Ukiyo-e, a major genre of early modern Japanese woodblock prints produced during the Edo period (17th–19th centuries), played a central role in the formation and circulation of visual culture in Japan. As mass-produced prints depicting landscapes, actors, customs, and historical themes, ukiyo-e developed a highly conventionalized titling system that encoded motifs, locations, scene types, and seasonal markers. These titles function not only as labels for individual works but also as structured metadata within catalogues and museum records. Within metadata schemas such as Dublin Core, they correspond to semantically dense dc:title fields (i.e., title metadata fields containing multiple encoded elements).

Because these titles are formulaic, domain-specific, and frequently expressed through historical linguistic variants, their cross-lingual interpretation as structured metadata fields presents significant challenges. Conventional machine translation systems rely on large-scale parallel corpora and therefore struggle with extremely small, specialized, and semantically rigid text types such as Japanese print titles. Even recent multilingual generative models often fail to preserve stable semantic units (i.e., minimal meaning-bearing components such as entities, locations, and motifs) and concept-level consistency (i.e., alignment of expressions referring to the same underlying meaning), producing outputs that drift from historical conventions or collapse distinct semantic units into generalized paraphrases. Such limitations directly affect downstream tasks—metadata integration, multilingual retrieval, and catalogue harmonization—particularly in linked data environments where cross-database interoperability depends on stable semantic alignment.

To address these challenges in a low-resource cultural heritage setting, this study proposes a metadata-aware, two-stage framework that integrates cross-lingual semantic alignment (i.e., mapping expressions across languages to shared meanings) with Chinese-side generative modeling, enabling stable and stylistically coherent production of Chinese titles from Japanese inputs.

Research on cross-lingual semantic spaces has advanced through unsupervised or weakly supervised alignment methods (Artetxe et al. 2018; Conneau et al. 2017) and multilingual sentence encoders such as Sentence-BERT (Reimers / Gurevych 2019). However, these methods assume large or domain-general corpora and thus cannot be directly applied to historically situated, structurally constrained heritage texts. In contrast, floating semantic units within Ukiyo-e titles rarely correspond to modern grammatical segmentation: they function as genre markers, thematic motifs, or encoded visual cues. Within the digital humanities, theoretical work has emphasized the interpretive nature of small, curated datasets (Drucker 2014) and highlighted epistemic risks associated with large generative systems (Bender et al. 2021). Art historical scholarship has likewise noted the instability and variability of Ukiyo-e titles across catalogues and metadata systems (Matsuda 2001). This study operates within these intersecting concerns, exploring how small-domain datasets can be systematically constructed and mobilized to support controlled cross-lingual generation for heritage metadata.

The project begins by assembling a small but carefully validated set of Japanese–Chinese title pairs from catalogues, print colophons, and museum metadata. Rather than treating titles as indivisible strings, the study adopts a unit-level approach that decomposes Ukiyo-e titles into recurrent semantic components—such as 名所 (“famous place” motif), 之図 (“depiction” marker), 見立 (“mitate” allusive parody), and 秋ノ景 (“autumn scene” motif). These units do not correspond to tokens recoverable by general-purpose Japanese tokenizers; instead, they reflect Edo-period genre conventions and must be identified through domain knowledge. Each unit is then embedded into a shared Japanese–Chinese semantic space, where it forms stable cross-lingual neighborhoods with one or more functionally equivalent Chinese expressions. This alignment strategy not only captures structural logic but also provides interpretable mappings to Chinese art-historical naming patterns. To mitigate the scarcity of attested examples, generative AI is used to propose additional expressions, paraphrases, and contextual variants, all of which undergo rigorous human review and normalization to preserve historical fidelity and semantic precision.

Building on this domain-sensitive shared semantic space, the study intro- duces a two-stage framework combining semantic alignment and Chinese-domain generation. When a Japanese title is provided, it is first segmented into semantic components and projected into the shared embedding space, where the system retrieves the most semantically proximal Chinese units. These retrieved units then serve as controlled prompts for a Chinese T5 (Text-to-Text Transfer Transformer; Raffel et al. 2020) generative model trained on a large corpus of traditional Chinese art- historical titles. This corpus largely reflects naming conventions that matured during the late Ming and Qing periods—the same broad historical era in which Ukiyo-e flourished in Japan’s Edo period. During this time, China and Japan maintained sustained cultural contact through imported texts, pictorial themes, and shared aesthetic vocabularies, resulting in partially parallel titling structures such as landscape descriptors, seasonal markers, and motif-based labels. Lever- aging this historically adjacent Chinese titling tradition allows the generator to produce stylistically coherent Chinese titles while remaining aligned with estab- lished art-historical norms. Retrieval provides a crucial constraint that prevents stylistic drift and reduces hallucination by anchoring the generative process to domain-appropriate semantic units. By decoupling cross-lingual semantic inter- pretation from Chinese-side generation, the framework both avoids overfitting to scarce Japanese data and ensures domain-sensitive control. Crucially, because each generated output is paired with its originating Japanese components, the workflow organically produces a growing Japanese–Chinese parallel corpus. This corpus forms a foundation for wider metadata enrichment, cross-collection link- age, and future bidirectional generation—including the inverse task of producing Japanese titles from Chinese art-historical inputs.

The preliminary results demonstrate that even a very small, manually cu- rated dataset—when formalized through a domain-sensitive shared semantic space—can support stable and interpretable cross-lingual mappings across di- verse title contexts. Rather than attempting full-sentence translation, the unit- level modeling approach highlights structural correspondences between Edo- period titling conventions and Chinese art-historical expressions, enabling con- trolled generation in a genuinely low-resource setting. These findings challenge the assumption that high-quality cross-lingual generation depends on large-scale parallel corpora and instead underscore the methodological significance of small, interpretively constructed datasets within heritage-oriented computational work- flows. They further demonstrate that effective cross-lingual metadata alignment can be achieved not through data volume but through representational frame- works sensitive to historical genre constraints.

This study makes three principal contributions to digital humanities and cross-lingual cultural heritage research. First, it proposes a small-data methodol- ogy for constructing domain-specific Japanese–Chinese resources, foregrounding the indispensable role of human annotation and interpretive labor in modeling historically situated texts. Second, it develops a domain-aligned shared semantic space capable of representing the structural logic of Edo-period title components and their functional equivalents in Chinese art-historical nomenclature. Third, it introduces a two-stage embedding-guided alignment-informed generation frame- work that enables controlled production of stylistically coherent Chinese titles and, as a natural byproduct, yields a growing Japanese–Chinese parallel corpus. Future work will expand the curated dataset, evaluate the framework across additional genres and cataloguing traditions, and explore applications in multi- lingual metadata alignment for museum collections, linked open data infrastruc- tures, and cross-collection interoperability—ultimately supporting broader cross- lingual access, cultural interpretation, and scholarly exchange in East Asian art history.