newsxlm-markuplm-ae

MarkupLM Baseline for News Attribute Extraction from HTML

This repository contains the fine-tuned MarkupLM baseline evaluated in the NewsXLM paper for news page HTML node classification. This version uses the original HTML text without translating it into English.

• Base model: MarkupLM base
• Fine-tuning dataset: NewsXLM

It labels HTML nodes as title, publication_date, text, author, tag, or none.

🚀 Usage

Installation

pip install -r requirements.txt

Prepare your HTML

html = '''
<html><body>
  <nav><a>Home</a><a>News</a></nav>
  <main><article>
    <h1>Scientists discover new planet beyond our solar system</h1>
    <div class="author"><span class="icon-user"></span>John Smith</div>
    <time>November 3, 2026, 10:30 AM</time>
    <div class="body">
      <p>Scientists announced the discovery on Tuesday.</p>
      <p>The planet is twice the size of Earth.</p>
    </div>
    <div>Tags: <a>Science</a><a>Space</a></div>
  </article></main>
  <footer>All rights reserved</footer>
</body></html>
'''

Attribute extraction

The extraction pipeline groups the results by attribute:

from transformers import pipeline

extractor = pipeline(
    "extraction",
    model="ispras-crawlers/newsxlm-markuplm-ae",
    trust_remote_code=True,
)

extracted = extractor(html)
print(extracted)

Example result:

{
    "title": ["Scientists discover new planet beyond our solar system"],
    "publication_date": ["November 3, 2026, 10:30 AM"],
    "text": ["Scientists announced the discovery on Tuesday.", ...],
    "author": ["John Smith"],
    "tag": ["Science", "Space"]
}

Node classification

The node-classification pipeline returns a prediction for each HTML node.

from transformers import pipeline

node_classifier = pipeline(
    "node-classification",
    model="ispras-crawlers/newsxlm-markuplm-ae",
    trust_remote_code=True,
)

nodes = node_classifier(html)
print(nodes)

Example result:

{
    "/html/body/main/article/h1": {
        "pred": "title",
        "probs": {
            "title": 0.98,
            "publication_date": 0.00,
            "text": 0.00,
            "author": 0.00,
            "tag": 0.00,
            "none": 0.02
        },
        "text": "Scientists discover new planet beyond our solar system"
    }, ...
}

• The model supports tail granularity: text following an element is classified separately from the element's own text. Tail text is returned using the standard XPath suffix /following-sibling::text()[1]. Paths refer to the cleaned DOM.
• Both pipelines process chunks in batches of 32 by default to limit GPU memory use; set chunk_batch_size when creating the pipeline or calling it.

Raw model output

import torch
from transformers import AutoModelForTokenClassification, AutoProcessor

repo_id = "ispras-crawlers/newsxlm-markuplm-ae"

processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForTokenClassification.from_pretrained(repo_id)
model.eval()

inputs = processor(html, return_tensors="pt")
with torch.inference_mode():
    outputs = model(**inputs)

Citation

If you use NewsXLM in your research, please cite:

@article{newsxlm,
  author  = {Bedrin, P. A. and Varlamov, M. I. and Yatskov, A. K.},
  title   = {NewsXLM: A Multilingual Dataset and Model for Information Extraction from News Web Pages},
  journal = {Proceedings of ISP RAS},
  year    = {2026},
  volume  = {38},
  number  = {2},
  pages   = {149--164},
  doi     = {10.15514/ISPRAS-2026-38(2)-10}
}