SpecsQA
收藏资源简介:
SpecsQA是由三星人工智能研究团队构建的基准数据集,旨在评估半结构化文档上的问答系统性能。该数据集包含117个精心设计的问题,数据来源于2025年11月14日抓取的英国三星官方网站快照,涵盖26个产品类别的2162个网页,整合了自然语言描述与规格表格等多模态信息。其创建过程涉及网页爬取、结构化解析和人工问题标注,确保了数据的真实性与复杂性。该数据集主要应用于增强检索生成系统领域,专门解决在半结构化语料库中需要进行精确过滤、聚合及跨文档推理的复杂问答任务,弥补了现有基准在语料库级检索评估上的不足。
SpecsQA is a benchmark dataset developed by the Samsung AI Research team, designed to evaluate the performance of question answering (QA) systems on semi-structured documents. This dataset includes 117 meticulously crafted questions, with data originating from a snapshot of the official Samsung UK website crawled on November 14, 2025, covering 2,162 webpages across 26 product categories and integrating multimodal information such as natural language descriptions and specification tables. Its construction workflow encompasses web crawling, structured parsing, and manual question annotation, ensuring the authenticity and complexity of the dataset. This benchmark is primarily applied in the retrieval-augmented generation (RAG) domain, specifically targeting complex QA tasks that require precise filtering, aggregation, and cross-document reasoning within semi-structured corpora, thereby filling the gap in existing benchmarks for corpus-level retrieval evaluation.
数据集概述:SpecsQA
所属项目:DualGraphRAG(Dual-View Graph-Enhanced Retrieval Augmented Generation Framework)
数据集名称:SpecsQA
来源:2025年11月从三星英国官网(Samsung UK website)爬取的快照。
内容规模:
- 共计 2,162 个产品页面。
- 覆盖 26 个消费电子产品类别。
标注数据:
- 117 个人工标注的问题,分为三类:
- 反向查询(Inverse Queries):要求列出所有产品的穷举性问题。
- 多条件查询(Multi-condition Queries):基于多个技术属性进行筛选。
- 推理与比较(Reasoning & Comparison):需要跨产品进行聚合或比较的问题。
数据用途:作为技术问答(technical question answering)的评估基准,专注于产品规格信息的半结构化问答任务。
数据许可:CC BY-ND 4.0
数据文件:
scraped_data/:原始HTML产品页面(以.tar.xz格式打包)。questions.json:117个标注问题的JSON文件。scraping/:爬取原始HTML数据的工具。databuilder/:预处理工具,用于将原始数据转换为可用于索引的格式。

- 1Query Symbolically or Retrieve Semantically? A Dataset and Method for Semi-Structured Question Answering三星人工智能华沙; 三星人工智能剑桥 · 2026年




