Result Details
GNN-Based Token Reduction for LLM Semantic Element Detection in E-Commerce Product Pages
Applying Large Language Models (LLMs) to whole-page semantic element detection in e-commerce is prohibitively expensive: typical product pages contain hundreds to thousands of DOM elements, and LLM costs scale linearly with input tokens. We propose a hybrid approach that uses a Graph Neural Network (GraphSAGE) as a pre-filter to reduce the candidate set before LLM processing. Our preliminary results on the Klarna Product Page Dataset show approximately 90% reduction in elements sent to the LLM (from hundreds to 5–10 per page) while maintaining 94.5% nomination accuracy. This position paper presents the core idea, motivates the token-saving focus, and reports preliminary results to solicit feedback before thorough experimental evaluation in an extended journal submission.
E-Commerce | Graph Neural Networks | Large Language Models | Semantic Element Detection | Token Reduction | Web Automation
@inproceedings{BUT212245,
author="{} and Hamza {Salem} and Radek {Burget}",
title="GNN-Based Token Reduction for LLM Semantic Element Detection in E-Commerce Product Pages",
booktitle="Lecture Notes in Computer Science",
year="2026",
journal="Lecture Notes in Computer Science",
number="16625 LNCS",
pages="311--314",
publisher="Springer Science and Business Media Deutschland GmbH",
address="Lyon",
doi="10.1007/978-3-032-29372-5\{_}31",
isbn="9783032293718",
url="https://link.springer.com/chapter/10.1007/978-3-032-29372-5_31"
}