OATE at ATE-IT: A Hybrid Approach for Automatic Terminology Extraction in Italian

This paper describes a hybrid system developed for the ATE-IT shared task at EVALITA 2026. The system combines supervised sequence labeling based on a BERT+CRF architecture with semantic post-processing driven by term embeddings. Graph-based ranking methods were also explored but performed poorly since was not suited because of the nature of the dataset. Experimental analysis shows that embedding-based semantic similarity, enhanced with large language model (LLM)–generated domain knowledge, represents the dominant signal for terminology filtering in this setting.