Production AI Systems for News Personalization and Journalistic Workflows

bonnier_news · Media & Entertainment · 2024 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: News personalization and journalistic AI workflows

Bonnier News needed AI systems that could personalize content across many publication brands, support journalists with LLM tools, and adapt language models to Swedish media tone without losing editorial control or data sovereignty.

Constraints:

• The publisher spans roughly 200 brands with distinct editorial voices.
• Personalization must scale across brands while still allowing brand-specific tuning.
• Journalists need optional tools rather than replacement workflows.
• Sensitive journalistic material may need internal hosting rather than external APIs.

Architecture

Bonnier News builds a central AI/data-science platform for personalization and journalism tools. Its white-label personalization engine represents articles and readers with embeddings, uses vector similarity and reranking instead of metadata-heavy curation, and can be rolled across brands. LLM workflows generate trigger questions, aggregate multi-brand stories, and propose headlines. A WASP/Linkoping collaboration continues pre-training Llama on Swedish media data, then studies instruction tuning and internal hosting for brand-sensitive work.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  brand-content-feeds["Brand content feeds"] --> article-embedding-pipeline["Article embedding pipeline"]
  brand-content-feeds["Brand content feeds"] --> news-aggregator["News aggregator"]
  article-embedding-pipeline["Article embedding pipeline"] --> vector-similarity-search["Vector similarity search"]
  reader-interest-estimator["Reader-interest estimator"] --> article-embedding-pipeline["Article embedding pipeline"]
  reader-interest-estimator["Reader-interest estimator"] --> vector-similarity-search["Vector similarity search"]
  vector-similarity-search["Vector similarity search"] --> reranking-layer["Reranking layer"]
  reranking-layer["Reranking layer"] --> recommendation-surface["Recommendation surface"]
  news-aggregator["News aggregator"] --> llm-summarization-workflow["LLM summarization workflow"]
  llm-summarization-workflow["LLM summarization workflow"] --> trigger-question-generator["Trigger question generator"]
  headline-generation-tool["Headline generation tool"] --> journalist-workflow["Journalist workflow"]
  swedish-media-pretraining-corpus["Swedish media pretraining corpus"] --> llama-continued-pretraining["Llama continued pretraining"]
  llama-continued-pretraining["Llama continued pretraining"] --> instruction-tuning-research["Instruction tuning research"]

Components

Data flow

A reader consumes articles across Bonnier brands. The personalization system estimates which articles they were most interested in from dwell time, article length, app reading signals, and other heuristics. It embeds those articles, searches vector space for similar content, optionally reranks nearest-neighbor results from multiple top articles, and serves further-reading recommendations or aggregated multi-brand story summaries. Separate journalist tools generate headline options or trigger questions, while domain-adapted Llama research uses Bonnier text data for Swedish media-specific pretraining.

Why this architecture

Bonnier uses embeddings and vector distance because article content produced stronger personalization signal than manual metadata, and a white-label engine lets many brands start from a shared service. Domain-adapted Llama research addresses the separate problem that generic LLMs miss Swedish media tone and may not be acceptable for confidential journalism.

Models used: embedding models · LLMs for summaries/questions/headlines · Llama

Scale

• Bonnier News has far above 200 brands.
  “there's a I mean there's a there's far above 200 Brands”transcript, stated
• One subscription package spans about 50 brands.
  “you can buy a subscription where you get access to something like 50 Brands”transcript, stated
• The personalization engine is described as rolled out at scale.
  “it's rolled out at scale okay so it's been used for yes a lot okay”transcript, stated
• Headline generation experiments ask for five alternative headlines.
  “can we ask an llm to generate five different headlines and then just let the the winning one take over once we publish it”transcript, stated

Alternatives considered (roads not taken)

Manual metadata as the primary personalization signal

Rejected because: The team found vector distance over content superior to manual metadata for personalization.

Single centroid vector for a user

Rejected because: Averaging distinct interests can land in an uninteresting point in embedding space.

Generic LLM headline generation without domain adaptation

Rejected because: Journalists reported that generic model outputs did not capture each publication’s tone.

Training a Swedish model from scratch

Rejected because: The research starts from existing Llama models and continues pretraining rather than building a foundation model from zero.

Patterns

retrieval-ranking (recommendation)

Bonnier retrieves nearest neighbors for top user-interest articles and reranks the combined results.

rag-retrieval (retrieval)

The personalization architecture uses embedding search over article/document vectors, analogous to RAG retrieval.

fine-tuning-adaptation (modeling)

The WASP collaboration continues pretraining Llama on Swedish media data for domain adaptation.

human-in-the-loop-gate (editorial)

Journalists remain the decision-makers for headline and editorial tooling, providing feedback on tone and use.

privacy-preserving-processing (governance)

The internal LLM strategy is motivated by confidential journalism that should not leave the organization.

context-engineering (prompting)

Bonnier treats prompt instructions as terse process instructions rather than conversational requests.

Tradeoffs

Productization strategy

Chose White-label personalization engine over Fully bespoke personalization per brand.

Rationale: A common engine lets brands start quickly; brand-specific fine-tuning can come later.

Personalization signal

Chose Content embeddings and vector distance over Manual metadata and taxonomies.

Rationale: The team found content vectors more useful and less labor-intensive than manual metadata.

Domain model strategy

Chose Continue pretraining Llama over Training from scratch or relying only on external APIs.

Rationale: This reuses open-source foundation capability while adding Swedish media domain knowledge and internal hosting options.

Gotchas (what bit them)

Manual metadata underperformed content embeddings for personalization.

Root cause: Metadata structures require manual labor and did not provide the strongest similarity signal.

Resolution: Use vector distance over article content and reading patterns.

Single-vector user representations can erase diverse interests.

Root cause: Averaging sports, tech, and other distinct interests may land between all of them.

Resolution: Experiment with multiple top articles and reranking rather than only a centroid.

Generic LLMs miss publication voice.

Root cause: Models lack exposure to each brand’s editorial tone and conventions.

Resolution: Research continued pretraining on Bonnier media data and brand-perspective adaptation.

Continued pretraining can lose instruction tuning.

Root cause: Domain pretraining produces a base model rather than an instruction-tuned model.

Resolution: Research how to add instruction tuning back without losing quality.

Evaluation

Bonnier compares personalization against trending/manual baselines, uses journalist feedback for LLM tools, and identifies a need for Swedish/editorial usability benchmarks.

Metrics: lift over trending/popular baselines · manual selection parity · journalist/editor usability · brand tone fit

Outcomes

• The personalization engine is rolled out at scale and performs better than simple trending/popular baselines.
• Bonnier built trigger questions and story aggregation over multi-brand content.
• The Llama/WASP research creates a path toward internal Swedish media LLMs.

The case spans far above 200 brands, a package of roughly 50 brands, a rolled-out personalization engine, and headline generation experiments with five candidates.