Automated News Analysis and Bias Detection Platform

asknews · Media & Entertainment · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (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 intelligence and media analysis

AskNews needed to help analysts and journalists process global news at scale, surface contradictions and perspectives across languages, and reduce bias without becoming another opaque editorial authority.

Constraints:

• the platform ingests hundreds of thousands of articles daily
• coverage spans many languages and countries
• cost makes cloud API processing of every article unattractive
• news bias should be surfaced without the system deciding which outlet is right
• citations and transparency are essential for media trust

Architecture

AskNews runs open-source LLMs such as Llama at the ingestion edge to read articles and extract structured metadata: people, facts, allegations, attribution, locations, source origin, reporting voice, relationships, and bias indicators. Articles are embedded into vectors, clustered into topics, and aggregated into a news knowledge graph. Within a cluster, the system enforces diversity by selecting perspectives across countries and languages, then uses higher-intelligence LLM reasoning to compare sources, identify alignments and contradictions, and generate meta-news summaries with citations. Users interact through AskNews/News Plunker graph chat, broad topic queries, API workflows, and forthcoming natural-language alerts.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  article-ingestion-edge[Article ingestion edge] --> open-source-extraction-model[Open-source extraction model]
  open-source-extraction-model --> embedding-and-clustering-layer[Embedding and clustering layer]
  open-source-extraction-model --> news-knowledge-graph[News knowledge graph]
  embedding-and-clustering-layer --> diversity-selection-layer[Diversity selection layer]
  diversity-selection-layer --> contradiction-analysis-layer[Contradiction analysis layer]
  contradiction-analysis-layer --> cited-synthesis-layer[Cited synthesis layer]
  news-knowledge-graph --> graph-chat-interface[Graph chat interface]
  cited-synthesis-layer --> graph-chat-interface
  graph-chat-interface --> programmatic-api[Programmatic API]

Components

Data flow

Open-web and licensed articles are ingested; an open-source edge model extracts facts, people, locations, source metadata, allegations, and reporting voice. The extracted article representations are embedded, clustered into topics, and folded into a knowledge graph. Topic-level queries select diverse country/language perspectives, compare contradictions and alignments, and return cited summaries or graph-chat answers for analysts, journalists, forecasters, and security-risk users.

Why this architecture

AskNews uses open-source models at the edge because per-article cloud API processing would be costly at 500,000 articles per day and because transparency matters for a news-bias product. Embeddings and clustering turn article-level extraction into topic-level comparison. Diversity selection is used to avoid summarizing a conflict from one national or language perspective. High-intelligence LLM reasoning is reserved for cross-source contradiction and graph-level questions, where the system can compare structured metadata rather than reread every article from scratch.

Models used: Llama 2 · Llama 3.1 · open-source LLMs · GPT-4 · closed-source high-level LLMs where trusted

Scale

• AskNews extracts metadata from about 500,000 articles per day.
  “France um so we assign a bunch of metadata to each article that we've read we do that for 500,000 per day and at”transcript, stated
• The embedding layer compares about 500,000 vectors per day.
  “and if you can do that for 500,000 you have 500,000 vectors you can compare vectors mathematically quite easily and”transcript, stated
• The founder says about 60% of processing is devoted to non-English alternative languages.
  “biggest strengths and we devote a lot of processing to other alternative languages besides English approximately 60% we have a transparency dashboard so”transcript, stated
• A highlighted customer is a global security-risk analysis firm.
  “our Cur current customers uh one of our top customers is um Riley risk and they're doing you know Global Security risk analysis and it's you know a a firm”transcript, stated

Alternatives considered (roads not taken)

Run every article through OpenAI-style cloud APIs

Rejected because: At 500,000 articles per day this would be expensive and harder to track transparently.

Label outlets by reputation before reading articles

Rejected because: The founder argues that prejudging outlets reinforces bias; the system should inspect each article and present what it says.

Decide which contradictory source is right

Rejected because: AskNews surfaces conflicts and citations so users can investigate instead of making a final truth judgment for them.

Patterns

structured-output-contract (output-control)

AskNews transforms each article into structured metadata covering people, facts, allegations, attribution, locations, reporting voice, and relationships.

batch-offline-processing (operations)

The platform processes hundreds of thousands of articles per day at ingestion time before users query topics.

retrieval-ranking (retrieval)

Article embeddings are clustered into topics so graph chat and meta-reports operate over semantically related coverage.

model-routing (model-operations)

AskNews uses open-source models for edge enrichment and can use higher-level closed-source models later for graph questioning when users trust them.

Tradeoffs

Use open-source models at the edge

Chose Llama/open-source edge enrichment over cloud API enrichment for every article.

Rationale: Open-source edge processing reduces per-article costs at scale and supports transparency for news extraction.

Surface contradictions instead of adjudicating truth

Chose conflict presentation with citations over system-decided right/wrong labels.

Rationale: The platform is meant to guide investigation and preserve user judgment rather than act as the final authority.

Enforce diversity by country and language within clusters

Chose algorithmic representation of competing perspectives over single-country or single-language topic summaries.

Rationale: Topic summaries should avoid national/language bubbles and show how different perspectives report the same issue.

Gotchas (what bit them)

Source origin is surprisingly difficult to determine.

Root cause: News sources do not always state their country of origin explicitly.

Resolution: Extract and assign source-origin metadata as part of the edge enrichment pipeline.

Bias detection can itself reinforce bias if it starts from outlet reputation.

Root cause: Pre-labeling outlets before article analysis causes users to judge articles by source reputation rather than content.

Resolution: Start clean with each article, extract the content, and present conflicts instead of declaring outlets categorically right or wrong.

Cross-language research creates large manual labor before analysis can even begin.

Root cause: Global news discovery previously required compiling, finding, translating, and tracking sources across languages.

Resolution: Translate and distill global coverage into one location while preserving citations to originals.

Transparency claims depend on the model and extraction layer being inspectable.

Root cause: A news product that claims unbiased extraction needs users to trust the extraction model itself.

Resolution: Train and open-source models so entity extraction and diversification can be audited.

Evaluation

The transcript describes qualitative validation through transparency dashboards, citations, customer feedback from analysts, source-diversified contradiction surfacing, and research use in misinformation alignment; it does not provide conventional offline accuracy metrics.

Metrics: 500,000 articles processed per day · approximately 60% processing devoted to non-English languages · citation click-through to original articles · customer feedback from global security-risk analysts · use by University of Texas misinformation-alignment research

Outcomes

• AskNews processes about 500,000 articles per day into metadata, vectors, clusters, and graph structures.
• Users can query topics, surface contradictions, and trace cited claims back to source articles.
• Analysts save time on cross-language compiling, finding, translating, and tracking.
• Customers include global security-risk analysts, forecasters, journalists, and misinformation researchers.

The platform is built around daily global news ingestion, many languages, country/language diversity, and graph chat over topic-level knowledge.