TL;DR
This paper tackles the challenge of digitally preserving and analyzing historical encyclopedias by developing a pipeline to match and link entries between different editions of the Nordisk familjebok. By linking geographical entries to Wikidata, they reveal shifts in Sweden's geographic focus over time, influenced by global events.
What Problem Does This Paper Address?
Historical encyclopedias, like the Nordisk familjebok, are invaluable cultural artifacts that capture the knowledge and perspectives of their time. However, their printed format makes them difficult to search and analyze comprehensively. Digitization efforts have made these texts accessible, but extracting structured information and understanding how knowledge evolves across editions remains a significant challenge. This paper specifically addresses the problem of automatically identifying and linking equivalent entries across different editions of the Nordisk familjebok, enabling the study of changes in content and focus, particularly in geographical representation.
Methods
The authors developed a multi-stage pipeline for processing the Nordisk familjebok:
- Scraping and Segmentation: They scraped the digitized text from Project Runeberg and developed a three-step approach to segment the raw text into individual entries, handling cases where bold tags were absent due to OCR limitations. This involved bold matching, index matching, and a binary classifier.
- Location Classification: A binary classifier, using KB-BERT for tokenization and a logistic regression model, was trained to identify entries describing locations.
- Entry Matching: To match entries between the first and second editions, they employed sentence embeddings generated by the KB-SBERT model. These embeddings were stored in a Qdrant vector database, and cosine similarity was used to find the closest matches, with a threshold of 0.9 to filter false positives.
- Wikidata Linking: Geographical entries were linked to Wikidata items. This involved querying Wikidata with entry headwords, retrieving candidate entities, and then using KB-SBERT to encode both the encyclopedia entry and the corresponding Swedish Wikipedia text to find the best match based on cosine similarity (with a threshold of 0.6). Finally, the Wikidata QID and geographical coordinates were retrieved.
Datasets
- Nordisk familjebok: The first (1876-1899) and second (1904-1926) editions of the Nordisk familjebok were used. These were digitized and OCRed by Project Runeberg.
- Wikidata: Used as a knowledge graph to link geographical entries and retrieve coordinate information.
- Swedish Wikipedia: Used to provide context for Wikidata entities.
A manually annotated dataset of 200 entries was created for training and evaluating the location classifier and for setting the matching threshold.
Key Insights and Ideas
- Evolution of Knowledge: Comparing editions of encyclopedias reveals how societal and global changes influence the selection and emphasis of information.
- Geographic Shifts: The study highlights a notable shift in the geographic focus of the Nordisk familjebok from the first to the second edition, with a decrease in European focus and an increase in attention towards North America, Africa, Asia, Australia, and northern Scandinavia.
- Influence of Global Events: The authors hypothesize that events like World War I and industrialization in regions like Norway and northern Sweden may have influenced these geographic shifts.
- Challenges in Historical Text Processing: The research underscores the difficulties in processing historical texts, including OCR errors, inconsistent formatting, and the need for language models adapted to older language variations.
- Wikidata as a Linking Resource: Wikidata proves to be a valuable resource for enriching historical text data with structured geographical information, enabling large-scale analysis.
Key Results
- The second edition of the Nordisk familjebok contains roughly double the number of entries compared to the first edition.
- The location classifier achieved F1 scores of 0.90 for the first edition and 0.92 for the second edition, demonstrating strong performance.
- The entry matching approach outperformed a baseline headword match, though the improvement was not always significant.
- Wikidata linking performance was moderate, with about half of the linked locations falling within 25 km of their correct coordinates.
Limitations and Future Work
- Limited Validation Sets: The validation sets used for evaluation were relatively small, which could affect the statistical significance of the results.
- Language Model Bias: Large language models trained on contemporary Swedish might exhibit bias or generate classification errors when applied to older texts.
- Entry Matching Errors: Similar semantic structures between entries can lead to erroneous matches.
- Wikidata Linking Accuracy: The accuracy of linking to Wikidata was moderate, with some errors in matching nearby entities.
Future work could involve fine-tuning models on older Swedish texts, exploring alternative matching algorithms (e.g., the Hungarian algorithm), and expanding the analysis to include more editions or other historical encyclopedias.
Summary updated