Chart Faithfulness and Insightfulness Benchmark (ChartFI-Bench)
收藏资源简介:
ChartFI-Bench是由复旦大学等机构构建的高质量图表描述基准数据集,旨在系统评估多模态大语言模型生成图表描述的忠实性与洞察力。该数据集包含896对图表-描述样本,其图表视觉复杂度高且描述语义丰富,数据来源于arXiv学术论文,并经过系统的过滤流程与人工核验以确保质量。数据构建过程以事实准确性、显著特征强调、领域知识引导及图文互补性四个维度为指导原则,精心筛选和标注而成。本数据集主要应用于评估和提升自动化图表描述生成模型的性能,旨在解决现有基准在复杂图表和深层语义描述评估方面的不足,推动可访问性、跨模态检索及数据洞察提取等相关领域的发展。
ChartFI-Bench is a high-quality chart description benchmark dataset constructed by Fudan University and other institutions, aiming to systematically evaluate the faithfulness and insightfulness of multimodal large language models when generating chart descriptions. This dataset contains 896 chart-description sample pairs, with highly complex visual charts and rich descriptive semantics. The data is sourced from arXiv academic papers, and has been subjected to systematic filtering and manual verification to ensure its quality. The dataset was constructed following four core guiding principles: factual accuracy, emphasis on prominent features, domain knowledge guidance, and image-text complementarity, with samples carefully screened and annotated. This dataset is primarily utilized to evaluate and enhance the performance of automated chart description generation models, aiming to address the limitations of current benchmarks in evaluating complex charts and deep semantic descriptions, and advance the development of relevant fields including accessibility, cross-modal retrieval, and data insight extraction.




