Organization Entity Extraction Prompt for News Articles

This prompt is built for media monitoring pipelines, market intelligence workflows, and news aggregation systems that need to extract organization entities from articles and resolve them to canonical forms. The ideal user is a data engineer or intelligence analyst who must turn unstructured news text into structured, queryable records of every organization mentioned—including legal names, ticker symbols, and parent-subsidiary relationships. The required context is a full news article body, and optionally a target date or a reference knowledge base of known organizations to improve disambiguation accuracy.

Use this prompt when you need a structured JSON payload that downstream systems can ingest without manual cleanup. The prompt handles DBA names, abbreviations, post-merger entity confusion, and ambiguous references by requiring explicit disambiguation rationale in the output. It is designed for batch processing of articles where completeness matters more than sub-second latency. The output schema includes fields for canonical_name, ticker, parent_organization, mention_type, and disambiguation_rationale, making it suitable for populating entity databases, generating market landscape reports, or feeding alerting systems with precise organization references.

Do not use this prompt for real-time streaming with sub-second latency requirements—the disambiguation reasoning step adds inference time that is incompatible with high-throughput, low-latency pipelines. Do not use it for documents where organization mentions are incidental rather than central to the extraction task, such as personal blog posts or fiction. For articles where organizations are mentioned only in passing, a simpler NER span detection prompt will be faster and cheaper. If your use case involves extracting entities across multiple languages simultaneously, pair this prompt with a language-detection pre-processing step and consider the Multi-Language Entity Recognition Prompt Template instead.

Before deploying this prompt, validate that your downstream system can handle null fields when an organization has no ticker or no known parent. The prompt will return an empty string for missing optional fields rather than omitting them, which prevents schema breakage in strict ingestion pipelines. For high-stakes use cases—such as financial compliance monitoring or regulatory reporting—always route low-confidence extractions to a human review queue using the confidence score included in each entity record.

The prompt template below is designed to be dropped into your AI harness. It expects a news article as input and returns a structured JSON payload of organization entities. Every placeholder is enclosed in square brackets and must be replaced by your application logic before the model call. The template enforces strict output typing, confidence annotation, and explicit handling of ambiguous or missing data, making it safe for downstream ingestion into media monitoring dashboards, knowledge graphs, or alerting pipelines.

textCopy
You are an expert entity extraction system specializing in financial and corporate news. Your task is to read the provided news article and extract every organization entity mentioned. For each organization, you must return a structured JSON object that includes the legal name, any mentioned ticker symbol, and the parent organization if a subsidiary relationship is implied or stated. You must disambiguate between the organization and any other entity type (person, location, product).

## INPUT
[ARTICLE_TEXT]

## OUTPUT_SCHEMA
Return a single JSON object with a key "organizations" containing an array of objects. Each object must have the following fields:
- "mention_text": string (the exact text span from the article)
- "canonical_name": string (the resolved legal name, or null if unresolvable)
- "ticker": string | null (the stock ticker if explicitly mentioned, otherwise null)
- "parent_organization": string | null (the canonical name of the parent entity if a subsidiary relationship is stated or strongly implied, otherwise null)
- "confidence": number (a score between 0.0 and 1.0 indicating extraction confidence)
- "disambiguation_note": string | null (explain reasoning if the mention is ambiguous, e.g., 'Apple' as company vs. fruit, or null if unambiguous)

## CONSTRAINTS
- Do not extract organizations that are only referenced as a location (e.g., "Washington" as a city, not Washington Post).
- If an organization is mentioned multiple times, return only one record with the most complete information.
- For DBAs, abbreviations, or legacy names, set "canonical_name" to the current legal name if known from the article context; otherwise, set it to null and note the ambiguity in "disambiguation_note".
- If a ticker is not explicitly stated, do not guess. Set "ticker" to null.
- If no organizations are found, return {"organizations": []}.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK_LEVEL
[RISK_LEVEL]

To adapt this template, replace [ARTICLE_TEXT] with the full body of the news article, ensuring any pre-processing (like paywall removal or HTML stripping) is done in the application layer. The [FEW_SHOT_EXAMPLES] placeholder should be replaced with 2-4 annotated examples that demonstrate the desired behavior for edge cases, such as post-merger entity names, common abbreviations (e.g., "IBM" for "International Business Machines Corporation"), and parent-subsidiary disambiguation. The [RISK_LEVEL] placeholder should be set to a string like "high" or "low" to dynamically adjust the model's caution; for instance, in a high-risk financial compliance setting, you might append an additional instruction to flag any entity with a confidence score below 0.9 for human review. After the model returns the JSON payload, always validate it against the expected schema before allowing it to reach any downstream system. A malformed or missing field can break ingestion pipelines, so implement a retry or repair step if the output fails validation.

This prompt is designed to be called from an application layer, not used as a standalone chat. The typical integration pattern is: fetch a news article, assemble the prompt with the article text and your canonical organization registry, call the model, validate the JSON output against a strict schema, and route low-confidence or malformed results to a human review queue before ingestion. Treat the model call as one step in a pipeline, not the entire pipeline.

Model choice and tool use. Use a model with strong JSON mode or structured output support (e.g., GPT-4o with response_format, Claude 3.5 Sonnet with tool use, or Gemini 1.5 Pro with controlled generation). Do not rely on the model to produce valid JSON without a schema constraint—always pass the expected output schema as part of the API call. If your provider supports it, define the output as a typed object with required fields: organization_name, legal_name, ticker, parent_organization, subsidiary_relationship, mention_context, confidence, and disambiguation_notes. For models without native structured output, include the schema inline in the prompt and add a post-processing validator that rejects malformed payloads.

Validation and retry logic. After receiving the model response, validate that: (1) the JSON parses without error, (2) every required field is present, (3) confidence is a float between 0.0 and 1.0, (4) ticker is either a valid uppercase symbol or explicit null, and (5) organization_name matches at least one span in the source article text. If validation fails, retry once with the same prompt plus the validation error message appended as a correction instruction. If the second attempt also fails, route the article to a human review queue with the raw model output and validation errors attached. Do not silently ingest invalid entity records.

Grounding and hallucination checks. For each extracted organization, verify that the mention_context field contains a verbatim substring from the article. If your application has access to a canonical organization database (e.g., an internal master data table or a service like OpenCorporates), cross-reference the legal_name and ticker against that source. Flag any organization where the model asserts a ticker or parent relationship that cannot be confirmed in your canonical source. This is especially important for post-merger entities, DBA names, and subsidiaries where the model may confidently produce plausible but incorrect resolutions.

Human review and escalation. Set a confidence threshold (we recommend 0.85 for fully automated ingestion) below which extracted entities are routed to a review queue. Additionally, always escalate when: the model flags disambiguation_notes with phrases like 'possible confusion with,' 'ambiguous reference,' or 'post-merger uncertainty'; the article mentions multiple organizations with similar names; or the extracted parent_organization conflicts with your internal hierarchy data. The review interface should show the original article span, the model's extraction, and any canonical database matches side by side.

Logging and observability. Log every extraction call with: prompt version, model identifier, input article hash, output entity count, validation pass/fail, retry count, and whether the result was auto-ingested or escalated. This telemetry lets you track extraction quality over time, detect model drift after provider updates, and identify news sources or organization types that consistently produce low-confidence results. Wire these logs into your existing monitoring stack so that a sudden drop in validation pass rate triggers an alert rather than silently degrading your entity database.