Twenty’s Plenty: Semantic Scaffolding and Span Architecture for 19-Label NER in Medieval Latin Charters

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

A study introduces a high-quality, 19-label named entity recognizer for medieval Latin charters, built with only 298 annotated sentences. This system combines "semantic scaffolding," which uses descriptive English label phrases to activate multilingual knowledge, with a custom span-based architecture. The architecture employs XLM-ROBERTa-large, 4-head attention pooling for long descriptions, and a hybrid loss system including Asymmetric Focal-Dice and InfoNCE contrastive terms. Semantic scaffolding enabled fine-tuned GLiNER to achieve 80.8% overlap F1, while the custom architecture reached 83.4% overlap F1. The research also empirically shows that domain-specific pre-training on medieval Latin offers no performance benefit after task-specific fine-tuning. While frequent categories like PER (95.7% F1) and LOC (93.5% F1) perform well, rare legal categories such as LEG (53.1% F1) and TRANS (52.6% F1) remain challenging.

Key takeaway

For NLP Engineers developing NER systems for historical or low-resource languages, this research indicates you can achieve strong performance without extensive domain-specific pre-training. Focus your efforts on innovative techniques like semantic scaffolding and custom span-based architectures, which proved effective with only 298 training sentences. Be prepared for lower performance on rare, position-dependent legal categories, and consider targeted strategies for these specific challenges.

Key insights

High-quality 19-label NER for medieval Latin charters is achievable with limited data via semantic scaffolding and custom architecture.

Principles

• Semantic scaffolding activates latent multilingual knowledge.
• Domain-specific pre-training offers no advantage post-fine-tuning.
• Rare, position-dependent categories remain challenging.

Method

A custom span-based architecture uses XLM-ROBERTa-large, 4-head attention pooling, and a hybrid loss (Asymmetric Focal-Dice, InfoNCE) with semantic scaffolding.

In practice

• Use descriptive English label phrases as prompts.
• Prioritize fine-tuning over extensive domain pre-training.
• Implement span-based models for complex entity descriptions.

Topics

• Named Entity Recognition
• Medieval Latin
• Semantic Scaffolding
• Span Architecture
• XLM-ROBERTa
• Low-Resource NLP

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Semantic marking of named entities in The Lusiads — Hybrid NER workflows improve accuracy for historical literary texts, overcoming limitations of models trained on modern corpora.
• Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan — Synthetic data effectively bootstraps NMT structural learning for low-resource languages, but requires authentic data for semantic depth.
• IHUBERT: Vector-Based Semantic Deduplication and Domain-Balanced Pretraining for Persian Resources — Semantic deduplication and domain-balanced pretraining significantly enhance Persian PLM performance across diverse NLU tasks.
• Lexicon-Grammar Web — A web-based interface provides interactive access to a large-scale European Portuguese grammatical database.
• FineREX: Fine-Tuned NER-RE for Human Smuggling Knowledge Graphs — Domain-specific fine-tuning of LLMs significantly enhances knowledge graph construction quality and efficiency for complex, jargon-heavy legal texts.
• IdioLink: Retrieving Meaning Beyond Words Across Idiomatic and Literal Expressions — Idiom understanding in retrieval requires semantic abstraction beyond lexical overlap.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.