DSBKTE at ATE-IT: From Token Classification to Zero-Shot Generation: Two Approaches to Italian ATE at EVALITA 2026

This paper presents our participation in Subtask A of the ATE-IT shared task at EVALITA 2026, which targets sentence-level Automatic Terminology Extraction from Italian administrative documents in the waste-management domain. The dataset includes a balanced collection of administrative acts (e.g., municipal regulations, service charters, tenders) and informative texts (e.g., public notices), combining formal institutional language with less rigid informational prose. These heterogeneous documents are characterized by syntactically complex sentences and dense specialized terminology. Given an input sentence, systems must identify all domain-relevant single- and multiword terms and output them in a predefined JSON format, requiring generalization beyond a fixed vocabulary. We investigate and compare two strategies: (i) a supervised neural approach based on an Italian BERT model fine-tuned for token-level term span detection, and (ii) a zero-shot approach using large language models to extract terminology without task-specific training. We report official results and provide an analysis of typical errors to highlight the relative strengths and limitations of supervised versus zero-shot methods for terminology extraction in complex institutional texts.