2026-10-10T10:09:09Z https://www.scidb.cn/oai

10.57760/sciencedb.46110 2026-08-18T14:49:09Z

2026-08-18 SciChart-KR doi:10.57760/sciencedb.46110 en SciChart-KR is a benchmark dataset for structured knowledge extraction from scientific charts. It is designed to evaluate whether multimodal models can recover numerical observations, chart-expressed condition-result relations, and scientific entity roles from figures in scholarly literature.The dataset is built from arXiv papers submitted in June 2025. The raw paper pool is collected through the official arXiv API using a date-only query, without keyword, subject-category, author, or venue filters. After PDF parsing, image extraction, figure-caption anchoring, chart-type filtering, single-panel screening, and manual verification, SciChart-KR contains 1,200 scientific charts across four chart types: line, bar, pie, and radar charts, with 300 charts for each type.SciChart-KR focuses on reusable chart-derived knowledge structures rather than only question answering, caption generation, or chart-to-table conversion. Each record includes source metadata, figure captions, anchored textual context, structured chart data, condition-result chains, and chain-conditioned scientific entity labels. The dataset contains 21,433 annotated condition-result chains and seven entity categories: Problem, Method, Dataset, Material, Tool, Indicator, and Value.The annotation process follows a model-assisted annotation and expert adjudication protocol. Draft JSON annotations are first generated by a model and then independently reviewed by two domain annotators in a blind setting. Inter-annotator agreement before adjudication reaches Cohen’s κ values of 0.85 for condition-result chains and 0.87 for entity types. Disagreements are resolved through annotator discussion and, when necessary, final adjudication by a senior reviewer.SciChart-KR is released as an evaluation-only benchmark. It supports systematic evaluation of causal-chain extraction, chain-conditioned entity recognition, and end-to-end scientific chart knowledge representation. Chart knowledge representation; Chart causal chain; Named entity recognition Donghuan Song RESTRICTED dataset Science Data Bank