OBIMD: A Multi-modal Dataset for Contextual Interpretation of Oracle Bone Inscriptions

Abstract Oracle bone inscriptions, the earliest known form of Chinese writing, hold immense historical and linguistic significance. However, existing digital datasets are typically limited to isolated characters and lack contextual and structural information essential for comprehensive analysis. We present the Oracle Bone Inscriptions Multi-modal Dataset (OBIMD), a large-scale, publicly available corpus to provide pixel-aligned rubbing and facsimile images, character-level annotations, and sentence-level transcriptions with corresponding reading sequences. OBIMD encompasses 10,077 oracle bone inscription images spanning five phases of the Shang Dynasty, featuring 93,652 annotated characters, 21,667 recorded missing-character positions, 21,941 sentence units, and 4,192 non-sentential elements. By integrating visual, structural, and linguistic modalities, OBIMD supports multi-modal learning and diverse tasks such as facsimile enhancement, character retrieval, and syntactic reconstruction. It constitutes a foundational resource for oracle bone inscription recognition and interpretation, enabling scalable and systematic analysis of ancient Chinese writing.