Multi-label entity disambiguation using ~100 auto-discovered types

AI Impact Summary

A neural network is now used to assign each word to multiple automatically discovered types (about 100), enabling multi-label entity disambiguation. This broadens disambiguation surface beyond single-type choices, providing richer signals for downstream tasks such as search, retrieval, and QA, but introduces new requirements for evaluating label quality, drift, and integration complexity. Expect increased inference latency and storage for per-token type vectors, plus the need to adapt downstream pipelines to consume and fuse multi-type outputs rather than a single canonical type.