Welcome to NEXT-EVAL, a comprehensive toolkit for the rigorous evaluation and comparison of methods for extracting tabular data records from web pages. This framework supports both traditional algorithms and modern Large Language Model (LLM)-based approaches. We provide the necessary components to generate datasets, preprocess web data, evaluate model performance, and conduct standardized benchmarking.
NEXT-EVAL is an open-source library accompanying the NeurIPS 2025 paper:
📄 NEXT-EVAL: Next Evaluation of Traditional and LLM Web Data Record Extraction [https://arxiv.org/abs/2505.17125]
This library provides tools for generating web-to-table datasets, applying extraction techniques (both traditional and LLM-based), and evaluating them with a standardized framework. It is intended to advance benchmarking and reproducibility in web data extraction research.
Experience NEXT-EVAL in action! Try our interactive web demo where you can upload HTML files or fetch live URLs to preprocess HTML in three different ways (Slim HTML, Hierarchical JSON, and Flat JSON) and compare traditional algorithms with LLM-based extraction methods:
npm install @wordbricks/next-evalConvert real-world webpage HTML into compact formats optimized for LLM processing:
- HTML to Slim HTML: Clean and simplify raw HTML for model input
- HTML to Hierarchical JSON: Structure webpage HTML into nested JSON preserving original structure
- HTML to Flat JSON: Structure HTML into flat JSON format where key is xpath and value is text
import { processHtmlContent } from "@wordbricks/next-eval/html/utils/processHtmlContent";
const htmlString = `<!DOCTYPE html>
<html lang="en">
<body>
<div class="container">
<h1>Main Page</h1>
<div class="card">
<div class="card-title">User Profile</div>
<div class="card-content">
<ul>
<li><strong>Name:</strong> Jane Doe</li>
<li><strong>Email:</strong> [email protected]</li>
<li>
<strong>Skills:</strong>
<ul>
<li>JavaScript</li>
<li>Python</li>
<li>HTML & CSS</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</body>
</html>`;
const { html: slimmedHtml, textMapFlat, textMap } = await processHtmlContent(htmlString);
console.log("[Slim HTML]", slimmedHtml);
console.log("[Hierarchical JSON]", textMap);
console.log("[Flat JSON]", textMapFlat);Generate tabular data from web content using LLM-based extraction with customizable prompts:
import { getLLMResponse } from "@wordbricks/next-eval/llm/utils/getLLMResponse";
const temperature = 1.0; // Control randomness (0.0 to 2.0)
// Option 1: Using Slim HTML format
const { text: slimText, usage: slimUsage } = await getLLMResponse(slimmedHtml, "slim", temperature);
// Option 2: Using Hierarchical JSON format
const { text: hierText, usage: hierUsage } = await getLLMResponse(textMap, "hier", temperature);
// Option 3: Using Flat JSON format
const { text: flatText, usage: flatUsage } = await getLLMResponse(textMapFlat, "flat", temperature);
console.log("Slim HTML result:", slimText);
console.log("Hierarchical JSON result:", hierText);
console.log("Flat JSON result:", flatText);Comprehensive evaluation with precision, recall, F1-score, and detailed overlap analysis:
import { calculateEvaluationMetrics } from "@wordbricks/next-eval/evaluation/utils/calculateEvaluationMetrics";
const predictedRecords = [
[
"/body/section[1]/div[4]/span[1]",
"/body/section[1]/div[4]/span[2]",
"/body/section[1]/div[4]/a[1]",
],
[
"/body/section[1]/div[2]/span[1]",
"/body/section[1]/div[2]/span[2]",
"/body/section[1]/div[2]/a[1]",
],
[
"/body/section[1]/div[3]/span[1]",
"/body/section[1]/div[3]/span[2]",
"/body/section[1]/div[3]/a[1]",
],
];
const groundTruthRecords = [
[
"/body/section[1]/div[3]/span[1]",
"/body/section[1]/div[3]/a[1]",
"/body/section[1]/div[3]/button[1]",
],
[
"/body/section[1]/div[2]/span[1]",
"/body/section[1]/div[2]/a[1]",
"/body/section[1]/div[2]/span[3]",
],
[
"/body/section[1]/div[5]/span[1]",
"/body/section[1]/div[5]/span[2]",
"/body/section[1]/div[5]/a[1]",
],
];
const { precision, recall, f1, totalOverlap, matches } = calculateEvaluationMetrics(
predictedRecords,
groundTruthRecords
);
console.log(`Precision: ${precision.toFixed(3)}`);
console.log(`Recall: ${recall.toFixed(3)}`);
console.log(`F1-Score: ${f1.toFixed(3)}`);
console.log(`Total Overlap: ${totalOverlap}`);
console.log(`Matches: ${matches}`);If you use NEXT-EVAL in your research, please cite:
@inproceedings{next-eval2025,
title={NEXT-EVAL: Next Evaluation of Traditional and LLM Web Data Record Extraction},
author={arXiv},
year={2025},
url={https://arxiv.org/abs/2505.17125}
}We welcome contributions to improve tool coverage, add datasets, or refine evaluation metrics. Please see CONTRIBUTING.md for guidelines.
NEXT-EVAL_.From.Web.URLs.to.Structured.Tables.Extraction.and.Evaluation.mp4
Have questions or ideas? We'd love to hear from you. Contact us at [email protected]
Inspired by our research? We are looking for innovative thinkers to join our team. Please email your resume to [email protected] and be sure to mention our paper.
To see what else we're building, explore our latest technologies at nextrows.com