关键词: 多模态, 馆藏数字资源, 语义互联, 双轨信息组织, MARC, 元数据仓储, 字段映射

Abstract:

[Purpose/Significance] Library collections have expanded from printed books to manuscripts, letters, photographs, rubbings, audio-visual materials and born-digital files. This expansion creates a structural difficulty at the level of description. MARC, the Chinese MARC format specified in GB/T 33286-2016, is designed around physical carriers: its fields record physical form, publication data and access points, but offer no designated position for the technical attributes of digital objects, such as image resolution, page count, the presence of a text layer, sampling rate or encoding parameters. Frameworks oriented to digital environments, such as BIBFRAME and IFLA LRM, provide modelling ideas, yet adopting them requires reorganizing existing fields and workflows, and their use in union cataloguing and interlibrary data exchange faces constraints of local adaptation. Most collections are in a transitional state in which physical carriers and digital objects coexist. This paper asks whether two descriptions can run in parallel, so that resources of different carriers and origins each obtain an appropriate position within one system and can check each other. [Method/Process] The paper proposed dual-track information organization, in which tracks are divided by the object of description rather than by the source of resources. The carrier track continues to use MARC, retaining compatibility with union cataloguing and with existing bibliographic records. The digital object track adopts a customized metadata scheme, because the description of a digital object must begin from the digital file itself rather than from a bibliographic conversion. The scheme organizes fields into three tiers: a core set holding identity and access information, a form set recording the structure and technical attributes of the digital object and supplying parameters to image services and processing modules, and a domain set carrying content-related information specific to a resource type; fields are further grouped into ten categories that determine storage location and the destination of mapping to external formats. A field registry is the single authoritative source of field definitions, so that adding a resource type extends mounting relations rather than redefining fields. The two tracks are mapped at the level of MARC subfields rather than whole fields, which prevents information carried in repeated subfields from being lost: of the 115 fields examined, 74 are mapped to MARC and 41 technical and system management fields are stored independently. The tracks are linked by resource identifiers and bibliographic record numbers, and the mapping relation is checked continuously during processing. Two complementary scenarios are then defined－Scenario A, in which a standard MARC record already exists for the physical carrier, and Scenario B, in which it does not－ and the framework is carried through in data production on three resource types held by a public library: manuscripts, rare books and photographs. Multimodal recognition is positioned inside the framework as a means of field production: pages with regular text are processed by local optical character recognition, pages with artistic lettering or mixed layouts are read by a vision-language model, images are cropped into semantic regions before being submitted for recognition, and recognition output is written into designated fields and confirmed by cataloguers. [Results/Conclusions] In Scenario A, existing MARC fields supplied values to the digital object track through the mapping table, form parameters collected during digitization were written back, and information recorded in bibliographic notes－colophons, dedicatory inscriptions, ownership stamps and biographical notes－ was registered in independent domain set fields corresponding to the relevant MARC note fields instead of being crowded into the note area. In Scenario B, a digital object without a bibliographic record generated a MARC draft with fields for standard number, title and statement of responsibility, publication, physical description, summary and electronic location, which entered the formal bibliographic database after review under tiered admission rules, with corrections written back to the digital object track. Field-by-field comparison found no conflicting values for the same field in the three cases, because values are maintained at a single point and referenced elsewhere. Three conclusions follow: the tiered organization of fields makes extension rule-based; field-level mapping and identifier correspondence make the two tracks mutually verifiable, since the digital object track fills the gap left by MARC in describing digital features while MARC supplies authoritative values in return; and the complementarity of the two scenarios lowers the threshold for adoption, because existing holdings can be digitized incrementally under Scenario A while new resources can be digitized before cataloguing under Scenario B. The study has several limitations. First, verification covers one case for each of three resource types and does not report an overall accuracy rate for the complementarity of the two tracks. Second, the domain set vocabulary and the mapping table need to be extended to more fields and resource types. Third, further work should test the framework on larger samples and on audio-visual materials.

Key words: multimodality, digital collection resources, semantic interconnection, dual-track information organization, MARC, metadata repository, field mapping