Hybrid NER: Contextual and Rule-Based Named-Entity Recognition for the Minutes of the Austrian Council of Ministers (1848–1918)

Dimitra Grigoriou

Austrian Academy of Sciences, Austria · dimitra.grigoriou@oeaw.ac.at

TopicsEnglish annotation structures systems and methods artificial intelligence and machine learning natural language processing text encoding and markup language creation deployment and analysis text mining and analysis History Humanities computing Informatics Library & information science

Keywordsnamed entity recognition machine learning transformers rule-based augmentation historical documents

The minutes of the Council of Ministers of Austria and Austria-Hungary (1848–1918) are an extraordinary, yet almost entirely neglected source that provides insight into the inner dynamics of the Habsburg Monarchy's highest governing body. Established in the wake of the March Revolution of 1848, the Austrian Council of Ministers had the oversight of political and administrative reorganization thereafter. The minutes of the Austrian Empire document not only decisions and policies but also patterns of communication that defined governance.

Digitized collection of documents from the 1848–1867edition series have basic TEI/XML structured data without semantic annotation. To improve accessibility of these ~3000 documents, it was necessary to identify a suitable NER model to detect entities such as persons, places, and organizations. These were tagged using a Python script. Various free open-source language models were tested, such as spaCy'sde_core_news_lg and dbmdz/historic-ner. Both models showed significant limitations when used on historic data—for example, failing to identify "Graf Stadion" (Count Stadion) or "Freiherr Bach" (Baron Bach) as people—and problematic biases, such as identifying "Juden" (jews) as an organization. This led to the development of a custom model capable of overcoming the challenges of historical entity recognition and mitigating the biases of general-purpose NER systems.

We created an NER model by training and fine-tuning the spaCy de_core_news_lgmodel. We reused and prepared data from Nerdpool (https://nerdpool-api.acdh-dev.oeaw.ac.at/). On a held-out part of this data, evaluation resulted in a precision of 90.1%, a recall of 86.5%, and an F1 score of 88.3%.

Even with these promising metrics, it was not possible to rely entirely on the model with our complete historical collection, since some entities, especially historical names and titles, regularly got skipped or were incorrectly identified. One reason contributing to this issue is that the pretraining of the German base spaCy model is based on static word vectors which do not capture token contexts. Less frequent or out-of-vocabulary (OOV) historical entities were therefore under-represented in the training data and often misclassified.

We decided to proceed with fine-tuning a BERT-based model (google-bert/bert-base-cased). The transformer architecture produces contextual embeddings; it allows for a dynamic “understanding” of tokens within their linguistic environment. The shift to a transformer-based model was expected to improve the recognition of less frequent and context-dependent entities in our historical corpus. Using the same training data, we fine-tune our BERT-based model and publish it on https://huggingface.co/demigrigo/mpr_bert_german_ner. The evaluation metrics improve and yield a 95.1 % precision, 94.2% recall, and an 94.65% F1 score.

Recognition was nearly complete until extended honorifics preceding personal names or titles used in place of personal names posed new challenges. More specifically, while our BERT-based model successfully identified constructions such as "Baron Prohaska" and "Fürst Hohenlohe", it struggled with more complex forms. Constructions with multiple titles attributed to one referent, were often misclassified, similarly as instances where the function title substitutes a name, as in: "Diesen Vorgang findet der Finanzminister nicht recht" (The Minister of Finance does not find this action right.) (Sitzung 508, 1848).

We want to propose a combination of our BERT-model with our spaCy model. We implement a hybrid NER pipeline, combining a transformer-based model with a spaCy-based rule-augmentation layer to improve the recognition of entities. Our core model is our fine-tuned BERT-type German NER system executed via the transformers.pipeline API with aggregated entity spans. The spaCy model is loaded with its internal NER component disabled and serves primarily as a tokenizer and span-management framework.

The pipeline produces entity predictions with the transformer model; this reduces segmentation issues by using the aggregation_strategy="simple" option (grouping subword-level predictions). The predicted spans are aligned to the spaCyDoc object.

Our spaCy component addresses the systematic misclassification of historical personal designations. It targets multi-token noble and bureaucratic titles that precede or stand in place of personal names. We compile a lexicon of historical and administrative titles, normalizing them into token sequences. For token sequences, search in backward-oriented windows up to the maximum observed title length follows. If a match is found, the person entity span is expanded the left.

After span extension, overlapping candidates are resolved using spaCy'sfilter_spans, ensuring longest-span precedence and suppressing redundant entities. The outcome is a unified entity annotation which brings together the contextual transformer-level predictions with deterministic rule-based constraints designed for historical NER.

This allows for better handling of personal references in documents where hierarchical titles, such as "Feldmarschall-Leutnant, Minister des Inneren" (Lieutenant Field Marshal, Minister of the Interior), or "General Freiherr von …" (General Baron von…), form an integral part of entity names also commonly used independent of proper names. The workflow thus overcomes some limitations of transformer-based German NER systems, which perform poorly on multi-token, morphologically rich, and domain-specific person appellations not present in contemporary benchmarks.

This project demonstrates that historical named-entity recognition benefits from a hybrid method combining transformer-based contextual modeling with rule-based augmentation. By fine-tuning a BERT model on historically enriched data and enhancing it with spaCy-driven span-extension logic, we reliably identify persons, places, and institutions in 19th-century Habsburg administrative records, particularly where noble or bureaucratic titles precede or replace personal names. The resulting pipeline provides a scalable, adaptable solution for processing semantically sparse TEI/XML corpora while reducing biases in general-purpose NER tools. The hybrid system improves significantly on recognition of multi-token honorifics and title-based references, but extremely long and nested formulations still pose a problem. Cases include: "Der Kriegsminister brachte in Antrag, den Leutnant von Herrn Erzherzog Ferdinand Maximilian Chevauxlegers, Julius Schill, welcher wegen seines Benehmens bei dem Volkstumulte …" (The Minister of War submitted a motion concerning Lieutenant Julius Schill of His Imperial Highness Archduke Ferdinand Maximilian’s Chevauxlegers Regiment, who, because of his conduct during the popular riot …) (Sitzung 6, 1848) illustrate compound aristocratic designations involving recursive titles, military unit identifiers, and appositional name structures that exceed conventional entity boundaries. These constructions, which are untypical in modern corpora, point to inherent limitations in the span-extension heuristics and underline the linguistic specificity of historical political discourse. Further training with these extreme examples only will be done in the near future.

References

Engel-Janosi, F. (1970) Foreword. In: The Minutes of the Austrian Council of Ministers 1848–1867, Introductory Volume: Council of Ministers and Council of Ministers Minutes 1848–1867: Historical and Archival Analysis. Available at: https://mrp.oeaw.ac.at/pages/show.html?document=MRP-1-0-00-0-00000000-vorwort.xml (Accessed: 13 November 2025).
Devlin, J., Chang, M.-W., Lee, K. and Toutanova, K. (2018) BERT: Pre-training of deep bidirectional transformers for language understanding. Available at: http://arxiv.org/abs/1810.04805
Grigoriou, D. (2025) demigrigo/mpr_bert_german_ner (Version 1.0) [Model]. Hugging Face. Available at: https://huggingface.co/demigrigo/mpr_bert_german_ner (Accessed: 3 November 2025).
Honnibal, M. and Montani, I. (2017) spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing.
Austrian Centre for Digital Humanities (2021) NERDPool: Data Pool for Named Entity Recognition. Available at: https://nerdpool-api.acdh-dev.oeaw.ac.at/ (Accessed: 13 November 2025).
Sitzung 6 (1848) In: ÖMR, Abteilung I, Die Ministerien des Revolutionsjahres 1848, Band 1. Available at: https://mrp.oeaw.ac.at/pages/show.html?document=MRP-1-1-01-0-18480407-P-0006.xml (Accessed: 3 November 2025).
Sitzung 508 (1851) In: ÖMR, Abteilung II, Das Ministerium Schwarzenberg, Band 5. Available at: https://mrp.oeaw.ac.at/pages/show.html?document=MRP-1-2-05-0-18510604-P-0508.xml (Accessed: 3 November 2025).

PreviousBeyond Similarity: Bridging the Interpretability Gap in Intertextual Analysis with LLM-Driven Annotation — A Case Study of the Analects and the Book of Han NextReading in Public: Interpretive Communities and Small-Data Reception of KPop Demon Hunters