Blog

Realtime Company Tagging, Entity Relevance, and Address Detection

Over the past few months we’ve released several major improvements to TextRazor’s entity models. These include expanded company coverage, more accurate entity relevance scores, and more precise detection of addresses and locations. We're excited to share a few details.

Realtime Company Indentifier Mapping

Reliable financial identifiers are important when working with company data. Wikipedia and Wikidata links are useful, but their coverage can be patchy, particularly for newly listed or less widely documented businesses. For example, "Apnimed, Inc." began trading on NASDAQ as APMD on July 31st 2026 but, at the time of writing, has no Wikidata or Wikipedia page.

TextRazor can now resolve to Bloomberg OpenFIGI and LSEG PermID for all publicly listed companies across all major global exchanges, and has expanded coverage of private companies too. For "Apnimed", we're now returning links to Crunchbase: apnimed, PermID: 5064638184, FIGI: BBG023ML26P7.

The new system is supported by a real-time company monitoring agent that tracks listed companies, their official websites, and relevant news coverage. This information provides continuously updated context for our entity-disambiguation engine, helping it distinguish between companies with similar names and recognise the contexts in which each business is discussed.

We have also extended our incremental indexing pipeline to track companies preparing to list, often within days of their identifiers being assigned. This allows applications built on TextRazor to identify new IPOs before trading begins.

Bloomberg FIGIs uniquely identify tradable financial instruments and can be mapped to associated tickers and exchanges through the OpenFIGI API. Combined with PermID and our existing entity links, these mappings make it easier to connect company mentions in unstructured text with reference financial datasets and trading systems.

Entity Relevance

TextRazor returns a relevance score with each named entity, indicating its "salience" or how important that entity is to the document’s overall meaning. We've developed a new salience system that carefully analyzes signals including the context, frequency and position of each mention, together with background data from our knowledgebase. This helps highlight the most important entities to the document overall.

Previously, this score relied heavily on the similarity between the document and our existing knowledge of the entity. This provided a useful indication of whether an entity was on topic, but it was less reliable in more complicated documents with multiple related entities, or when processing entities for which we have limited prior knowledge.

We have developed a compact transformer model dedicated to relevance scoring that uses substantially more contextual information, while adding minimal processing latency. As a result, TextRazor can now generate reliable relevance scores even for entities it has identified but never encountered before.

More accurate relevance scores also improve downstream topic tagging and classification. Entities with low relevance now contribute less to the final topic or category score, reducing the influence of incidental mentions.

Relevance scores are particularly useful when filtering entities that have been identified confidently but are not central to the document. For example, a location mentioned only in passing may receive a score close to 0. The primary subject of an article may score close to 1, while supporting entities are more likely to score around 0.5.

The new system is now enabled for all entity extractor requests. Scores remain calibrated on a scale from 0 to 1 and are broadly comparable with scores produced by the previous system.

Better Place and Address Detection

TextRazor can now detect longer, more specific street addresses in addition to broader geographic entities such as cities, regions, and countries. Previously, a string such as "TextRazor Ltd, 86-90 Paul Street, London EC2A 4NE" would return the Company match and "London". Enabling the new option will return "86-90 Paul Street, London EC2A 4NE" as a single Place match, alongside its individual components.

This provides a useful first pass for systems that require more precise geographic targeting. The complete address string can then be passed directly to a geocoding service.

To try out the new system just add "entities.includeAddressPlaces" to your requests.

As always, we’d be very grateful for your feedback, in particular how these new features work with your particular documents.