SMTE at ATE-IT: Ensemble Term Extraction with Italian BERT, spaCy, and Vocabulary-Based Filtering

This paper describes the SMTE system developed for the shared task ATE-IT (Subtask A: Automatic Term Extraction) at EVALITA 2026 in Italian institutional texts in the municipal waste-management domain. Given a sentence, the goal is to extract domain-specific single- and multi-word terms while complying with task constraints such as lowercasing, sentence-level de-duplication, and the removal of nested terms. SMTE follows an ensemble approach that combines (i) a fine-tuned Italian BERT token-classification model trained with BIO tags to detect term spans, (ii) a trained spaCy sequence labeller to provide complementary candidates, and (iii) a vocabulary built from gold training terms to support normalization and filtering. A final post-processing stage cleans and canonicalises predictions, removes truncated spans and nested terms, and formats the output JSON for evaluation and submission. We report our official evaluation results and discuss strengths, typical errors, and directions for future improvements.