CASENT is a lightweight multi-label entity classification model designed for extremely large label space (e.g., UFET and WikiData). It can also be used for entity extraction and tagging when integrated with a span detector.
CASENT offers several advantages compared to previous methods: 1) Standard maximum likelihood training; 2) Efficient inference through a single autoregressive decoding pass; 3) Calibrated confidence scores; 4) Strong generalization performance to unseen domains and types.
