A top European financial daily faced a persistent problem: journalists were too busy to tag their articles properly, leaving valuable archived content hard to find. AI-powered semantic tagging dramatically improved metadata quality and search performance — without adding a single step to the journalist's workload.
One of Europe's most authoritative sources of financial and business intelligence, this French daily has built its reputation on the depth and reliability of its journalism. Since 2009, the publisher has operated a multi-channel Eidosmedia platform to manage print production, web portals, e-paper delivery, and editorial archives — a sophisticated infrastructure befitting a publication whose content retains commercial value long after the day of publication.
For a business publication serving paying subscribers, the ability to surface the right article at the right moment is not a convenience — it is a core part of the product. Archived articles on markets, companies, and economic policy represent a significant asset, but only if readers can actually find them.
Effective search relies on accurate, consistent metadata: topic tags, named entities, IPTC classifications, and geographic markers that allow retrieval systems to index content precisely. In theory, this tagging is the author's responsibility. In practice, journalists working under deadline pressure routinely skipped the step entirely, or applied tags too superficially to be useful. The result was a growing archive of high-value content that was, in search terms, effectively invisible.
What was needed was a way to generate richer, more consistent metadata without imposing additional burden on editorial staff — or accepting the continued erosion of archive quality.
The publisher adopted the Luxid semantic analysis platform, developed by French technology company Temis, integrated directly into the Eidosmedia editorial workflow.
The process is fully automated up to the point of human review. When a journalist releases a story for publication, the article's XML file is cleaned and reformatted, then transmitted to the Luxid platform where two distinct analytical processes run in parallel. The first performs entity and concept extraction — identifying people, organisations, locations, and topics — and maps them to standard classification schemes including IPTC newscodes. The second performs similarity analysis, cross-referencing the article against the publication's own archive to surface related content.
The enriched metadata is then returned to the editorial workflow, where the journalist is presented with the suggested tags and invited to review them. The interaction is deliberately lightweight: a matter of deselecting terms that do not apply, and optionally adding geographic data via an autocomplete tool linked to an external database. Once confirmed, the finalised metadata is appended to the article and it proceeds to publication and archiving.
The design principle is key: rather than asking journalists to create metadata, the system asks them only to validate it — a task that takes seconds rather than minutes, and one that editorial staff accepted readily.
The impact was immediate and measurable across two dimensions.
First, metadata quality improved significantly: articles were consistently tagged with richer, more accurate descriptors than the editorial team had previously managed manually. Second, the process became reliably completed — because reviewing a pre-populated list requires far less effort than originating one from scratch, compliance rates rose sharply.
The gains were felt directly in the accessibility of the publication's archived content. Subscribers could locate relevant articles more easily, strengthening the perceived value of the archive as part of the overall subscription proposition.
The solution also established a robust metadata infrastructure on which the publisher could build more ambitious capabilities — including advanced search, content recommendations, and personalisation features — in subsequent phases of development.