TableEval
收藏资源简介:
TableEval是一个针对复杂、多语言和多结构表格问答的基准数据集,旨在评估大型语言模型在真实世界表格问答任务上的能力。该数据集包括从政府、金融、学术和工业报告等四个领域收集的具有各种结构(如简洁、层次结构和嵌套结构)的表格。此外,TableEval还包含简体中文、繁体中文和英文的跨语言场景。为了最小化数据泄露的风险,所有数据均来自最近的现实世界文档。该数据集包含六种高级任务和16种细粒度问题类型,从简单的查找查询到数值计算、比较分析和多跳问题。数据集还包含多轮对话任务,要求模型在多个步骤中动态推理。为了评估模型在表格问答任务中的表现,我们提出了SEAT(Structured Evaluation for Answers in TableQA),这是一种新的表格问答评估框架,它利用LLM与精心设计的提示来比较生成的响应与结构化的参考答案。SEAT通过一个两步过程评估响应的语义正确性:(1)从模型的响应中提取每个子问题的关键答案,并将每个答案与参考答案进行比较,以识别匹配或不同的元素。如果没有提供多个子问题,则将原始问题视为单个子问题。(2)以结构化的JSON格式呈现评估结果,并汇总最终分数,以便轻松跟踪和验证。这种方法提供了一种可靠、可扩展的评估方法,超越了表面级的匹配,提供了更准确的评估。
TableEval is a benchmark dataset for complex, multilingual, and multi-structure table question answering, designed to evaluate the capabilities of large language models (LLMs) on real-world table question answering tasks. This dataset includes tables with various structures (e.g., simple, hierarchical, and nested structures) collected from four domains: government, finance, academic, and industrial reports. Additionally, TableEval covers cross-lingual scenarios in Simplified Chinese, Traditional Chinese, and English. To minimize the risk of data leakage, all data is sourced from recent real-world documents. The dataset contains six high-level tasks and 16 fine-grained question types, ranging from simple lookup queries to numerical calculations, comparative analysis, and multi-hop questions. It also includes multi-turn dialogue tasks that require models to perform dynamic reasoning across multiple steps. To evaluate model performance on table question answering tasks, we propose SEAT (Structured Evaluation for Answers in TableQA), a novel table question answering evaluation framework that leverages LLMs and meticulously designed prompts to compare generated responses against structured reference answers. SEAT evaluates the semantic correctness of responses through a two-step process: (1) Extract key answers for each sub-question from the model's response, and compare each answer with the reference answer to identify matching or differing elements. If multiple sub-questions are not provided, the original question is treated as a single sub-question. (2) Present the evaluation results in structured JSON format and aggregate final scores for easy tracking and verification. This approach provides a reliable, scalable evaluation method that transcends surface-level matching to deliver more accurate assessments.
TableEval 数据集概述
数据集简介
TableEval是首个支持简体中文、繁体中文和英文的跨语言表格问答基准测试集,源自论文《TableEval: A Real-World Benchmark for Complex, Multilingual, and Multi-Structured Table Question Answering》。
核心特性
- 真实领域数据:涵盖财务披露、学术论文、行政记录和行业报告
- 多语言支持:简体中文/繁体中文/英文三语种
- 复杂表格结构:包含层级表头、嵌套单元格、合并布局等
数据规模
- 表格数量:617个经人工校验的Excel表格
- 问答对数量:2,325对
- 任务分类:
- 6大类任务(信息检索、数值分析、推理、数据分析、多轮对话、表格结构理解)
- 16个细分子任务
评估框架SEAT
- 子问题级评估:细粒度分析模型表现
- 可视化验证:通过LLM提取答案并比对参考结果
- 评估指标:采用F1-score,与人工判断高度一致
模型表现(截至2025-03-06)
| 模型名称 | 平均分 | 信息检索 | 数值分析 | 推理 | 数据分析 | 多轮对话 | 表格结构理解 |
|---|---|---|---|---|---|---|---|
| o1-preview | 83.43 | 88.30 | 87.08 | 82.88 | 77.89 | 83.38 | 81.03 |
| claude-3-5-sonnet | 83.32 | 89.62 | 91.06 | 85.76 | 84.01 | 87.94 | 61.51 |
| deepseek-r1 | 82.46 | 90.15 | 88.56 | 87.91 | 77.79 | 78.29 | 72.05 |
完整榜单包含19个模型在6类任务上的表现数据
使用方式
- 环境配置:Python 3.11 + 依赖库(requirements.txt)
- 评估流程:
- 配置API密钥(支持OpenAI兼容接口)
- 生成模型响应(run_prediction.py)
- 执行评估(run_evaluation.py)
文件结构
config/ ├── api_config.yaml # API配置 ├── prompts.yaml # 提示模板 data/ ├── tables/ # Excel表格 ├── TableEval-test.jsonl # 测试集 outputs/ ├── prediction/ # 模型输出 ├── evaluation/ # 评估结果
引用格式
bibtex @misc{zhu2025tableevalrealworldbenchmarkcomplex, title={TableEval: A Real-World Benchmark for Complex, Multilingual, and Multi-Structured Table Question Answering}, author={Junnan Zhu et al.}, year={2025}, eprint={2506.03949}, primaryClass={cs.CL} }




