Year: 2024
Type: report
Abstract: Since its release in 2013, census microdata (I-CeM) for England and Wales covering censuses carried out between 1851 and 1921 have proved key to many areas of (British) economic history and cognate fields such as public health, transport, demography, history of entrepreneurship, migration, etc. 2 The heretofore unprecedented level of granularity of the data, which includes demographic, occupational, household structure and migration information for individuals, has unlocked many new research projects.However, I-CeM is limited by its lack of precise spatial linkage: indeed, only the parish was reliably recorded and standardised between 1851 and 1911.From these, time-consistent simili-parish units have been derived, and are currently the most precise units individuals can be linked to. 3 Therefore, the parish is currently the most granular geographic information provided.These were collated into consistent-parish geographies with an associated shapefile, but with an average size of 12.25 km², the level of detail they provide is not sufficient for granular studies Census enumerators did, however, record households' addresses, which were subsequently transcribed into I-CeM.The lack of standardisation in addresses, changes in the urban fabric and the toponymic evolution of the last 150 years complicate their use.This paper proposes a method to geocode individual census records of I-CeM using the address recorded for each household in England and Wales.Historical addresses are matched both to contemporary buildings and streets, and historical gazetteers.Locations are approximated for street-level matches, and the quality of the results is assessed manually in two different ways to demonstrate the use of this method and results for downstream tasks.The rest of this paper is structured as follows: in the first part, I review the literature pertaining to automated address matching and census geocoding.In the second part, I describe the proposed method to improve existing solutions.This section covers the sources used, the preprocessing workflow, the matching process, and the approximation method.In the third part, several data quality checks are introduced to assess the efficacy and reliability of the method.The final section presents the results and draws conclusions from these findings.A key axiom to this method is that dealing with historical data (which is by nature imperfect), it does not seek to geocode all historical addresses to their actual location.Rather, it seeks to strike a balance between precision (distance between actual address location and geocoded location) and recall 1 Thanks go to Gethin Rees from the British Library for providing access to Goad Maps, Bruce Hunt for his dataset of London Street Name changes, Dr January-McCann for the Welsh List of Historical Place Names and Nick Groome from the Ordnance Survey for AddressBase.Many CAMPOP members also contributed their thoughts and feedback in the preparation of this paper.Please do not cite this paper without explicit permission. 2 K. Schurer and E. Higgs, 'Integrated Census Microdata (I-CeM), 1851-1911' (UK Data Archive, 2023).3 Max Satchell and Ian Gregory, 'I-CeM Consistent Parish Geography', 2024.'b'ham' 'birmingham' 0.045 P Addresses having been standardised they can now be simplified/restricted.That simplification aims to extract from the input address the toponymic information that will be helpful for matching and discard any superfluous information.Hence, I use a Named Entity Recognition (NER) model to perform this task.17 NER models are able after training to build a degree of semantic understanding of addresses (that is to say, not merely quantifying the most common characteristics/components of addresses but being able to understand the role of tokens in the context of their address).This is a key expansion, as such a "cleaning" wouldn't be possible with a rule-based approach: the number of cases to be considered is simply too vast.Hence, only an AI approach seems able to perform that task: here, an off-the-shelf NER model (from the Spacy library) is used.This is not the only AI approach possible: one could train an LLM to reformulate addresses, à la Smits et al, but after experimentation, NERs were determined to be the most effective approach.Spacy 18 is trained on a varied corpus of addresses to cover the identified edge-cases as well as possible.17
Source: Apollo (University of Cambridge)
DOI: 10.17863/cam.111637
Author: Proffit, Guillaume
Language: English
Subfield: Geography, Planning and Development
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W6962809437
Cites:
Cited by:
Related to: