metadata

language:
  - en
license: mit
size_categories:
  - 100K<n<1M
task_categories:
  - question-answering
pretty_name: WebSRC
dataset_info:
  features:
    - name: domain
      dtype: string
    - name: page_id
      dtype: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: image
      dtype: string
  splits:
    - name: train
      num_bytes: 18678357896
      num_examples: 307315
    - name: dev
      num_bytes: 3401184457
      num_examples: 52826
  download_size: 536103491
  dataset_size: 22079542353
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev
        path: data/dev-*
tags:
  - webdataset

Dataset Card for WebSRC v1.0

WebSRC v1.0 is a dataset for reading comprehension on structural web pages. The task is to answer questions about web pages, which requires a system to have a comprehensive understanding of the spatial structure and logical structure. WebSRC consists of 6.4K web pages and 400K question-answer pairs about web pages. This cached copy of the dataset is focused on Q&A using the web screenshots (HTML and other metadata are omitted). Questions in WebSRC were created for each segment. Answers are either text spans from web pages or yes/no.

For more details, please refer to the paper WebSRC: A Dataset for Web-Based Structural Reading Comprehension. The Leaderboard of WebSRC v1.0 can be found here. The original GitHub Repo is also available.

This flat version of the dataset was specifically compiled to aid Large Multimodal Model (LMM) development, especially in digital domains that need to reason about screens.

Structure

domain: str, broad category of the website
page_id: str, unique ID for the particular page
question: str, the question to answer
answer: str, the actual answer
image: str, a base64 encoded binary string of the image.

The image is converted back to a PIL.Image with:

import base64
import io


def decode_base64_to_image(base64_string):
    img_data = base64.b64decode(base64_string)
    img = Image.open(io.BytesIO(img_data))
    return img

Data Statistics

Questions are roughly divided into three categories: KV, Compare, and Table. The detailed definitions can be found in the original paper. The numbers of websites, webpages, and QAs corresponding to the three categories are as follows:

Type	# Websites	# Webpages	# QAs
KV	34	3,207	168,606
Comparison	15	1,339	68,578
Table	21	1,901	163,314

The statistics of the dataset splits are as follows:

Split	# Websites	# Webpages	# QAs
Train	50	4,549	307,315
Dev	10	913	52,826
Test	10	985	40,357

Note: The test split is not included in this upload. See the original repo for compiling the test set, and how to obtain scores for the test split via submission.

Reference

If you use any source codes or datasets included in this repository in your work, please cite the corresponding papers. The bibtex are listed below:

@inproceedings{chen-etal-2021-websrc,
    title = "{W}eb{SRC}: A Dataset for Web-Based Structural Reading Comprehension",
    author = "Chen, Xingyu  and
      Zhao, Zihan  and
      Chen, Lu  and
      Ji, JiaBao  and
      Zhang, Danyang  and
      Luo, Ao  and
      Xiong, Yuxuan  and
      Yu, Kai",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.343",
    pages = "4173--4185",
    abstract = "Web search is an essential way for humans to obtain information, but it{'}s still a great challenge for machines to understand the contents of web pages. In this paper, we introduce the task of web-based structural reading comprehension. Given a web page and a question about it, the task is to find an answer from the web page. This task requires a system not only to understand the semantics of texts but also the structure of the web page. Moreover, we proposed WebSRC, a novel Web-based Structural Reading Comprehension dataset. WebSRC consists of 400K question-answer pairs, which are collected from 6.4K web pages with corresponding HTML source code, screenshots, and metadata. Each question in WebSRC requires a certain structural understanding of a web page to answer, and the answer is either a text span on the web page or yes/no. We evaluate various strong baselines on our dataset to show the difficulty of our task. We also investigate the usefulness of structural information and visual features. Our dataset and baselines have been publicly available.",
}

Dataset Compilation

Hunter Heidenreich; hunter (DOT) heidenreich at rootsautomation dot com