Low-resource machine translation has emerged as a critical subfield within computational linguistics [7]. Early approaches relied on rule-based and statistical methods, but recent advances in neural machine translation have shifted the focus toward data-driven models [8]. However, these models require large-scale parallel corpora, which are often unavailable for low-resource languages (LRLs) [9]. Researchers have explored techniques such as transfer learning, multilingual training, and zero-shot translation to mitigate data scarcity [10]. Despite these innovations, issues such as domain mismatch, code-switching, and cultural untranslatability remain largely unaddressed [11]. For example, while multilingual models like mT5 can handle multiple languages, they often struggle with dialectal variations and culturally embedded expressions unique to specific regions like Shandong.

Translation is not merely a linguistic task but also a cultural one [12]. Scholars such as Nida (1964) and Venuti (1995) have long emphasized the importance of cultural equivalence in translation [13]. In digital contexts, this requires systems that can recognize and adapt culture-specific terms, historical context, and rhetorical styles [14]. For projects like the Qilu Library, where texts are deeply embedded in local traditions, a culturally-aware translation system is essential. Recent studies have begun exploring the integration of external knowledge sources, such as ontologies and cultural databases, to enhance MT systems’ ability to handle culture-specific items, though practical implementations remain limited [15].

While models like mBART, mT5, and OPUS-MT have shown promise in multilingual settings, their performance on low-resource languages with dialectal variations remains limited [16]. There is a notable lack of frameworks that integrate cultural experts into the MT pipeline, or that use active learning to continuously improve system output based on user feedback [17]. Furthermore, most existing systems prioritize linguistic accuracy over cultural appropriateness, leading to translations that may be grammatically correct but culturally insensitive or misleading [18]. This gap is particularly critical for cultural heritage projects where contextual fidelity is as important as textual accuracy.

The first step in constructing the translation system is the development of a specialized parallel corpus. This includes not only digitizing and aligning existing Qilu Library materials in Chinese and target languages but also implementing a multi-layered annotation strategy [19]. For instance, classical texts require special markup for archaic terms and historical references, while folk narratives need dialectal annotations and cultural notes. To make the dialect processing more concrete, consider a Shandong dialectal term such as “夜来” (yè lái), which means “yesterday” in standard Mandarin. In the corpus, this term would be annotated with its dialectal provenance (e.g., 夜来), along with a contextual note explaining its usage. This structured annotation enables the model to learn region-specific mappings and supports the generation of culturally-glossed translations (e.g., “yesterday (Shandong dialect: yè lái)”). The corpus building process should also incorporate oral histories and expert-validated translations of key cultural terms to ensure contextual accuracy. Collaboration with cultural institutions is crucial for enriching the corpus with metadata and contextual notes, which can later serve as training data for culture-aware translation models. Moreover, techniques such as web scraping of relevant contemporary materials and controlled data augmentation using back-translation can help expand the corpus size while maintaining domain relevance [20].

A hybrid model combining pre-trained multilingual NMT systems with fine-tuning on domain-specific data is recommended. The system should leverage transfer learning from high-resource languages while incorporating domain adaptation techniques specifically designed for cultural content. This includes implementing a cultural knowledge graph that maps key concepts, entities, and their relationships within the Qilu cultural domain. The structure of this knowledge graph would use entities such as historical figures (e.g., Confucius, Mencius), philosophical concepts (e.g., “仁” Benevolence, “礼” Ritual), cultural artifacts (e.g., “泰山石刻” Mount Tai Stone Inscriptions), literary works, and geographic locations as nodes. The edges would define relationships like “is a student of,” “is a core concept of,” “is created in,” or “is located in.” For example, when translating terms like “儒家思想” (Confucianism), the system can reference connected concepts in the knowledge graph to generate contextually appropriate translations [21]. This allows the system to generate not only lexically accurate but also contextually and culturally coherent translations for compound terms or phrases. Additionally, incorporating specialized modules for handling proper nouns and technical terms through transliteration and contextual disambiguation can significantly improve named entity recognition and consistency across different text types and historical periods.

To ensure cultural and linguistic accuracy, a human-in-the-loop mechanism is essential throughout the translation pipeline. This involves not only post-editing by bilingual cultural experts but also their active participation in the pre-translation phase through glossary development and translation memory curation. The distinct roles within this workflow should be clarified: Bilingual cultural experts are primarily responsible for high-level tasks requiring deep domain knowledge, such as defining core cultural terminology, developing and validating the initial cultural knowledge graph, setting translation guidelines for complex texts (e.g., classical philosophy), and performing quality assurance on critical outputs. In contrast, the community engaged via crowdsourcing is better suited for scalable tasks that benefit from collective local knowledge, such as validating dialectal expressions, collecting regional variants of folk narratives, flagging potentially ambiguous translations for expert review, and expanding the corpus with contemporary usage examples. The workflow should be designed as an iterative process where human feedback is systematically incorporated into model retraining. For instance, a dedicated interface can allow experts to flag culturally problematic translations, which are then used to update the knowledge graph and fine-tune the model. Community engagement through crowdsourcing platforms can further enhance dialectal and regional validation, particularly for texts with strong local characteristics. This continuous collaboration between human experts and the AI system creates a virtuous cycle of improvement, gradually enhancing the system’s cultural competence while maintaining efficiency.

Active learning strategies can significantly enhance resource efficiency by prioritizing the translation of uncertain or complex segments for human review. The system can be designed to calculate confidence scores for each translation output, automatically flagging low-confidence segments, such as those containing rare cultural terms or complex syntactic structures, for expert verification. This approach ensures that human effort is focused where it adds most value, thereby improving model performance over time with minimal expert input. Adaptive training cycles can then leverage this human feedback to incrementally update the model, allowing it to learn from corrections and incorporate new linguistic patterns. For example, the system can be retrained weekly with newly verified translations, gradually expanding its coverage of cultural concepts and dialectal variations while reducing error rates in problematic areas.

Traditional metrics like BLEU often fail to capture cultural adequacy, as they primarily measure surface-level similarity to reference translations. This study advocates for a comprehensive evaluation framework that combines multiple assessment methods. Human evaluation rubrics should be developed to systematically assess fluency, accuracy, and cultural appropriateness, with special attention to the handling of culture-specific items. Semantic similarity measures using embeddings can provide additional insights into meaning preservation beyond n-gram matching. Furthermore, implementing a detailed error classification system based on cultural and linguistic typologies enables targeted improvements, for instance, identifying whether errors stem from linguistic misunderstanding, cultural misinterpretation, or contextual insensitivity. This multi-faceted approach ensures that translation quality is assessed holistically, with equal emphasis on technical accuracy and cultural resonance.

To ensure the system’s longevity, it is crucial to establish clear governance and maintenance protocols from the outset. Key implementation challenges must be acknowledged and planned for, including the ongoing cost and availability of bilingual cultural experts for annotation and validation, and the need for dedicated technical expertise to maintain, update, and troubleshoot the system’s complex components (e.g., the knowledge graph, active learning module, and model retraining pipelines). This includes version control for both models and corpora, regular quality audits, and systematic updates based on user feedback and evolving linguistic standards. Developing open-access tools for community contributions can foster a sustainable ecosystem around the translation system, encouraging ongoing improvement and adaptation. Integration with broader digital heritage platforms, such as national digital library initiatives and international cultural databases, enhances the system’s utility and visibility while facilitating resource sharing. Additionally, implementing a modular architecture allows for component-level upgrades and adaptations, enabling the system to evolve alongside technological advancements and expanding requirements. These measures collectively ensure that the translation system remains relevant, accurate, and accessible over the long term, maximizing its impact on cultural preservation and dissemination.