MHEL-LLAMO: A New Multilingual Historical Entity Linking Method Based on Large Language Models

This article introduces an unsupervised multilingual historical entity linking method called MHEL-LLAMO, which combines confidence-driven prompt chain technology with dual-encoder retrieval and achieves leading performance on multiple historical text benchmarks. Subsequent floors will elaborate on the challenges in this field, method innovations, performance results, technical details, application value, and other content.

Historical entity linking faces many challenges: 1. Language characteristics of historical texts: spelling variations, outdated vocabulary, grammatical structure differences, and contextual historical specificity make modern NLP tools difficult to apply directly; 2. Dilemmas in multilingual scenarios: different languages have distinct historical evolution paths, and low-resource languages lack annotated data; 3. Limitations of traditional methods: relying on supervised learning, while annotated data for historical texts is costly to obtain and hard to scale, so unsupervised/weakly supervised methods have become the focus of research.

Proposed by a European team, MHEL-LLAMO's core innovation is a confidence-driven unsupervised prompt chain method, combined with dual-encoder retrieval and LLM ranking, forming an end-to-end solution. Its architecture is divided into two stages: 1. Candidate Entity Retrieval: Use the multilingual dual-encoder model BELA to retrieve candidate entities and metadata from the knowledge base; 2. Candidate Ranking and NIL Prediction: Rank candidates via LLM prompt chain technology, predict whether they are NIL (entities not present in the knowledge base), and provide confidence scores.

MHEL-LLAMO's prompt chain decomposes the task into two subtasks: NIL detection and candidate ranking, instead of a single-round prompt. The confidence mechanism is key: it estimates decision confidence by analyzing the probability distribution and consistency patterns of LLM outputs. This information is used to dynamically adjust the number of candidates—high confidence reduces candidates to improve efficiency, while low confidence increases candidates to enhance recall, balancing efficiency and accuracy.

MHEL-LLAMO was evaluated on four benchmarks: HIPE-2020 (German, English, French), NewsEye (German, Finnish, French, Swedish), AJMC (German, English), and MHERCL (English, Italian). When using Mistral-Small-24B, key F1 scores include: HIPE-2020 English 0.723, French 0.692; NewsEye French 0.662; MHERCL English 0.700. Low-resource languages (Finnish, Swedish) showed competitive performance. For resource-constrained scenarios, Mistral-8B (English, French, German) and Gemma-3-12b-it (Swedish) still provide competitive performance.

The open-source implementation of MHEL-LLAMO considers engineering details: since dual encoders and LLMs have different dependencies, it is recommended to create two conda environments to run the retrieval and ranking stages separately, resolving dependency conflicts and supporting independent optimization. The code structure includes three scripts: candidate generation, filtering and prompt chain, and evaluation. Users can configure hyperparameters such as the number of candidates, confidence threshold, and model selection via command-line parameters, and a detailed configuration guide is provided.

MHEL-LLAMO is of great significance to digital humanities research: entity linking is a key step in the semanticization of historical texts. This tool requires no annotated data, supports multiple languages, and has reliable performance, lowering the threshold for semantic analysis of historical texts. Libraries, archives, and historical research institutions can use it to quickly build historical knowledge graphs, supporting richer historical research and cultural exploration.

MHEL-LLAMO opens up new directions for historical entity linking. Future explorations can include: 1. More powerful multilingual pre-trained models; 2. More refined confidence calibration methods; 3. Transfer learning strategies across historical periods. Additionally, this work shows that general AI needs to be combined with domain knowledge (historical linguistics, digital humanities) to maximize its value.