annotation tag contamination reduced trafilatura's evaluation score #68

The HTML samples in the WebMainBench 545-sample subset were saved from a browser
with annotation plugins active. This injected non-standard DOM elements and
attributes into the saved HTML, affecting parser evaluation fairness. Most of htmls have <marked-tail> / <marked-text>, trafilatura would probably strip these tags and recall less text.

Fix Method

_ANNOTATION_TAG_RE = re.compile(
    r'</?(?:marked-tail|marked-text|marked-inline)[^>]*>', re.IGNORECASE
)
_ANNO_ATTR_RE = re.compile(r'\s+data-anno-uid="[^"]*"', re.IGNORECASE)


def clean_html(html: str) -> str:
    """Strip browser annotation plugin artifacts from saved HTML."""
    html = _ANNOTATION_TAG_RE.sub('', html)
    html = _ANNO_ATTR_RE.sub('', html)
    return html

Evaluation Results

Contamination Sources

1. Content Annotation Plugin (<marked-tail> / <marked-text>)

A Chrome-based content annotation extension (similar to ContentCat) wraps every
text node in custom elements to assign unique IDs for selection and labelling:

<!-- Original HTML -->
<pre><code>import numpy as np
arr = np.array([1, 2, 3])
print(arr)</code></pre>

<!-- After annotation plugin injection -->
<pre><code><marked-text data-anno-uid="abc">import numpy as np</marked-text>
<marked-tail data-anno-uid="def">arr = np.array([1, 2, 3])</marked-tail>
<marked-text data-anno-uid="ghi">print(arr)</marked-text></code></pre>

• <marked-tail> wraps lxml "tail text" (text after a child element)
• <marked-text> wraps lxml "text" (direct text content of an element)

Prevalence in the 545-sample subset:

2. Immersive Translate Extension

A browser translation extension left residual elements in the saved HTML.

Approximately 5% of samples contain actually injected translation text, which
can cause extractors to output duplicate or translated content.