This paper presents a Conditional Random Fields (CRF) sequence-labelling system for sentence-level Automatic Terminology Extraction in the ATE-IT shared task at EVALITA 2026, targeting Italian municipal waste-management documents. The system models term spans using linguistically informed features, including POS tags, lemmas, dependency relations, and morphological patterns, and applies task-specific post-processing to satisfy lowercasing, sentence-level de-duplication, and non-nested output constraints. On the official test set, our submission achieves micro-F1 scores of 0.519 and 0.529, corresponding to ranks 5 and 3 across submitted runs, and yields higher precision than the baseline. However, recall remains comparatively lower, and we observe a substantial gap between training and test performance, suggesting overfitting and limited generalization to unseen documents. We also find only modest gains over a zero-shot large language model approach, indicating that large pre-trained models may encode substantial implicit terminology knowledge even without task-specific supervision. We conclude by outlining directions for improving robustness, including richer semantic representations, data augmentation, and ensemble strategies to reduce the generalization gap.
MKTE at ATE-IT: CRF-Based Term Extraction for Italian Waste Management Documents
Di Nunzio G. M.
2026
Abstract
This paper presents a Conditional Random Fields (CRF) sequence-labelling system for sentence-level Automatic Terminology Extraction in the ATE-IT shared task at EVALITA 2026, targeting Italian municipal waste-management documents. The system models term spans using linguistically informed features, including POS tags, lemmas, dependency relations, and morphological patterns, and applies task-specific post-processing to satisfy lowercasing, sentence-level de-duplication, and non-nested output constraints. On the official test set, our submission achieves micro-F1 scores of 0.519 and 0.529, corresponding to ranks 5 and 3 across submitted runs, and yields higher precision than the baseline. However, recall remains comparatively lower, and we observe a substantial gap between training and test performance, suggesting overfitting and limited generalization to unseen documents. We also find only modest gains over a zero-shot large language model approach, indicating that large pre-trained models may encode substantial implicit terminology knowledge even without task-specific supervision. We conclude by outlining directions for improving robustness, including richer semantic representations, data augmentation, and ensemble strategies to reduce the generalization gap.
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
