This document outlines several AI use cases at FD Mediagroep, a Dutch media company. It describes a personalized newsletter that recommends articles to readers using collaborative filtering. It also describes an entity linker that identifies organizations in news articles using named entity recognition and linking models. Finally, it discusses a "Smart Radio" project to personalize audio content delivery in their radio app.
3. FD mediagroep
Het Financieele Dagblad
Het Financieele Dagblad is dé nieuws- en inspiratiebron die op elk
moment van de dag financieel-economische betekenis geeft aan
ontwikkelingen in de wereld.
Company info geeft altijd real time toegang tot actuele bedrijfs- en
prospectinformatie van 2,5 miljoen bedrijven in Nederland
FD/BNR Networks
FD/BNR Networks brengt ideeën, meningen en talentrijke
ondernemers en professionals samen door de organisatie van o.a.
forums, debatten, netwerksessies, en evenementenreeksen.
BNR Nieuwsradio
BNR Nieuwsradio is de enige radiozender in Nederland waar
ambitieuze en ondernemende mensen 24 uur per dag op de hoogte
worden gehouden van relevant nieuws.
FD Mediagroep helpt om betere beslissingen te nemen door overzicht, inzicht én
(vooruit)zicht te bieden met journalistieke merken en hoogstaande kwalitatieve producten.
6. FD personalized newsletters
• Rich data:
– Logged-in data with unique client ID
• Recommendation model: Item-to-item collaborative
filtering
Published
last week
Older
articles
PySpark:
+ Convert reader-article matrix to IndexedRowMatrix
+ columnSimilarities(): fast compute of cosine
similarity between columns
14. Rich data
1. Knowledge Base (Company.info):
– >2.8M organization profiles
– Rich metadata: sector information, financial information,
people, buildings, etc…
2. News articles:
– >1200 sources (Dutch, online)
• From Het Financieele Dagblad to the Groninger Gezinsbode
– Around ~4k articles per day
– Multiple years of hand-labeled archive
15. Model
• Two stage approach:
– Named-entity recognition
– Entity Linking using binary classification
• Predict whether mention m should be linked to entity candidate c
20. Evaluation
• Take data, make train/test-split
– NER: ~85%
– EL: ~85%
• But: Data is noisy/biased
– + Manual inspection
20
21. Results
– Replace previous manual process
– Number of articles per day: approx. +160%
– Number of linked orgs: approx. +310%
– More “long tail” articles
21
23. Smart Radio – audio personalization
• Current audio listening experience:
– Passive listening
– Long audio data with mix of topics
• New way:
– On demand and personalized listening
TV
Radio
26. Take-home messages
• AI projects add values to existing businesses
• Diversify content delivered to customers: enable
“long tail” articles/audio fragments
• In combination with rich data, simple AI models work
• Our solutions are re-usable