In this paper we present LocLinkVis (Locate-Link-Visualize); a system which supports exploratory information access to a document collection based on geo-referencing and visualization. It uses a gazetteer which contains representations of places ranging from countries to buildings, and that is used to recognize toponyms, disambiguate them into places, and to visualize the resulting spatial footprints.
One important direction for improving access to su ciently
large collections, such as those curated by archives and
libraries, is geo-referencing: linking information to
geographical location [
Existing Geographic Information Retrieval (GIR) systems
designed for archives, however, are limited by their
representation of locations. These systems represent place as point
(i.e. a single latitude-longitude pair) [
In this paper we present LocLinkVis; a system which supports exploratory information access to a document collection based on geo-referencing and visualization. It uses a gazetteer which contains representations of places ranging from countries to buildings, and which is used to recognize toponyms, disambiguate them into places, and to visualize the resulting spatial footprints. OpenStreetMap1 (OSM) is the only data source that is needed to ll the gazetteer. Our current LocLinkVis prototype covers the Netherlands and neighboring regions, but it can easily be populated by arbitrary subsets of OSM data (which has planet-wide coverage2). The current prototype runs comfortably on a commodity laptop, but would bene t from more powerful hardware if it were to provide global coverage. For the documentside of the system our proof-of-concept focuses on existing digital collections of parliamentary proceedings from Canada, the Netherlands, and the United Kingdom. The current LocLinkVis prototype is limited to a Dutch gazetteer and corpus, but will soon be expanded with the Canadian and UK corpora to match the coverage of the general PoliticalMashup3 search engine.
The systems design of LocLinkVis originates from the following requirements. The system should e ciently: 1. detect toponyms ranging from countries to buildings; 2. disambiguate which geographical features the toponyms refer to; 3. build an index of documents by the locations they mention; 4. use a map interface's viewport and zoom level as query parameters; 5. visualize results on a map, at a suitable level of abstraction.
Figure 1 shows the components of the resulting system, and which communication occurs between them. The numbers
OSM's coverage in 2014
in this schematic indicate which requirement is satis ed in (or between) which part(s) of the system.
The gazetteer builder takes OSM data as input, in the
form of e cient binary dumps as well as incremental XML
updates. Its primary purpose is to transform OSM's
normalized (i.e. with internal references) Node, Way, and
Relation instances to denormalized GeoJSON4 features. This
JSON-based format for geospatial data structures is
better suited towards retrieving and rendering spatial data,
whereas OSM's data schema favors editing it. Second, while
the OSM data is being imported, instances of the same
geographical type (e.g. city or road) that share a toponym and
geometrically intersect, are merged into a single gazetteer
feature. OSM contains data at ne level of granularity, e.g.
to indicate that a particular road section goes through a
tunnel. Our aim is to simplify (the side-e ects of) such details
when they are undesirable for spatial document search.
Our gazetteer, like its paper media predecessors, provides
an easily searchable list of toponyms, each accompanied
by facts about its location amongst other aspects.
Modern gazetteers, such as the Getty Thesaurus of Geographic
Names5, commonly represent the location of places as
coordinate pairs. With thanks to the OSM contributors, we
can now provide detailed geometries for the features in our
gazetteer. Besides the data that is copied and denormalized
from OSM, we add to each feature the identi ers of the OSM
instances it was derived from, and build an index of features
by all their known names (including historic toponyms).
Toponym detection entails identifying which tokens from
an input document refer to a place. LocLinkVis takes a
lexicon-based approach in which all known toponyms from
the gazetteer are encoded into a nite state automaton by
the Aho-Corasick algorithm [
NIL detection is the process by which candidate toponyms may be identi ed as not-referring to any location that is represented in the gazetteer. We opted for a corpus-independent approach, which starts by counting the frequency with which toponyms from the gazetteer are mentioned in a Wiktionary6 dump of the desired language. Wiktionary, for this purpose, serves as a compact, but high-coverage, cross section of lexical forms within a language. Possible toponyms that are mentioned frequently, it seems safe to assume, refer either to well-known places or are homographs of common words or other names. LocLinkVis uses Support Vector Machines that have been trained on a small number of training examples to make this distinction. Besides Wiktionary occurrence, the feature space that is used for classi cation includes variables obtained from SPARQL queries to DBpedia, as well as queries to the gazetteer.
Location type classi cation takes the lexical context of mentioned toponyms into account for disambiguation. It computes for any toponym (not limited to our gazetteer) the probabilities that it refers to an instance of the types Country, Province, (body of) Water, Municipality, Neighborhood, Road, and Building. It does so on the basis of a window of ve tokens preceding and following the toponym in question. Numerical tokens are replaced by a normalized form. The same is done for other toponyms within the token window in the training examples, and for unambiguous toponyms in unseen input. LocLinkVis o ers the choice between a Naive Bayes and a Maximum Entropy classi er, and uses between 25 and 85 training examples per class. Spatial analysis is useful to operationalize Tobler's \ rst law of geography," that \everything is related to everything
else, but near things are more related than distant things."
[
The geo-demo user interface (UI), available at http://geodemo.politicalmashup.nl, allows users to interactively validate our geo-referencing approach. Using this UI, arbitrary text can be submitted to the geospatial web server to view the intermediate results at each stage of the geo-referencing process. It shows which toponyms have been detected, which are thought not to be represented in the gazetteer, and the (detailed) output of location type classi cation and spatial analysis for all candidate features. A slippy map7 covers the screen, with input elds and georeferencing details being available from an expandable menu.
The search engine provides indexed access to one or more document collections. In our particular case these are corpora of national parliamentary proceedings in the rich PoliticalMashup XML format, but plain text documents would also be su cient. In order to enrich a document with geospatial annotations, the text it contains is submitted to the geospatial web server, which returns annotations of the form <Feature URI, Span, Confidence, Bounding Box>. A geospatial index ensures that the search engine can be queried with an arbitrary bounding box, and return documents from which any of the annotated locations intersect with the bounding box. If the documents also feature a temporal property (e.g. creation date or period of activity), it can enable further meaningful ltering of the results. LocLinkVis' search UI is described in the following section.
The main aim of our search UI is to provide an interactive geospatial overview of document results. In contrast to traditional GIR systems, which have focused on associating documents with a single location (often in the form of a point), this UI displays all places that are mentioned in the documents (see Figure 2), and with the same delity to which they are rendered on the background map. To assist in searching document collections with an interesting temporal aspect, we have added a timeline widget which further lters the result set.
The timeline widget, see Figure 3, shows the number of places that are mentioned in the result set and intersect with the map viewport. A categorical variable (i.e. facet ) can be used to break the total frequency down into result categories. The user may make a selection on this timeline, which acts as a temporal lter within the larger query. Additionally, besides manual adjustment of the selected time period, controls are provided which enable moving window playback which shows the dynamic of how often places were mentioned through time.
By these means the search UI provides a bidirectional binding between the spatial and temporal aspects for the exploration of a search result set. When a user interacts with the slippy map (i.e. by panning and zooming), the chart in the timeline widget is updated to re ect the new viewport. Vice versa, when a time period is selected or adjusted, the search result set is (re-) ltered and the highlighted locations on the map are updated accordingly. This enables the user to seamlessly move between shift and narrow search strategies, as well as between the spatial and temporal dimensions of the result space.
The search UI is under active (iterative) development at the time of writing. Hence, our prototype at http://geosearch. politicalmashup.nl will be changed periodically and may not always be available. We are, for instance, currently experimenting with the placement of result snippets in the interface. When a user clicks a highlighted location, document results from a query could either be displayed on or near the clicked location, or be shown within a xed screen area.
In this paper a novel GIR-based system was presented that supports exploratory information access to a document collection by means of geo-referencing and interactive visualization. LocLinkVis distinguishes itself from existing systems that have similar goals by working with a broader range of location types, and by representing a location's geometry with the same delity as to which it is rendered on OpenStreetMap. It also features a UI which allows the user to seamlessly move between the shift and narrow search strategies, as well as between the spatial and temporal dimensions of the result space.
This work is situated in the context of a project which aims to improve (exploratory) information access to parliamentary proceedings and related cultural heritage collections. We are working towards a living lab with high-quality, reliable prototype systems and ubiquitous interaction logging. This approach allows us to use a combination of (remote) laboratory experiments and analysis of in-the-wild usage to investigate how we can best assist historians, political scientists, journalists, and media scholars in their complex search and sensemaking activities.
Our current e ort to provide accurate results for
historical collections mainly consists of collecting historical place
names from OSM, e.g. to compensate for spelling variants
in older texts. The gazetteer, however, does not yet contain
information about during which time period a place was
known by a particular name, and only contains the most
recent geometric representation that is available. This can
lead to a distorted picture e.g. when borders have changed
much, or when a monument has not always been in its
present-day location. At least for Dutch municipalities the
information needed to display the historical situation
accurately is available [
Acknowledgements This research was supported by the Netherlands Organization for Scienti c Research (ExPoSe project, NWO CI # 314.99.108; DiLiPaD project, NWO Digging into Data # 600.006.014). We would like to thank the anonymous reviewers of this paper for their useful remarks.