 342) |
| Shortest epitaph | 48 characters |
| Longest epitaph | 2,156 characters |
| Vocabulary size (unique words) | Approx. 18,500 |
| Average sentence length | 12.4 characters |
| Domain-specific terms | Approx. 1,200 (official titles, toponyms, funeral formulae) |
The length distribution reveals that the majority of epitaphs (62%) fall within the range of 300–800 characters, while short epitaphs (<300 characters) and long epitaphs (>800 characters) account for 18% and 20%, respectively. This diversity in length ensures that models trained on this corpus can handle a wide range of text complexities. The vocabulary comprises approximately 18,500 unique words, with a substantial proportion (approximately 1,200 terms) being domain-specific terminology related to official titles, toponyms, and funeral formulae. This high density of specialised terms underscores the necessity of domain-specific modelling for accurate processing.
To meet the technical requirements for developing domain-specific models and to ensure research reproducibility, the training process in this study was conducted using the enterprise-grade machine learning framework provided by Volcengine. This platform offers hosted infrastructure for distributed training and large-scale model deployment.
Key technologies and implementation approaches for constructing epitaph models.
To further enhance the model’s proficiency in the epitaph domain, this study implemented two key strategies: continued pre-training and instruction fine-tuning. Complementing these training strategies, retrieval-augmented generation uses a local knowledge base as an inference-level enhancement module.
Data preparation: To prevent catastrophic forgetting while preserving the general ability of the underlying model, this study mixed general domain data such as ancient literary texts with a specialised corpus of epitaphs during the continued pre-training phase. The cleaned and organised epitaph data was converted into standard JSON lines format, which is {“text”: “xxxxx”}. Training configuration: Using the pre-training tools provided by the Volcengine platform, the study employed a parameter-efficient fine-tuning method, particularly low-rank adaptation (LoRA). The model was trained on an NVIDIA A100 GPU with a learning rate of 2e-5 and a batch size of 16, for a total of 3 epochs. This balanced approach enabled the model to effectively use high-quality domain data, thereby enhancing its understanding and processing capabilities for tasks related to epitaphs.
Dataset construction: The core of instruction fine-tuning is to create a high-quality, large-scale, and domain-specific dataset. The data were further processed and converted into a conversational JSON format: {“messages”: [{“role”: “user”, “content”: “Question”}, {“role”: “assistant”, “content”: “Answer”}]. Three key methods were used to construct this dataset for epitaphs:
Template-based data integration: Using pre-defined templates to systematically generate instruction–response pairs from the epitaph corpus. Self-instructed data augmentation: Using the model itself to generate additional instruction–response pairs, thereby expanding the dataset and introducing diversity. Document-based Q&A generation: Asking questions based on the content of the epitaph to generate corresponding answers and create an instruction–response dataset. Training configuration: With the supervised fine-tuning function of the Volcengine platform, this study deepened the model’s understanding of the field of epitaphs. The model was optimised using the AdamW optimiser, setting the rank of LoRA to 8 and the learning rate to 1e-5, training it for 5 epochs with a warm-up ratio of 0.05.
Knowledge base construction. The structured metadata of epitaphs (such as official records, geographic data, and historical chronology) was divided into text blocks of approximately 400 tokens. The Doubao Embedding model was used to vectorise these text blocks and establish an index in the vector database. Retrieval mechanism. During the inference stage, the user’s query was transformed into a vector, and the top two most relevant historical records were retrieved through cosine similarity search. By dynamically concatenating the retrieved context with prompt words, this significantly improved the model’s ability to extract key information such as the tomb owner’s name, life story, and time of writing, and successfully associated this information with validated historical knowledge, thereby preventing AI from producing hallucinations.
To evaluate the performance of our domain-specific epitaph model, a human evaluation was conducted.
Three domain experts participated in the evaluation. All evaluators held doctoral degrees in Chinese Philology or Classical Chinese Literature, with specialised expertise in Tang dynasty epitaphs and palaeography. Each evaluator had more than five years of experience in the transcription, annotation, and interpretation of ancient Chinese epitaph texts. Prior to the formal evaluation, each annotator completed a training session comprising twenty example to ensure a consistent understanding of the evaluation criteria.
To assess the necessity of domain-specific fine-tuning and to contextualise the performance of the proposed model, it was compared against three representative general-purpose large language models:
All baseline models were evaluated under zero-shot conditions, without domain-specific fine-tuning or task-specific training.
The evaluation assessed the model’s performance across four task types, each reflecting a core capability required for epitaph processing. For each task, evaluators rated the model’s output on a ten-point scale (1 = completely incorrect or unusable, 10 = fully accurate with no errors). The four tasks were defined as follows:
Three Tang dynasty epitaphs, those of Cui Yi, Liu Yi, and Li Hua, were selected as evaluation materials. These epitaphs were chosen based on the following criteria: (1) they were representative of typical Tang dynasty epitaph structure and content; (2) they vary in length (short, medium, and long) to assess model robustness across text complexity; and (3) their authoritative transcriptions and interpretations have been validated by professional researchers in the field, ensuring reliable ground truth for evaluation.
Each evaluator independently assessed the model’s outputs for all three epitaphs across the four task dimensions. The final score for each epitaph–task combination was determined by the evaluator based on the ten-point rubric. To ensure consistency, evaluators were provided with a detailed scoring guideline with annotated examples for each score level. Inter-rater reliability was assessed using the intraclass correlation coefficient, yielding a value of 0.89 (95% CI: 0.82–0.94), indicating excellent agreement among evaluators. Discrepancies in scoring were resolved through consensus discussion. Evaluators were blinded to model identity to mitigate potential evaluator bias.