未提及具体数据集名称
收藏资源简介:
本文提出了一种名为AAA的框架,旨在通过人机协作模型实现可扩展的网站无障碍性审计。AAA框架包含两个关键创新:GRASP,一种基于图的跨模态采样方法,通过学习视觉、文本和关系提示的嵌入来确保页面覆盖的代表性;MaC,一种基于多模态大型语言模型的副驾驶,通过跨模态推理和智能辅助支持审计员在高强度任务中的工作。为了支持未来的研究,我们还发布了四个针对WAA管道不同阶段量身定制的新数据集。
This paper proposes a framework named AAA, which aims to enable scalable website accessibility auditing via human-AI collaborative models. The AAA framework encompasses two core innovations: GRASP, a graph-based cross-modal sampling method that learns embeddings of visual, textual, and relational prompts to ensure representative coverage of web pages; and MaC, a copilot based on multimodal large language models, which supports auditors in high-intensity tasks through cross-modal reasoning and intelligent assistance. To support future research, we also release four novel datasets tailored for different stages of the WAA pipeline.
AAA数据集概述
数据集基本信息
- 数据集名称:AWA Web Accessibility Benchmark
- 发布来源:AAAI 2026 Special Track AI for Social Impact论文《Towards Scalable Web Accessibility Audit with MLLMs as Copilots》
- 存储位置:https://doi.org/10.5281/zenodo.17548393
数据集组成
AWA网络可访问性基准包含四个子数据集:
APR
- 全称:Accessibility-relevant Page Recognition
- 用途:可访问性相关页面识别
CCT
- 全称:CAPTCHA of Cognitive Tests
- 用途:认知测试验证码
TPS
- 全称:Triple-representativeness Page Sampling
- 用途:三重代表性页面采样
- 重要性:GRASP方法唯一使用的数据集
CPE
- 全称:Complete Process Extraction
- 用途:完整流程提取
使用方法
数据准备
- 至少需要下载TPS数据集
- 放置在项目同级目录的
data/TPS文件夹中
实验配置
- 对于APR、CCT和CPE实验,需参考论文附录中的提示词
- 使用任意MLLMs运行实验
引用信息
bibtex @inproceedings{aaa, title={Towards Scalable Web Accessibility Audit with {MLLMs} as Copilots}, author={Ming Gu and Ziwei Wang and Sicen Lai and Zirui Gao and Sheng Zhou and Jiajun Bu}, journal = {Proceedings of the AAAI Conference on Artificial Intelligence}, year={2026}, url={https://arxiv.org/abs/2511.03471}, }

- 1Towards Scalable Web Accessibility Audit with MLLMs as Copilots浙江大学 · 2025年



