World V2.0 is now Generally Available: A Point-in-Time Database of Global Events

V1.2 used spaCy to identify named entities, but its results were often imprecise. We needed cleaner names and types for researchers who filter events by entity, and for the models that extract relationships, find locations and map events to tickers.

In V2.0, we fine-tuned GLiNER V2.5 to identify people, organizations, products, locations and geopolitical entities, then normalize their names. Each entry in the entities field includes a type, mention count and confidence score for the extracted entity.

An article can be relevant to a company without mentioning its name. News about Zepbound, for example, can matter to Eli Lilly. V1.2 missed some of these connections, leaving researchers without events they could use in company-level signals and backtests.

V2.0 matches company names and ticker symbols, then uses Atlas relationships, personalized PageRank and a reranker to find and score additional companies. Each match includes identifiers, a score and an explanation. Read our ticker mapping analysis.

V1.2 classified events using embedding similarity, which produced noisy ontology labels. Researchers need more precise classifications to select events by topic or industry, set confidence thresholds and assemble datasets for investment research.

V2.0 uses EmbeddingGemma and trained heads across 15 taxonomies. It returns categories and probabilities in JSON. In V2.0 we keep only classifications above a 50% confidence to reduce JSON payload size.

One event can affect the entities involved differently. In a lawsuit, for example, a ruling may be good for one company and bad for another. An event-level sentiment score alone cannot express both assessments for researchers studying the parties involved.

We use a ChronoGPT vintage trained on information available before each event to score whether it is good or bad for each named entity. We correct for an entity-specific baseline and normalize using historical data available before the event. This makes the scores point-in-time safe. The sentiments field gives each entity a label and standardized score for the event.

Names alone don't tell you how entities are connected. Researchers and AI agents also need to know, for example, which company owns a product or has agreed to acquire another company. Those relationships also build the Atlas knowledge graph.

We fine-tuned GLiNER to extract relationships and built an internal reranker to assess them. World stores each result as a subject–predicate–object triple with scores from both models. Atlas aggregates relationships across events into a point-in-time knowledge graph.

World contains events reported in many languages. Researchers working in English need a way to analyze those events together, search their contents and compare findings across markets while retaining access to the original text for inspection.

We use Google's TranslateGemma to translate non-English event titles, descriptions and key points into English. The original text remains in event, while the translation is stored in english. For events originally in English, the field is unchanged from the source text.

Events can involve more than one place. A single location can miss the geographic scope of a cross-border development and leave researchers unable to find an event when searching for another country or geographic region mentioned in the event.

We identify an event's primary geographic context from its title and description, then resolve named locations and geopolitical entities to coordinates. The locations field records each place, how it was found and which is the primary location for the event.

An event's title and description don't always capture every detail a researcher or AI agent needs. Key points add facts about the people, companies and products involved, including how they are connected.

We added event.points, a list of factual points drawn from an event's sources. The points sit alongside the title and description in the original event field, and their English translations are available in english for analysis across languages and markets.

Large events can be reported by thousands of sources. In V1.2, retaining raw timestamps could make event records large and slow to process, while researchers still needed to know when the coverage appeared and how it developed over the lifetime of the event.

V2.0 removes the raw timestamps field and adds availability as a top-level field. It records source-document counts by hour, giving researchers a time series of coverage that they can use without carrying and processing every individual timestamp.