Doctopus: A System for Budget-aware Structural Data Extraction from Unstructured Documents
Name
3722212.3725103.pdf
Size
2.82 MB
Format
Adobe PDF
Checksum (MD5)
09ac6a980f32d19247c20a27fc74a174
Author(s) • • • • • •
Zhong, Yuanhao
Deng, Yuhao
Chai, Chengliang
Gu, Ruixin
Yuan, Ye
Wang, Guoren
Cao, Lei
Date Issued
June 22, 2025
Publisher
Association for Computing Machinery
Citation
Yuanhao Zhong, Yuhao Deng, Chengliang Chai, Ruixin Gu, Ye Yuan, Guoren Wang, and Lei Cao. 2025. Doctopus: A System for Budget-aware Structural Data Extraction from Unstructured Documents. In Companion of the 2025 International Conference on Management of Data (SIGMOD/PODS '25). Association for Computing Machinery, New York, NY, USA, 275–278.
Version
Final published version
Abstract
Extracting structured data from unstructured documents is essential for applications like analytical SQL queries and decision-making. Strategies such as pre-trained language models (PLMs) can be employed, but they often fall short in quality. Large language models (LLMs) have shown effectiveness in attribute extraction but are costly, making them impractical for large-scale document sets. To best trade off quality and cost, we present Doctopus, a system designed for accurate and cost-effective attribute extraction. Overall, Doctopus combines LLMs with non-LLM strategies to achieve an optimal quality-cost balance. First, the system employs an indexbased approach to efficiently identify and process only relevant chunks. Afterwards, it further estimates the quality of multiple strategies for each attribute. Finally, based on the cost and estimated quality, Doctopus dynamically selects the optimal strategies through budget-aware optimization. With a real-life scenario, we demonstrate that Doctopus allows users to extract attributes accurately and affordably. The corresponding video is available at https://youtu.be/Cxl_PfvZY10?si=NYoHt2SyD9KHqd6V.
Description
SIGMOD-Companion ’25, Berlin, Germany
Terms of Use
Article is made available in accordance with the publisher's policy and may be subject to US copyright law. Please refer to the publisher's site for terms of use.
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1145/3722212.3725103