Semantic HTML extraction: controlled FR/EN fixtures, seven tools and revised relation evaluation (v1.2.1)
Abstract
This bilingual French/English dataset archives Semantic HTML extraction protocol v1.2.0, published by Edikka on 20 September 2026. Twelve synthetic HTML fixtures form three controlled families: native headings versus styled generic elements; table header cells with scope versus data cells; and an essential caveat in a paragraph versus an aside. Visible text and surrounding markup are held constant within each pair except for the declared treatment. Language variants are not counted as independent replications. Seven pinned tools run twice on identical input bytes: Mozilla Readability 0.6.0, Trafilatura 2.2.0, readability-lxml 0.8.4.1, newspaper4k 0.9.3.1, jusText 3.0.2, html2text 2025.4.15 and markdownify 1.2.2. Five extraction/readability tools and two Markdown converters are described separately. Expected relationships, fixtures, native outputs, observations, hashes and standalone reproduction code are included. A descriptive addendum archives eight complete Edikka page snapshots captured on 11 September 2026. All eight come from one publisher; this convenience sample is neither representative of the Web nor a causal comparison. A newspaper4k replay instability on one English page is retained. Separate evidence records one scripted keyboard interaction with DOM and accessibility-tree observations before and after JavaScript execution. This is not an autonomous-agent benchmark or a rendered-crawl experiment. The protocol measures preservation, flattening or loss of information and relationships in these inputs. It does not measure accessibility conformance, search indexing, rankings, retrieval, model understanding or citation probability. FR : jeu de données bilingue comprenant douze fixtures contrôlées, sept outils aux versions fixées, huit pages Edikka archivées, les sorties brutes, les résultats défavorables, un scénario d’interaction clavier et le nécessaire de reproduction autonome. Les variantes linguistiques ne constituent pas des réplications indépendantes. Aucun gain SEO ou de citation IA n’est mesuré. Release v1.2.1 preserves extraction protocol 1.2.0 and introduces relation-evaluation revision 2.0.0. It re-examines all twelve archived cases for all seven tools without rerunning the extractors. Column-header and row-header observations are separated; each primary native output is identified by field and SHA-256 and accompanied by an excerpt. The original aggregate counted a caption marked role=head as evidence of explicit table headers. Its four counter-test entries were also fixed declarations, not executed checks. The correction executes four deliberate corruptions in each language, after a valid positive control, against the evaluator of archived Mozilla Readability output. These eight checks validate that evaluator; they are not eight new extractor tests or independent Web observations. Original results and a correction note are preserved with the revised data. Release 1.2.1 is a correction to evaluation and documentation, not a new collection or extractor run. Extraction protocol: 1.2.0; original run: 20 September 2026; revised evaluation: 2.0.0 on 21 September 2026. Earlier public archives retain their original bytes. CC BY 4.0 covers original Edikka-authored material within the scope of the supplied licence; third-party software, notices and components retain their own terms. Conflict of interest: Edikka publishes this protocol while offering web development, SEO and AI visibility services. English is the primary metadata language; fixtures and documentation include French and English. The record does not imply peer review.
