Share your HF username, email, affiliation, and intended use with the dataset maintainers. Access is granted automatically after submission. / 向数据维护者提供 HF 用户名、邮箱、所属机构及使用目的，提交后自动获得访问权限。

OB-Radix

Paper (PDF) · ACL Anthology · Code

Dataset Summary / 数据集简介

OB-Radix accompanies the ACL 2026 paper Specializing Large Models for Oracle Bone Script Interpretation via Component-Grounded Multimodal Knowledge Augmentation.

The dataset provides oracle bone character images, segmented component images, and Chinese and English interpretation annotations. It supports research on component recognition, structured knowledge retrieval, and oracle bone script interpretation. The paper describes an expert annotation workflow using LabelMe for component segmentation and expert review of annotations.

OB-Radix 是上述 ACL 2026 论文的配套甲骨文数据集，提供整字图片、分割后的部件图片及中英文推理标注，用于研究部件识别、结构化知识检索和甲骨文释义。论文介绍了使用 LabelMe 进行部件分割及专家复核的标注流程。

Release Snapshot / 版本概览

Content / 内容	Count / 数量
Character images / 整字图片	1,022
Character classes / 整字类别	934
Component images / 部件图片	1,853
Component image classes / 部件图片类别	478
Chinese interpretation records / 中文推理标注	1,005
English interpretation records / 英文推理标注	1,004
Component annotation records / 部件标注	2,146

The four image statistics match those reported in the paper. Annotation records, character identities, and image files are different counting units. Both character annotation tables cover 918 distinct Character labels. The component annotation table contains 503 distinct Radical labels; this vocabulary is not identical to the 478 component image classes.

上述四项图片统计与论文一致。标注记录、字符类别和图片文件属于不同统计单位。中英文推理标注均覆盖 918 个不同的 Character 标签；部件标注表包含 503 个不同的 Radical 标签，与 478 类部件图片的词表并不完全相同。

Recommended Uses / 推荐用途

Component image recognition and retrieval / 部件图片识别与检索
Character–component knowledge construction / 整字与部件关联知识构建
Knowledge-augmented interpretation generation / 知识增强的甲骨文释义生成
Chinese and English interpretation analysis / 中英文推理标注分析

These are supported research directions, not a claim that this snapshot alone reproduces all paper experiments. Validate annotation-to-image associations before constructing paired samples.

以上为适用的研究方向，不代表仅凭本版本即可复现论文全部实验。构建图文配对样本前，请核验对应关系。

Dataset Structure / 数据组织

OB-Radix/
├── README.md
├── character_explanations_CN.csv
├── character_explanations.csv
├── radical_explanation.csv
├── img_zi/
│   └── <character>/
│       ├── <character-image>.jpg
│       └── <component-image>.png
└── organized_radicals/
    └── <component-label>/
        └── <component-image>.png

img_zi/ groups full character images and component images by character. organized_radicals/ groups the same component images by component label. Its 1,853 PNG files are copies of the component images under img_zi/; they should not be counted as additional independent samples.

img_zi/ 按整字组织原图和部件图；organized_radicals/ 按部件标签组织同一批部件图。这两种组织方式中的 1,853 张 PNG 文件按文件名和文件内容对应，不应重复计为 3,706 个独立样本。

Data Fields / 字段说明

Character annotations / 整字推理标注

character_explanations_CN.csv contains Chinese annotations; character_explanations.csv contains English annotations.

Field / 字段	Description / 说明
`Character`	Character label / 字符标签
`Explanation`	Interpretation annotation, potentially including meaning, structural explanation, or usage / 推理标注，具体记录可能包含含义、构形说明或用法

Repeated character labels occur. Character is not a unique record identifier, and neither table supplies a per-image unique key. Do not assume row-wise bilingual or image alignment.

同一字符可能有多条记录，Character 不是记录唯一标识。表中没有逐张图片的唯一关联键，不能默认按行号将中英文或图片强制配对。

Component annotations / 部件推理标注

Field / 字段	Description / 说明
`Radical`	Component label; the project uses the term radical / 部件标签，项目中使用 radical 命名
`File_Path`	Component image reference / 部件图片引用
`Part_of_Character`	Associated character label / 所属整字标签
`Explanation`	Component interpretation / 部件语义与推理说明

File_Path currently contains two formats: paths containing / are relative to the repository root; bare filenames are resolved under img_zi/<Part_of_Character>/. Historical references may not resolve to a file. Always check existence rather than substituting another component image.

当前 File_Path 有两种格式：含 / 的路径相对于仓库根目录；只有文件名时，在 img_zi/<Part_of_Character>/ 下查找。历史引用可能无法找到文件，应显式检查，不应替换为其他部件图。

Loading the Dataset / 加载方式

Load the annotation tables with datasets / 使用 datasets 加载标注表：

from datasets import load_dataset

zh = load_dataset("6zjn6/OB-Radix", "character_interpretations_zh", split="full")
en = load_dataset("6zjn6/OB-Radix", "character_interpretations_en", split="full")
components = load_dataset("6zjn6/OB-Radix", "component_interpretations", split="full")

Images are available in img_zi/ and organized_radicals/. The examples load annotations only; full denotes the entire table, not an experimental split.

图片位于 img_zi/ 和 organized_radicals/。以上示例仅加载标注；full 表示整张表，不是实验划分。

Recommended Evaluation Use / 建议评测方式

This release does not include predefined train/test splits. For character-level unseen evaluation, group records by character identity, keep train/test character sets disjoint, and document the split procedure and seed. Use explicit sample mappings for bilingual comparisons. For component retrieval, separate support and query images and document handling of rare classes.

当前版本不包含预定义训练/测试划分。进行字符级未见评测时，应按字符分组，保证训练与测试字符集合不相交，并记录划分方法和随机种子。中英文对照需要显式样本映射；部件检索应分离支持集与查询集图片，并说明稀有类别的处理方式。

The original experimental splits and the full set of benchmark-specific annotations are not established by this snapshot alone. Reproducing the reported scores requires the corresponding experimental inputs, code, and configuration.

仅凭本快照不能确定原实验划分及各项基准的完整标注范围。复现论文结果还需对应的实验输入、代码和配置。

Access / 访问

To download from this repository, log in to Hugging Face and complete the access registration form. Access is approved automatically. Your HF username, email, affiliation, and intended use are shared with the dataset maintainers for access registration and understanding declared uses. The CC BY 4.0 license remains unchanged; this form does not track downstream usage.

从本仓库下载前，请登录 HF 并填写访问登记表，提交后自动批准。HF 用户名、邮箱、所属机构及使用目的将提供给数据维护者，用于访问登记和了解申报用途。CC BY 4.0 许可保持不变；登记不会追踪下载后的实际使用。

For Python downloads, first request access on this page, then run hf auth login locally.

使用 Python 下载前，请先在本页面完成登记，再在本机运行 hf auth login。

Citation / 引用

If you use OB-Radix in your research, please cite the ACL 2026 paper / 使用本数据开展研究时，请引用：

@inproceedings{zhang-etal-2026-specializing,
  title = {Specializing Large Models for Oracle Bone Script Interpretation via Component-Grounded Multimodal Knowledge Augmentation},
  author = {Zhang, Jianing and Li, Runan and Pang, Honglin and Xia, Ding and Zhu, Zhou and Zhang, Qian and Li, Chuntao and Yang, Xi},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  url = {https://aclanthology.org/2026.acl-long.1626/},
  doi = {10.18653/v1/2026.acl-long.1626},
  pages = {35221--35236}
}

License / 许可证

This dataset is released under CC BY 4.0.

本数据集采用 CC BY 4.0 许可证发布。

Downloads last month: 58

Total file size:

13.7 MB