Originally posted at blogs.it.ox.ac.uk on July 29, 2013 by Martin Wynne. IT Services at the University of Oxford has decided to delete a large number of historical blogs, and this is one of a number of posts related to the Oxford Text Archive which are being re-published here, after being laboriously retrieved from the archive provided by the Wayback Machine.
I note with interest that there are still attempts to kick-start the effective use of geo-spatial technologies and methods in the study of literature. Some years ago, I attended a very interesting workshop at the University of Nottingham on ‘Places in Literature’. The event brought together researchers from various fields to investigate the feasibility of an interdisciplinary project to use geospatial technology to enhance literary research. The event featured presentations, discussion and some hands-on encoding and analysis, and was held in the Centre for Geospatial Science at the University of Nottingham (in a brand new building on the site of the old Raleigh bike factory). Most of the participants were from Nottingham, with a couple of us from Oxford, and two people from Glasgow. The Nottingham participants included specialists in geospatial scientice, English literature, English language, place names, cultural geography and history, and computational linguistics.
It was an excellent opportunity for us to push forward our understanding of the challenges and barriers to developing useful applications which could be used in research. The findings included the following:
- there is an unbounded number of ways in which narrative texts refer to places; probably most of the time the reference to place in a text is not a reference to a stage in a journey, or a description of the location of an event (e.g. “You’re not in London now!”);
- references are often non-explicit – Heart of Darkness by Conrad does not say where the action is taking place, although the reader is likely to infer London, Brussels and the Congo;
- places in fictional worlds may or may not relate in a reliable way to geography in the real world;
- all texts are historical; mapping places in texts is never the same as mapping in contemporary real-world applications;
- place names are subject to variation in languages, spelling, change over time, movement of borders, etc.;
- variation in granularity and fuzziness are inherent in place name references in texts;
- existing geo-coders appear not to be very good at recognizing place names in literary texts (the poverty of named entity recognition);
- existing geo-coders appear not to be very good at reliably assigning locations to place names (due to the ambiguity of place name references – Lancaster in Lancashire or Lancaster in Pennsylvania?).
It was my conclusion that with the current state of the art, it was not easy, and probably not possible, for literary scholars to use the state of the art geoparsing and mapping to0ls to improve their research or to ask new research questions.
I concluded that future work to improve this situation might involve:
- applying state of the art named entity recognition software to texts in order to more reliably identify place name references in texts;
- investigating heuristics to improve the geocoding (or the ranking of possible hits) of place names;
- producing tools that allow users to investigate, correct and examine the outputs at each stage, as these outputs are likely to require human intervention to improve accuracy to an acceptable level, and because these outputs are likely to be interesting to various types of research;
- investigating means of applying the techniques to large text collections, or combining them with web searches; developing a focussed, in-depth research topic to make use of these tools in respect of a specific set of texts;
- combining geocoding tools with historical place name gazetteers and maps;
- – allowing geospatial information to be combined with linguistic information in the text, e.g. locations linked to concordances and collocations of textual references to the location.
- embedding the relevant tools in language resources infrastructure such as CLARIN, so that the geographical tools can be combined with tailored NLP tools.
Current examplars which I have seen still use very specific text types based on travel itineraries, such as travel writing and guide books. I still haven’t seen an example which has effectively supported research in narrative fiction with automated or semi-automated geographical analysis, but I would be happy to be proved wrong!