📄 A Gazetteer and Georeferencing for Historical English Documents

Summary

Most georeferencing tools rely on modern gazetteers, so historical spellings of English places (for example Bearla for Barlow) are often missed or matched to the wrong spot. The DEEP project addressed this by digitising every one of the 86 volumes in the EPNS's Survey of English Place-Names and converting the semi-structured OCR text into the DEEP Historical Gazetteer (DHG). An XML pipeline built on LT-XML2 and LT-TTT2 uses rules that depend mainly on layout, font and style rather than linguistic analysis. These rules mark up the hierarchy of places, historical forms, their attestations (dated sources), and etymologies. Dates become numeric intervals, and place-name elements are linked to KEPN. Coordinates come from KEPN and other sources; places without coordinates borrow them from the unit that contains them. The resulting MADS records link to Unlock, GeoNames and KEPN. The Edinburgh Geoparser was adapted to search the DHG, with optional county and date constraints. A Dorset Feet of Fines text shows both the benefits and the remaining gaps. The work gives historians a practical way to map pre-modern documents.

Key points

• The source is the Survey of English Place-Names. Its first volume (Buckinghamshire) appeared in 1925 and the newest (part of Shropshire) in 2012. Later volumes add minor names, field-names and street-names, and some counties are still missing or only partly covered.
• The pipeline converts Word files to OpenOffice XML and tokenises the text while keeping style attributes. It then adds structural and NER mark-up with hand-written rule sets, followed by date and place-name-element grounding. No POS tagging is used.
• Fuzzy dates get begin/end values:
  • a circa date gets a 20-year window;
  • "early 13th century" becomes 1200–1225;
  • a regnal date such as Hy 1 maps to 1100–1135.
• The DHG is stored as MADS records. These carry external links to the EPNS site, placenames.org.uk, KEPN, Unlock and GeoNames. The data is not yet RDF, but it could be converted to linked data without much effort.
• There was no formal NER evaluation, because no annotated test set exists. Instead, place-name experts reviewed the output in repeated quality-assurance cycles that fed back into the rules.

Interesting ideas

• I1. Typography and position on the page can do the job of part-of-speech tags: bold marks etymological elements and italics mark historical forms, so rules can read these signals directly.
• I2. Vague time expressions are turned into explicit numeric intervals by fixed conventions: a window around "circa" dates, fixed spans for parts of a century, and reign lengths for regnal dates.
• I3. A place with no coordinates inherits them from the nearest containing unit in the hierarchy, which gives an approximate but justified answer.
• I4. The paper separates a hard filter (consider only entries in one county) from a soft prior (raise the rank of entries inside a box or circle). It also shows how the hard filter can be too strict.
• I5. One canonical record holds many attested spelling variants as search terms, each tied to its dated source.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): The tokenizer could keep formatting and layout attributes on tokens, letting later rules use them in headings, lists, or textbook exercises.
• I2 (world): Event ordering could store every fuzzy time phrase as a begin/end interval with a recorded rule, so comparisons stay explainable.
• I3 (world): When an entity lacks a property, it could inherit it from its part-of parent, with the inheritance step visible in the explanation trace.
• I4 (world): Entity disambiguation could treat context as a weighting prior by default and switch to a strict filter only on purpose, logging which mode was used.
• I5 (world): Knowledge-graph entities could store aliases together with where and when each was attested, so older or variant names in fables and tales resolve to the same entity.