Discovering types for entity disambiguation

OpenAI has published a research blog detailing a new approach to entity disambiguation that uses a neural network to assign words to one or more of approximately 100 automatically discovered 'types' (non-exclusive categories). Instead of relying on handcrafted ontologies or large pre-defined knowledge bases, the system learns relevant types from data, enabling it to resolve which specific entity a word refers to in context. The key innovation is that types are discovered automatically, not manually defined, and they are non-exclusive, meaning a word can belong to multiple types simultaneously. This allows the model to handle ambiguity more flexibly—for example, distinguishing between a bank (financial institution) and a bank (river side) by recognizing their different type memberships. For developers building AI workflows that involve text understanding, this research points toward more robust and adaptable disambiguation components that could slot into pipelines for knowledge extraction, search, or conversational AI. However, the blog is research-focused and does not announce a product or API release.