AXE: Low-Cost Cross-Domain Web Structured Information Extraction

TL;DR

AXE tackles the challenge of cross-domain web information extraction by reframing HTML as a pruneable DOM and grounding outputs to exact source nodes. The pipeline combines lossless HTML preprocessing, aggressive DOM pruning, and a small 0.6B LLM with three LoRA adaptors (QA and schema-focused) to produce accurate structured JSON outputs with strong zero-shot generalization. Grounded XPath Resolution ensures every extraction maps to a concrete DOM node, mitigating hallucinations and enabling traceable outputs. Empirical results on SWDE and WebSRC show state-of-the-art zero-shot F1 scores and dramatic token reduction, highlighting AXE’s practicality and cost efficiency for large-scale web information extraction.

Abstract

Extracting structured data from the web is often a trade-off between the brittle nature of manual heuristics and the prohibitive cost of Large Language Models. We introduce AXE (Adaptive X-Path Extractor), a pipeline that rethinks this process by treating the HTML DOM as a tree that needs pruning rather than just a wall of text to be read. AXE uses a specialized "pruning" mechanism to strip away boilerplate and irrelevant nodes, leaving behind a distilled, high-density context that allows a tiny 0.6B LLM to generate precise, structured outputs. To keep the model honest, we implement Grounded XPath Resolution (GXR), ensuring every extraction is physically traceable to a source node. Despite its low footprint, AXE achieves state-of-the-art zero-shot performance, outperforming several much larger, fully-trained alternatives with an F1 score of 88.1% on the SWDE dataset. By releasing our specialized adaptors, we aim to provide a practical, cost-effective path for large-scale web information extraction.

Figures (10)

• Figure 1: Example of input HTML page, the input is the raw HTML code, not the screenshot. Information requested in input schema example in Figure \ref{['fig:input_example']} is bordered with green boxes
• Figure 2: Example of an input schema, requesting information from the page.
• Figure 3: Example of a system output, with correct values for requested queries in input example provided in Figures \ref{['fig:amazon_example']}, \ref{['fig:input_example']}.
• Figure 4: The AXE Pipeline Architecture: showing the flow from Raw HTML through AI-driven Pruning and Extraction to Grounded XPath Resolution.
• Figure 5: Impact of chunk size on SWDE extraction performance. Larger chunks provide beneficial context, with performance peaking around 4000 tokens.