[Paper Review] Named Entity Recognition and Classification on Historical Documents: A Survey

Maud Ehrmann, Ahmed Hamdi+3 more|arXiv (Cornell University)|Sep 23, 2021

Topic Modeling178 references20 citations

TL;DR

This survey synthesizes challenges, resources, and methodologies for Named Entity Recognition (NER) and classification in historical documents, highlighting issues like OCR errors, linguistic variation, and lack of standardized datasets. It evaluates existing NLP approaches—ranging from traditional machine learning to deep learning—and identifies key research gaps, advocating for improved data curation, multilingual support, and domain-specific adaptation to advance digital humanities research.

ABSTRACT

After decades of massive digitisation, an unprecedented amount of historical documents is available in digital format, along with their machine-readable texts. While this represents a major step forward with respect to preservation and accessibility, it also opens up new opportunities in terms of content mining and the next fundamental challenge is to develop appropriate technologies to efficiently search, retrieve and explore information from this 'big data of the past'. Among semantic indexing opportunities, the recognition and classification of named entities are in great demand among humanities scholars. Yet, named entity recognition (NER) systems are heavily challenged with diverse, historical and noisy inputs. In this survey, we present the array of challenges posed by historical documents to NER, inventory existing resources, describe the main approaches deployed so far, and identify key priorities for future developments.

Motivation & Objective

• Address the growing need for semantic indexing in digitized historical documents to enable efficient information retrieval and exploration.
• Identify core challenges in NER for historical texts, including OCR noise, linguistic variation, and lack of standardized annotations.
• Inventory existing datasets, tools, and NLP techniques tailored for historical document processing.
• Evaluate the performance and limitations of current NER approaches—especially deep learning and transfer learning—on historical corpora.
• Outline key priorities for future research, such as improved data curation, multilingual support, and domain-specific adaptation.

Proposed method

• Systematically survey peer-reviewed literature, datasets, and tools relevant to NER in historical documents from 2000 to 2021.
• Categorize and analyze NER approaches based on their underlying techniques: rule-based, machine learning (e.g., CRF, SVM), and deep learning (e.g., BERT, BiLSTM-CRF).
• Assess the impact of OCR quality and HTR errors on NER performance, emphasizing post-correction and normalization techniques.
• Examine the role of transfer learning and pre-trained language models in improving NER on low-resource historical languages.
• Evaluate the effectiveness of multilingual and cross-lingual NER in historical contexts, especially for underrepresented languages.
• Synthesize findings into a framework for future NER development, emphasizing data quality, annotation standards, and tool interoperability.

Experimental results

Research questions

• RQ1 What are the primary challenges in applying NER to historical documents, particularly regarding OCR errors and linguistic variation?
• RQ2 Which existing datasets and tools are most effective for NER in historical texts, and what are their limitations?
• RQ3 How do different NLP techniques—especially deep learning models—perform on historical document corpora compared to modern text?
• RQ4 To what extent do transfer learning and multilingual pre-trained models improve NER performance on low-resource historical languages?
• RQ5 What are the key gaps and priorities for future research in NER for historical documents in the digital humanities?

Key findings

• OCR and HTR errors significantly degrade NER performance, with studies showing that up to 30% of named entities are misrecognized due to transcription errors.
• Traditional machine learning models like CRF and SVM remain effective on curated, low-noise datasets but struggle with noisy or highly variant historical texts.
• Deep learning models, especially fine-tuned BERT-based architectures, outperform classical methods on most historical NER benchmarks, particularly when pre-trained on related historical corpora.
• Multilingual and cross-lingual models show promise but require careful adaptation to historical language forms, especially for non-modern or rare language varieties.
• The lack of standardized, high-quality, and multilingual annotated datasets remains a major bottleneck for training and evaluating robust NER systems on historical documents.
• Post-correction of OCR output improves NER performance by up to 20% in some cases, indicating that preprocessing is critical for reliable entity recognition.