Daejeon, July 27–31
A critical bottleneck in the machine translation of Classical Chinese is zero anaphora, the pervasive omission of pronouns for established person entities. Even state-of-the-art (SOTA) Large Language Models (LLMs) struggle to link these zero pronouns to their antecedents, particularly within the long-context and interpretively-dense structures of historical text. This challenge persists because specialized models lack explicit coreference mechanisms (Cao et al., 2024; Yao et al., 2025), and the scarcity of annotated corpora makes supervised clustering infeasible. Furthermore, standard evaluation metrics assume a single ground truth, failing to account for the ambiguity inherent in classical philology (Moosavi & Strube, 2016; Wu et al., 2025).
We address these limitations by reframing the task not as labeling, but as referential explicitation. Grounded in translation theory (Vinay & Darbelnet, 1958), we treat coreference resolution as a source-side generative rewrite. We introduce Llama3-classical-ex, fine-tuned to materialize implicit references directly within the source text before translation. This approach transforms context-unaware inputs into explicit structures without requiring large-scale entity annotations.
We evaluate the impact of explicitation on translation quality using an LLM-as-a-Judge framework, rigorously assessing referential consistency and factual fidelity across diverse model families. Validated by alignment with human-expert judgement, our results demonstrate that resolving incomplete references prior to translation significantly reduces hallucinations. By prioritizing interpretable source rewrites, this pipeline offers a practical, Digital Humanities (DH)-oriented method for preserving the semantic integrity of under-represented cultural heritage texts.
This research introduces a novel referential explicitation framework structured into a four-stage pipeline encompassing data synthesis, model optimization, and domain-aware validation, as illustrated in Figure 1.
Figure 1 The generic framework of the research.
To overcome the scarcity of annotated corpora, we construct a synthetic dataset using biographical texts from the History of Song and History of Ming via Wikisource https://zh.wikisource.org/wiki/Wikisource
The dataset comprises 579 paragraphs for training—automatically annotated following the above rules—and 143 paragraphs for testing, which are manually annotated by a human expert. (see Figure 2).
Figure 2 Length distribution of original (pre-EX) and explicitated (post-EX) paragraphs in the datasets.
We develop Llama3-classical-ex, a specialized model for paragraph-level referential explicitation. Starting with our domain-specific baseline (Llama3-classical Baseline configuration: Built on the NiuTrans Open Source Classical Chinese Corpus. Continual Pre-training (CP) utilized monolingual texts. Supervised Fine-Tuning (SFT) employed a balanced 50/50 Classical-Mandarin bilingual dataset. Explicitation Training: The composite dataset integrates the baseline's bilingual Classical-Modern corpus with our 579 synthetically annotated paragraphs containing explicitations.
This joint training approach provides simultaneous supervision for translation and explicitation, mitigating catastrophic forgetting of general linguistic capabilities (Kirkpatrick et al., 2017). The model performs paragraph-level rewrites, explicitly embedding contextually implied references into the source text prior to translation to minimize information ambiguity.
We employ a hybrid evaluation framework to rigorously benchmark translation improvements:
Table 1 Comparative analysis on decisive rounds (Ndec). The full test set contains 143 samples per model (Ndec = 143 - Tie). Win rate is calculated over Ndec .The 95% C.I. and bootstrapped p-values (10k resamples) test the null hypothesis of 0.5 win rate.
Table 2 Alignment analysis on 140 stratified samples. Statistical significance is tested via the Wilcoxon Signed-Rank Test (H0 : Agreement ≤0.5, H1 : Agreement >0.5).
To diagnose the limitations of generative explicitation, we conducted a granular taxonomic analysis on the 17 instances where the human expert strictly preferred the baseline (pre-EX) translation. We classified failure modes into four non-exclusive categories (Table 3), as a single erroneous rewrite often triggers cascading translation faults.
The most prevalent issue was Factual Error (70.6%), where the model confused identities or genealogical relationships not present in the source. This was often compounded by Semantic Imprecision (35.3%), where the model incorrectly resolved collective nouns (e.g., zhujiang, "generals") into specific single entities. Less frequently, we observed Omission (23.5%) of tactical details and Over-Explicitation (23.5%), where the recursive insertion of titles (e.g., "Ming Ming Daizong") degraded the stylistic quality of the output.
Table 3 Qualitative analysis of error propagation. Frequency is calculated over the failure subset (N = 17); and category counts are “non-exclusive”. Examples trace the generation from original Source (SRC) to the Explicitated intermediate (EXP) to the final Translation (TRANS).
The explicitated inputs yield superior translation accuracy compared to the baseline, as validated by domain experts (65% alignment rate). The significant statistical alignment between the automated judge and human verdicts validates our scalable protocol for historical translation research.
While the framework proved robust, the error analysis highlights a trade-off: while explicitation resolves syntactic ambiguity, it introduces a risk of factual hallucination in complex contexts. The model occasionally "over-interprets" historical data, conflating kinship terms with personal names (see Table 3, Factual Error). However, the high overall win rates across models confirm that restoring critical context generally outweighs these edge cases. Future work will mitigate these errors by incorporating uncertainty estimation to suppress low-confidence rewrites.
This study addresses the persistent challenge of referential ambiguity in Classical Chinese translation by introducing a novel framework for source-side explicitation. By constructing a synthetic dataset from the History of Song and History of Ming and fine-tuning Llama3-classical-ex, we demonstrated that resolving zero anaphora and ambiguous entities prior to translation significantly enhances target text fidelity.
Our rigorous two-staged evaluation confirms the efficacy of this approach. The explicitated inputs yielded superior translations compared to the baseline across models, as validated by domain experts. Furthermore, the significant statistical alignment (p ≈ 0.002) between the automated LLM judge and human verdicts establishes the reliability of our evaluation pipeline, offering a scalable protocol for future historical translation research.
While the framework proved robust, our taxonomic error analysis highlighted specific limitations regarding semantic imprecision and identity conflation in complex genealogical contexts. These findings suggest that generative explicitation is a powerful but not infallible tool. Future work should focus on integrating uncertainty estimation mechanisms to filter low-confidence resolutions, thereby minimizing the propagation of factual errors in automated historical text processing.