Negative News Sentiment Pipeline

The ingestion layer must perform Named Entity Recognition (NER) across dozens of languages to identify the target supplier, its subsidiaries, and key executives. This requires transformer models fine-tuned on a cross-lingual knowledge base to resolve 'John Doe' in English and 'Juan Doe' in Spanish to the same entity ID. Without this, the pipeline suffers from entity fragmentation, where a single adverse event is not linked to the correct supplier record.

A binary classifier filters articles for relevance and severity before sentiment analysis. This model is trained on a proprietary taxonomy of adverse event types, including:

• Financial Crime: Fraud, money laundering, sanctions evasion
• Operational Failure: Recalls, data breaches, factory explosions
• Governance Violations: Bribery, executive misconduct, accounting irregularities
• ESG Incidents: Environmental spills, labor strikes, human rights abuses This filtering step prevents the downstream sentiment model from being distracted by neutral operational news.

Unlike generic sentiment models, this component performs fine-grained, aspect-based sentiment analysis. It decomposes an article into specific facets of the supplier—such as 'labor practices,' 'product quality,' or 'financial stability'—and assigns a polarity score to each. For example, an article might be negative on 'financial stability' but neutral on 'product quality.' This granularity prevents a single negative event from incorrectly poisoning the overall supplier score.

A major news event can generate thousands of syndicated articles in hours. The pipeline uses semantic similarity clustering (via sentence embeddings) to group articles reporting the same underlying event. A canonical event ID is assigned, and the system tracks the volume, velocity, and source authority of the cluster. This prevents a single event from being counted as thousands of independent risk signals and provides a true measure of media impact.

Raw sentiment signals are ephemeral. This component applies a time-decay function to historical events, ensuring that a resolved issue from two years ago does not carry the same weight as a breaking crisis. It combines the sentiment polarity, cluster velocity, source credibility, and event taxonomy to output a final composite severity score (0-100). This score is the single, actionable metric pushed to risk dashboards and alerting systems.

Not all news sources are equal. The pipeline maintains a dynamic source authority graph that weights articles based on the publisher's historical accuracy, domain expertise, and reporting rigor. A report from a primary regulatory filing carries more weight than an unverified social media post. This weighting is a critical input to the severity scoring function, ensuring that the system amplifies credible threats and dampens noise from low-quality sources.