Towards a Narrative GIS

Towards a Narrative GIS
McIntosh, John, jmcintosh@ou.edu University of Oklahoma,
De Lozier, Grant, ghaxed@gmail.com University of Oklahoma,
Cantrell, Jacob , jcantrell@ou.edu University of Oklahoma,
Yuan, May, myuan@ou.edu University of Oklahoma,
Introduction
Research in narrative intelligence applies artificial intelligence approaches to study human ability to organize experience into narrative form (Mateas and Sengers 2003). Narratives are traditionally defined as “a series of temporally ordered clauses” (Labov 1972, p360-361). The time-centric approach leads to a lesser consideration of space in narrative construction and analysis. In contrast, we advocate a geospatial narrative in order to stress the importance of space and time in understanding the ordering and spatial interaction of events.

We define a geospatial narrative as a sequence of events in which space and time are equally important. Narratives are stories that constitutes sequential organizations of events (Franzosi 2010). Each event in a narrative relates sequential or consequential occurrence in space and time. The conventional Geographic Information Systems (GIS) center on information about spatial states of reality, and temporal information is handled as add-ons to spatial objects. Alternatively, we conceptualize a narrative GIS that emphasizes representing and ordering events in space and time as well as functional abilities to construct meaningful geospatial narratives. While an event is a complex, fuzzy term, we start with one basic linguistic element of narratives: action, as the primitive data construct to start building a narrative GIS. By relating action events across space and time, a narrative GIS aims to discover spatiotemporal correlates among actions and relate actions across scales.

Depending on the perspectives, there are many kinds of events, e.g. instantaneous events, discrete events, cyclic events, transitional events, and others. In contrast to TimeMaps (Farrimond et al. 2008), our vision of a Narrative GIS goes beyond spatiotemporal visualization to spatial analytics. By using action events as the primitive data constructs, a narrative GIS can support spatial queries of sequential and consequential actions. A Narrative GIS is therefore capable of revealing how time unfolds change and space unfolds interactions (Massey 2005).

We use two distinctive corpuses of histories in building narrative GIS databases and narrative analytics as a proof of concepts: Dyer's Compendium of the War of the Rebellion and the Richmond Daily Dispatch. Frederick H. Dyers, a Civil War veteran compiled the Compendium based on materials from the Official Records of the Union and Confederate Armies and other sources. The compendium lists organizations and movements of regiment cavalries mustered by State and Federal Governments for services in the Union Armies. Collaborating with the digital scholarship group at the University of Richmond, we have started with four files from Dyer's Compendium: the 45th Massachusetts Infantry, the 107th Pennsylvania Infantry,the1stCaliforniaInfantry,andthe1stNewYorkCavalry. The Richmond Daily Dispatch was one of the primary news media in the south during the Civil War. The newspaper was one of the most widely distributed newspapers of the south and included news from the entire east coast. The Richmond Daily Dispatch retained the reputation as politically unbiased was published throughout the Civil War.

Methodology
Our idea of a narrative GIS consists of (1) semantic elements (who did what), (2) temporal elements (when), and (3) spatial elements (where). A geospatial narrative object integrates the three elements and enables search for and analysis of spatial and temporal relationships among narrative objects. Input data for narrative GIS vary widely from structured to unstructured sources. In this study, both input data are texts, albeit in very different structures. Dyer's Compendium concisely lists regiment movements. Richmond Daily Dispatch consists of news articles. Spatial and temporal connections among units in these texts are considerably different. Nevertheless, the conceptual framework of a narrative GIS demands the identifications of semantic, temporal, and spatial elements from the texts to form narrative objects and relationships. As such, our workflow includes six key steps: (1) extract text analysis units; (2) identify action verbs; (3) identify time for words and text units; (4) identify locations for words and text units; (5) combine all identified elements into a GIS database; and (6) build spatial and temporal relationships among narrative objects. A schematic view of the workflow is presented in figure 1

Figure 1

Full Size Image

We begin with electronic versions of the historical documents. The texts are split into subsets such as newspaper articles or book chapters for processing. These subsets are typically written as a unit and need to be analyzed that way for successful interpretation. For each processing unit, we apply natural language processing to tokenize sentences and identify parts of speech (e.g. verbs and nouns). The parts of speech provide important clues to extract information.

The work presented here is centered on the location, time, and other characteristics of events. The part of speech tagging is used to identify verbs. We are most interested in “action” verbs and refine our list of potential candidates by removing stative and modal verbs. Location referencing begins with recognition of a standard grammatical structure to the way locations appear in text. In general, locations are proper nouns that do not directly follow a determiner (except for physical features).

Candidate words are matched to all their possible real-world geographic referents in the “Gazetteer Matching” process. A number of different gazetteers are utilized in the matching including the US Populated Places gazetteer and State hydrography datasets from the USGS, historical counties, states, and territories files from National Historical Geographic Information System, and the US Census Bureau’s historical 100 largest cities dataset (US BGN; US NHD; NHGIS 2008; Gibson 2008). These data are assembled in GIS and each location is identified with a historically and spatially appropriate hierarchy. The names of geographic locations are often highly ambiguous. For example “Georgetown” has over 70 possible locations among U.S. cities. Disambiguating a word to its true location is an important and difficult task. A substantial amount of work has already been done on location disambiguation under the heading of “Toponym Resolution” (Leidner2007;Leidneretal2003). Figure 2 illustrates the steps in the location referencing process.

Figure 2

Full Size Image

The temporal processing steps aim to extract dates, durations of events and the relative temporal ordering between events. Historical texts contain temporal information in a variety of formats. Most obvious are explicit dates that include information such as the day, month and year. In addition, these texts often include clues to derive dates and relative ordering of events. For example words such as “yesterday“ or “last week” allow the date to be derived based on a temporal relation to an anchor date (Han 2006). Similarly, relative temporal expressions allow explicit dates to be determined based on temporal relations to the current temporal focus (Han 2009).

Figure 3 outlines the steps in our approach.

Figure 3

Full Size Image

We begin by extracting anchor dates such as the date of publication for a newspaper article and explicit dates found in the text. We use temporal indicator words to refine the date of events and help establish temporal ordering. Explicit dates contained in the text are modified by deictic or temporal expressions. Semantic relationships between events are extracted based on semantic indicators. When all of the temporal information is relative and there are no explicit dates to give an explicit order, thirteen temporal relationships are used to find the temporal ordering (Allen 1983).
Results and Concluding Remarks
Thus far, our effort has been focused on extraction of events from the natural language text sources, anchoring the events to geographical locations and in time, and extracting information on the actors and objects involved in the events. Figure 4 illustrates results from Dyer's Compendium of the War of the Rebellion.

Figure 4

Full Size Image

This figure shows the activity of the 6th New York Regiment Calvary in Maryland and Virginia in September and October of 1862. The processing identified the events including the regiment’s movements and splitting off of a reconnaissance mission from Lovettsville to Smithville while the main regiment moved to Kearnysville. While this example from a single source, it illustrates the potential for the system to support more complex geospatial narratives with the addition of information from other sources.
Figure 5 shows a visualization of a Richmond Daily Dispatch article.

Figure 5

Full Size Image

The article describes a letter from a Union colonel to his family. It discusses the Union’s plan to move troops to Alexandria Virginia the next evening. The article illustrates that in addition to working with events that had already occurred the approach can also be used to help investigate the thoughts and motivation leading to events that had yet to occur.
These examples of preliminary results demonstrate the basic use of a Narrative GIS. As we continue building the event narrative database, additional functions will be built in for narrative analytics. For example, we are interested in deciphering the local, regional and national processes on emancipation and to identify scalar effects on military, political, and individual processes. One approach will be extracting reports on battles and run-away slaves and analyze spatial and temporal correlations among these events. When we extract events of different categories in space and time, a Narrative GIS will allow us to analyze spatial and temporal relationships among these kinds of events to draw insights into the integration of multiple perspectives and interpretations of geospatial narratives.

References:
Allen, J. Waltz, D. 1983 Maintaining Knowledge about Temporal Intervals Communications of the ACM 26 11 32-843

Bird, S. Klein, E. Loper, E. 2009 Natural Language Processing with Python --- Analyzing Text with the Natural Language Toolkit Cambridge, MA O'Reilly Media

Farrimond, B. Presland, S. Bonar-Law, J. Pogson, F. 2008 Making History Happen: Spatiotemporal Data Visualization for Historians Second UKSIM European Symposium on Computer Modeling and Simulation Liverpool, UK IEEE 424-429

Fellbaum, C. 1998 WordNet: An Electronic Lexical Database Cambridge, MA The MIT Press

Franzosi, R. 2010 Quantitative Narrative Analysis Los Angelas, CA SAGE Publications, Inc.

Gibson, C. 2008 Population Of The 100 Largest Cities And Other Urban Places In The United States: 1790 to 1990. U.S. Census Bureau, Population Division (link)

Han, B. Gates, D. Levin, L. 2006 From Language to Time: A Temporal Expression Anchorer Proc. 13th International Symposium on Temporal Representation and Reasoning

Han, B. 2009Reasoning about Temporal Scenario in Natural LanguageIn the Proceedings of AAAI Workshop on Spatial and Temporal Reasoning

Jackendoff, R. 1992 Languages of the Mind Cambridge, MAThe MIT Press

Labov, W. 1972 Language in the inner cityPhiladelphia University of Pennsylvania Press

Leidner, J, L. Sinclair, G. Webber, B 2003 Grounding spatial named entities for information extraction and question answering Proceedings of the HLT-NAACL 2003 Workshop on Analysis of Geographic References Edmonton, CAN

Leidner, J 2007 Toponym Resolution in Text: Annotation, Evaluation and Applications of Spatial Grounding of Place Names. Diss. University of Edinburgh, School of Informatics. Institute for Communicating and Collaborative Systems

Massey, D. 2005 For Space Thousand Oaks, CA Sage

Mateas, M. Sengers, P. 2003 Narrative Intelligence. Amsterdam/Philadelphia John Benjamins Publishing Company

Miller, G. Beckwith, R. Fellbaum, C. Gross, D. Miller, K. 1993 Introduction to WordNet: An Online Lexical Database (Revised) Princeton University

Pouliquen, B. Kimler, M. Steinberger, R. Ignat, C. Oellinger, T. 2006 Geocoding Multilingual Texts: Recognition, Disambiguation and Visualisation In Proceedings of The Fifth International Conference on Language Resources and Evaluation(LREC) (link)

National Historical Geographic Information System 2008 Minnesota Population Center: University of Minnesota. Minneapolis, MN (link)

U.S. Board on Geographic Names: Domestic and Antarctic Names – State and Topical Gazetteer Download Files United States Geological Survey (link)

U.S. Geological Survey: National Hydrography Dataset United States Geological Survey (link)