Scholarlink

<p>The rise of Neural Language Models (NLMs) has advanced Natural Language Processing (NLP), yet their application to complex, low-resource domains like the legal field remains limited. This gap is especially pronounced for the Italian language, which lacks high-quality annotated datasets and presents significant syntactic challenges. This paper addresses the task of Named Entity Recognition (NER) in Italian Non-Disclosure Agreements (NDAs), a contract type of high operational relevance. We conduct an empirical comparison of general-purpose and domain-specific Small Language Models (SLMs) to evaluate the balance between computational efficiency and predictive accuracy. Our study is part of the broader ICARUS project, which introduced a novel, expert-annotated NDA dataset—the first resource of its kind for Italian. We systematically benchmark transformer-based architectures, including BERT, RoBERTa, and ELECTRA, alongside their domain-adapted counterpart, Legal-BERT. Performance is assessed using token-level metrics and a granular entity-level error analysis. Contrary to trends observed in some specialized domains, our findings suggest that, within this specific low-resource Italian NDA setting, domain-specific pre-training does not consistently provide a decisive advantage for legal NER. Furthermore, model scale does not linearly correlate with performance gains. This highlights that for niche, resource-constrained tasks, strategically fine-tuned SLMs offer a robust and efficient alternative to larger architectures. We provide a comprehensive account of our experimental protocols to ensure reproducibility and support progress in low-resource legal NLP.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative evaluation of small language models for named entity recognition in low resource Italian legal texts using non-disclosure agreements

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Raffaele Guarasci,
  • Marco Pota,
  • Roberto Abbruzzese,
  • Domenico Alfano,
  • Massimo Esposito,
  • Aniello Minutolo

摘要

The rise of Neural Language Models (NLMs) has advanced Natural Language Processing (NLP), yet their application to complex, low-resource domains like the legal field remains limited. This gap is especially pronounced for the Italian language, which lacks high-quality annotated datasets and presents significant syntactic challenges. This paper addresses the task of Named Entity Recognition (NER) in Italian Non-Disclosure Agreements (NDAs), a contract type of high operational relevance. We conduct an empirical comparison of general-purpose and domain-specific Small Language Models (SLMs) to evaluate the balance between computational efficiency and predictive accuracy. Our study is part of the broader ICARUS project, which introduced a novel, expert-annotated NDA dataset—the first resource of its kind for Italian. We systematically benchmark transformer-based architectures, including BERT, RoBERTa, and ELECTRA, alongside their domain-adapted counterpart, Legal-BERT. Performance is assessed using token-level metrics and a granular entity-level error analysis. Contrary to trends observed in some specialized domains, our findings suggest that, within this specific low-resource Italian NDA setting, domain-specific pre-training does not consistently provide a decisive advantage for legal NER. Furthermore, model scale does not linearly correlate with performance gains. This highlights that for niche, resource-constrained tasks, strategically fine-tuned SLMs offer a robust and efficient alternative to larger architectures. We provide a comprehensive account of our experimental protocols to ensure reproducibility and support progress in low-resource legal NLP.