Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=tjde20
International Journal of Digital Earth
ISSN: 1753-8947 (Print) 1753-8955 (Online) Journal homepage: https://www.tandfonline.com/loi/tjde20
Geo-analytical question-answering with GIS
Simon Scheider, Enkhbold Nyamsuren, Han Kruiger & Haiqi Xu
To cite this article: Simon Scheider, Enkhbold Nyamsuren, Han Kruiger & Haiqi Xu (2020):
Geo-analytical question-answering with GIS, International Journal of Digital Earth, DOI:
10.1080/17538947.2020.1738568
To link to this article: https://doi.org/10.1080/17538947.2020.1738568
© 2020 The Author(s). Published by Informa
UK Limited, trading as Taylor & Francis
Group
Published online: 12 Mar 2020.
Submit your article to this journal
View related articles
View Crossmark data
OUTLOOK PAPER
Geo-analytical question-answering with GIS
Simon Scheider , Enkhbold Nyamsuren, Han Kruiger and Haiqi Xu
Department of Human Geography and Spatial Planning, University Utrecht, Utrecht, The Netherlands
ABSTRACT
Question Answering (QA), the process of computing valid answers to
questions formulated in natural language, has recently gained attention
in both industry and academia. Translating this idea to the realm of
geographic information systems (GIS) may open new opportunities for
data scientists. In theory, analysts may simply ask spatial questions to
exploit diverse geographic information resources, without a need to
know how GIS tools and geodata sets interoperate. In this outlook
article, we investigate the scientific challenges of geo-analytical question
answering, introducing the problems of unknown answers and indirect
QA. Furthermore, we argue why core concepts of spatial information play
an important role in addressing this challenge, enabling us to describe
analytic potentials, and to compose spatial questions and workflows for
generating answers.
ARTICLE HISTORY
Received 26 July 2019
Accepted 1 March 2020
KEYWORDS
Geographic question
answering; GIS; core concepts
of spatial information; geo-
analytics; geocomputation
1. Motivation
A large variety of analytical resources available on the Web and elsewhere offers genuine new oppor-
tunities for empirical scientists (Kitchin 2013). Take the example of a health scientist (Richardson
et al. 2013). Millions of wearable sensors, exabytes of personal health records, billions of nodes on
Open Street Map (OSM) and countless geolocated social media posts make it very probable that
the spatial data needed for, say, finding out how environmental factors influence a person’s health,
stand ready for sophisticated ways of modeling.
Yet, the question of the health scientist can probably not directly be answered by this data. Data
may not directly fit the purpose and may require further processing or analysis to generate a valid
answer through geo-analytical tools. Geo-analytical tools, on the other hand, are difficult to employ
when distributed across countless software programs. The 40 most well-known GIS software
packages
1
together contain thousands of different tools (Ballatore, Scheider, and Lemmens 2018),
and thousands of modules are added by online repositories such as PyPy
2
for Python or CRAN
3
for R. For the health scientist, it may simply take too much time to first learn GIS to find out whether
it might answer his or her question using a particular data source.
What if the health scientist could simply ask a spatial question to find the right data and analysis
functionality in an instant? This would tremendously reduce the effort needed, and it would also be a
step towards exploiting geo-computational resources for analysts who are not GIS experts. Questions
capture what analysts want to know and let them share and discuss their results independently from
particular software or data formats (Vahedi, Kuhn, and Ballatore 2016). Yet, the kind of question-
based analysis sketched here is unfortunately not possible today. Although Question-Answering
© 2020 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License (http://
creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the
original work is properly cited, and is not altered, transformed, or built upon in any way.
CONTACT S. Scheider sscheider@uu.nl Department of Human Geography and Spatial Planning, University Utrecht, 3584 CB
Utrecht, The Netherlands
INTERNATIONAL JOURNAL OF DIGITAL EARTH
https://doi.org/10.1080/17538947.2020.1738568
(QA) systems have matured in recent times in both industry and academia (see Section 2.1), current
approaches still focus on machine translations of questions into queries on factoid knowledge bases.
To answer the question of our health scientist, however, it is not sufficient to query a database of
known facts.
Let us illustrate this argument. We argue that the challenge of building a GIS that can directly
answer questions (a question-based GIS) should not be conceived as a query task (as in ordinary
QA), but rather as a particular transformation task. The question ‘How far is it from Paris to Amster-
dam?’(Gao and Goodchild 2013) might be directly answered from a knowledge base, but it might
also be translated into the concepts ‘Spatial Distance’,
4
‘Paris’and ‘Amsterdam’, which in turn
can be easily mapped, for instance, to Google’s routing function and corresponding nodes on Google
Places. Consider, in contrast, the following question which is relevant in the context of Health
Geography (Richardson et al. 2013):
How much is Tom exposed to green space while running through Amsterdam?
First, it is unknown what the answer to this question is, as it depends on an analytic parameter
(Tom’s particular run). Furthermore, even if we grant that Tom’s run is stored in a knowledge base, it
is not obvious how an answer can be generated. For example, which combination of data sets and
GIS tools as listed in Figure 1would yield a valid answer? We call the latter kind of questions analytic
in the following.
It is important to realize that GIS is fundamentally about designing workflows as answers to ana-
lytic questions for which answers are yet unknown. The work of GIS analysts, therefore, cannot be
reduced to querying over known datasets. As an example, take the administrative Geography avail-
able on Wikidata
5
: Though Wikidata can easily answer the question in which state the US city San
Diego is located, such kind of knowledge is of little interest for a Geographer.
6
Similar to statistics,
GIS is a collection of methods for analysts, involving creativity in figuring out how data can be trans-
formed to obtain an answer to a novel kind of question. What is needed is, therefore, a way to rep-
resent the analytic potential of a dataset for answering questions, based on a good theory about
spatial questions as well as the possibilities of operational transformations provided by GIS.
We call this problem geo-analytical QA, which is part of a more general endeavour of indirect QA
(Scheider, Ostermann, and Adams 2017). ‘Indirect’means here that answers cannot be directly
filtered out by queries, but need to involve transformations first. While indirect QA is relevant to
all analytical sciences (including statistics and data science), not only to geographic information
science, the inherent transformation possibilities and semantic constraints depend on the kind of
information being transformed. This means that solutions will be specific to geographic information,
and geo-analytical QA is special because transformations necessarily involve spatial concepts. In this
Figure 1. Which combination of tools and data would yield an answer to the question of the environmental health scientist?
Example tools were taken from ArcGIS (http://desktop.arcgis.com/en/arcmap/). Data from the Amsterdam data portal (https://
maps.amsterdam.nl).
2S. SCHEIDER ET AL.
outlook article, we make a case for geo-analytical QA as an autonomous research endeavour in the
context of the Digital Earth, to which geospatial semantics (Janowicz et al. 2012), GIS workflow com-
position (Hofer et al. 2017) and spatial language processing (Hamzei et al. 2019) may contribute on
an equal footing. We closely investigate this challenge and clarify the role that core concepts of spatial
information could play in formulating and answering such types of questions. Since this is an outlook
article, our discussion of solutions needs to remain preliminary.
2. The challenge: building a question-based GIS
In this section, we compare the task with current approaches in order to carve out its main chal-
lenges. This prepares the ground for arguing why semantic concepts in general are needed, and
core concepts of spatial information in particular.
2.1. State of the art
Question answering has a long tradition in Artificial Intelligence (AI) and computational linguis-
tics, going back to the wave of expert systems research of the twentieth century (Simmons 1970).
Though these systems had limited impact, they touched upon many issues still relevant today,
including linguistic grammars and semantic frames for parsing questions (Ofoghi, Yearwood,
and Ma 2008). With the advent of the Web, a revival of QA systems occurred due to the avail-
ability of large query and answer sets (Lin 2002), with a potential to improve general information
retrieval (IR) systems (Laurent, Séguéla, and Nègre 2006). Today, we distinguish knowledge-based
Question Answering (KB QA), which derives answers from structured data (Diefenbach et al.
2018), and document-based QA, which finds answers within unstructured text (Kolomiyets and
Moens 2011).
A recent review of the two categories of QA systems can be found in Shah et al. (2019). Compared
to document-based QA systems that are more suitable for answering simple factoid questions, KB
QA systems can answer questions that require reasoning over multiple factoids. While document-
based QA systems can be rather easily generalized to different domains, KB QA systems require con-
siderable manual effort in creating KBs and, thus, have limited cross-domain applicability. However,
document-based QA systems require the availability of large corpora from which answers can be
extracted. These disadvantages may indicate which QA system is more suitable in a particular
case (Gupta and Gupta 2012). Recent efforts aim to create hybrid QA systems that combine elements
of both (Mitra et al. 2019; Sawant et al. 2019). In case of GIS, such hybrid systems may be most suit-
able. Since GIS needs to answer complex analytical questions, KBs are really required, yet the domain
also lacks a corpus of documents that can be leveraged.
The linked data cloud
7
and RDF
8
have been recently proposed for KB question-answering over
data cubes (Höffner, Lehmann, and Usbeck 2016), allowing answers to be retrieved over many
dimensions and resolution levels. The Semantic Web can be seen as a core technique for KB QA
because its particular strength lies in reasoning over taxonomic concepts (Höffner et al. 2017),
and large Web data bases such as DBpedia,
9
Yago
10
or WordNet
11
can be used for answer set gen-
eration (Bao et al. 2014). Main computational steps involve (1) the analysis of questions into phrases,
(2) the mapping of phrases (including named entities) to the KB, (3) entity disambiguation, and the
(4) construction and (5) firing of queries over the KB (Diefenbach et al. 2018).
Although introductory textbooks in GIS are focused around answering spatial questions (Hey-
wood, Cornelius, and Carver 2011, ch.1), geographic question answering as such has only been sub-
ject of a limited number of research endeavors. In the past, researchers have proposed conversational
and natural language interfaces to GIS (Cai et al. 2005). More recently, researchers have addressed
how spatial questions can be translated to spatial query languages (Chen 2014; Pulla et al. 2013).
Others also addressed how QA queries can be spatially and semantically expanded to arrive at a
more successful answer rate (Mai et al. 2020). Gao and Goodchild (2013) matched geo-analytical
INTERNATIONAL JOURNAL OF DIGITAL EARTH 3
tools to questions based on keywords. More recently, Zhang et al. (2018) proposed a knowledge-
based QA system for answering geographic questions that can be found in standardized tests of Chi-
nese high school students. Overall, we need to ascertain a lack of effort in creating a system for
addressing geo-analytical questions, which are so central to GIS.
Correspondingly, there is also a lack of studies on investigating these types of questions in par-
ticular. Few studies we have identified agree on informal categories (Heywood, Cornelius, and Carver
2011; Kraak and Ormeling 2013;O’Looney 2000; Allen 2016; Mitchell 2012), e.g. questions about
relationships (e.g. What is the relationship between the local microclimate and locations of fac-
tories?) or questions about implications (e.g. If we build a new theme park here, what will be the
effect on trafficflows?) (Heywood, Cornelius, and Carver 2011; Kraak and Ormeling 2013;O’Looney
2000; Mitchell 2012). While traditional QA systems also try to address questions about implication
and relationship (Kolomiyets and Moens 2011; Wang 2006), it is assumed that answers can be
directly retrieved from documents or logically inferred from knowledge bases. GIS commonly relies
on analytic operations on raw data to answer these questions. Kuhn and Ballatore (2015) and
Vahedi, Kuhn, and Ballatore (2016) revealed that the core concepts of spatial information, as
suggested by Kuhn (2012), may play a central role in question formulation, as well as in describing
these analytical operations.
2.2. Challenges in asking and answering geo-analytical questions
What is it that makes geo-analytical QA challenging? And why does the problem require more than
what current QA technology has to offer? In a nutshell, since answers are basically unknown, they
need to be given as workflows, which requires creativity in finding an answer. This, in essence, makes
the problem a non-trivial learning task, since it requires going beyond matching of question and
answer sets.
Creativity in answering. First, there is not only large degrees of freedom in how a spatial ques-
tion can be asked, but also an equally large variety in how a given spatial question can be answered.
The former problem is due to the flexibility of how a concept, such as distance, can be expressed in
natural language (‘How far?’,‘How close?’,‘what is the nearest …?’), leading to questions whose
formulation is too far offthe answer formulation in order to successfully match. Query expansion
by increasing the tolerance of spatial query expressions (Mai et al. 2020) and also machine learning
approaches have been used to circumvent this (Diefenbach et al. 2018; Bao et al. 2014). More
importantly, however, there is usually also a certain semantic ambiguity in how a given expression
in a question might be interpreted in terms of geospatial concepts. For example, the term ‘green
space’in the introductory question might equally well refer to an object (a park), a collection of
objects (trees), or patches of a certain landcover category. This ambiguity is inherent to GIS
and needs to be dealt with in geo-analytical QA. And a similar kind of ambiguity appears in answer
formulation, too. There is no such thing as a most definite or probable answer contained in a geo-
analytical KB. As a matter of fact, using a GIS, the same question can be answered in ways that are
very different yet equally valid, based on combining different tools with different sorts of data
sources (Chrisman 2002). For example, in order to address the question in Section 1,allofthe
tools in Figure 1may render equally valid answers, using data sources ranging from Open Street
Map (OSM) to official land-use statistics. The task is rather to capture the large variety of valid
approaches to answering a given question, not of reducing an answer set to a most probable
answer.
Question complexity.O’Looney (2000) suggests types of questions addressed by GIS and ranks
them from simple to complex in the following order: location, condition, routing, pattern modeling,
trend modeling, and what-if modeling. A ‘location’question is a simple where question. An example
of a condition question is ‘What is the condition of the water treatment plant at 270 feet?’.An
example of a pattern modeling question is ‘What is the pattern of public spending in areas where
the majority of residents are African American?’. While this particular typification of questions
4S. SCHEIDER ET AL.
may be contested, the acknowledgment of question complexity in GIS is of importance. Moreover,
question types are not mutually exclusive. A condition question can also be a location question, and a
pattern modeling question can also be a condition question. Therefore, answering a question in GIS
requires its decomposition into simpler ones that need to be answered separately.
Indirect answers. The reason why geo-analytical questions are challenging in the first place is that
their answers cannot be looked up. For those questions that QA systems usually handle, such as
‘Who is the director of Forrest Gump?’(Bao et al. 2014), the answer is known. Our task is rather
to match questions to answers which are unavailable, yet may be generated from what is known
using analytic functions. This latter task requires capturing the analytic potential of tools and data
to answer a question.Ineffect, this means to translate a spatial question into a query over a trans-
formation: The query should match ‘potential’datasets generated by some workflow (Figure 2), a
novel computational challenge which was called ‘indirect QA’in Scheider, Ostermann, and
Adams (2017).
Non-trivial learning. All this makes geo-analytical QA a non-trivial learning task. On the one
hand, the creativity and complexity involved, as well as the predominance of indirect answers,
make it very hard to obtain representative training samples for questions and answers. On the
other hand, training a QA system to give the most probable answer fails to capture precisely that
a very different, maybe improbable, yet valid answer might be given.
2.3. The geo-analytical QA problem
One might argue that every indirect QA problem can be turned into an ordinary QA problem,
simply by computing an answer and adding the answer to a database. However, such an approach
would hardly solve the problem: it is simply impossible to precompute answers to every possible
analytical question. It is for this reason that analytical QA really requires a new paradigm based
on possible computational transformations, the latter being specific for GIS. Which steps would
need to be taken for geo-analytical QA? As illustrated by Figure 2, there are three subproblems:
(1) Assessing the analytic potential of a geodata set. We cannot find out about the analytic potential
of a dataset for a question by simply querying over it. Instead we need to assess whether a trans-
formation of data exists which would answer a question. Similarly, we need to assess whether a
certain GIS tool can be used in this transformation.
(2) Assessing the possibility of a transformation requires synthesis of GIS workflows. This generates
possible answers. Workflows may be diverse but need to have a high quality, i.e. they need to be
Figure 2. The problem of indirect QA. Blue boxes denote information which needs to be generated by an indirect QA system.
INTERNATIONAL JOURNAL OF DIGITAL EARTH 5
valid from a methodological point of view. This captures the creative aspect of geo-analytical QA
in answering a question.
(3) Translating geo-analytical questions into queries over such workflows. To pick valid workflows as
answers to a given question, we need to decompose the question into underlying concepts which
match the outcome of some GIS-based transformation.
Restricting this to the case of geo-analytical questions is important: Our point is that all three pro-
blems are only solvable based on exploiting the semantic concepts that are contained in the questions
and which are underlying GIS. This means that solutions need to be searched within the radius of
geospatial semantics.
3. The role of core concepts
In the following, we suggest that core concepts of spatial information are indispensable not only for
understanding how geo-analytical questions and answers are composed but also for knowing
whether geodata is fit for answering. The reason is that they provide semantic constraints for posing
spatial questions, as well as operational constraints for describing analytic potentials and for finding
answers by constructing workflows.
3.1. Core concepts of spatial information
Core concepts of spatial information were proposed by Kuhn (2012), Kuhn and Ballatore (2015)as
generic interfaces to GIS in the sense of conceptual ‘lenses’through which the environment can be
studied. Though they have been used in the sense of abstract data types (ADT),
12
they are considered
results of human cognition and interpretation, and thus go beyond data types or formats. Core con-
cepts in this latter sense were not invented by Kuhn, but are known and used implicitly by everybody
who understands the essence of a GIS. Our task is to lift these concepts to an explicit level in order to
make use of the information contained in them. Though a formal specification of core concepts is
still ongoing work, and though the set of concepts have changed in the past to some extent,
13
there is a rather stable consensus of the following content concepts
14
on which we focus here:
.Fields are understood as particular kinds of functions (Galton 2004; Câmara, Freitas, and Casa-
nova 1995; Scheider et al. 2016) whose domain are locations which allow for metric distance
measurement, and whose range may be any kind of quality. Prime examples are temperature
fields. A field function can change in time (Scheider et al. 2016). As quality values are separated
by spatial distance, one can study change of a field as a function of spatial distance. Fields also
offer the possibility of determining quality values at arbitrary locations in their domain. Missing
quality values can therefore be estimated by interpolation. This concept is closely related to the
notion of a field in physics (Einstein 1934).
.Objects are understood as entities which have a spatial region and diverse qualities that can change
in time. This corresponds to the idea of ‘endurants’in philosophy (Galton 2004), which can
change their location and quality while remaining their identity. Objects are distinct from
other concepts in the sense that they have identity (usually a name) and that they are fully loca-
lized in each moment of their existence, even if this location may be fuzzy. In this way objects give
rise to trajectories, which are functions from time to location, and time series, which are functions
from time to quality (Scheider et al. 2016). Geographic places, such as shopping malls or parking
lots, are considered particular kinds of objects. Though they are not considered locations, they are
localizable themselves.
.Events are understood as entities that, besides having identity and having qualities like objects, are
not fully localized in each moment but happen during some time interval. Events thus correspond
to particular kinds of ‘occurrents’or ‘perdurants’in Philosophy (Galton 2004). Since they have a
6S. SCHEIDER ET AL.
start and an end, they allow us to determine duration, and they might have objects, fields or spatial
networks as participants. In GIS, we usually assume in addition that events are localizable similar
to objects. Prime examples are earth quakes, having a time, a location as well as a magnitude.
.Networks are understood as quantified relations between objects. In this way, networks are able to
measure a relationship between these objects. Similar to graphs, this relationship is quantified, e.g.
in terms of an amount of flow or a distance. Networks in this sense are e.g. commuter flow
matrices or distance edges in a road network.
Since core concepts have the mentioned properties, certain kinds of operations are naturally
applied to them. For example, since fields are total functions on a metric space, their quality can
be probed at every location and at every distance within that space, while object qualities cannot.
Objects, on the other hand, can be counted, have spatial parts (mereology) and neighbors (topology),
and furthermore give rise to sizes and closeness. Events can be ordered in time. Finally, since net-
works are relations between objects, they can always be projected to object qualities by fixing
some source or destination, giving rise e.g. to catchments and service areas. Similar to levels of
measurement and other semantic types (Chrisman 2002; Scheider and Huisjes 2019; Scheider and
Tomko 2016), core concepts work as constraints to spatial analysis (Sinton 1978).
In contrast to ADTs, however, core concepts cannot be directly operated on. They rather work as
spatial ‘lenses’through which analysts see the geographic world when interpreting geodata (Vahedi,
Kuhn, and Ballatore 2016). A given dataset, therefore, might be viewed with different lenses, making
it possible to switch semantic perspectives and rendering interpretations inherently ambiguous. For
example, the part of a road between two intersections can be regarded both as an object and a net-
work element, and weather phenomena such as a storm can usually be viewed from a field perspec-
tive, an event perspective, as well as an object perspective. This ambiguity is part of GIS practice and
thus needs to be taken into account. In Scheider et al. n.d.), we proposed core concept data types as a
way to capture the different ways of how geodata types can represent core concepts. The ambiguity
can be handled by allowing a dataset to be annotated by more than one core concept (see example in
Section 3.2).
Note that our argument is not to convince people to use core concepts as a direct interface to a QA
system. The idea is rather to use them as an internal representation of (possibly multiple) shared
understandings of GIS question-answering resources. In the following, we argue for core concepts
as part of (1) a type system for adding analytic potentials to data and tools, (2) an internal grammar
for formulating questions and interpreting them as queries, and as (3) a way to compose answer
workflows. These three aspects may provide the glue for solving geo-analytical QA.
3.2. The role of core concepts in describing analytic potentials of geodata sets
Though geodata types are commonly used in GIS, they are insufficient to assess the QA potential of
geodata. The fact that raster and vector types do not capture the underlying concepts relevant for
analysis is subject already of introductory GIS books. Figure 5(a), taken from Heywood, Cornelius,
and Carver (2011), illustrates how diverse examples of spatial concepts (hotels, ski lifts, forst areas,
roads and elevation surfaces) can be represented by both raster or vector data, and thus are orthog-
onal to these concepts. Furthermore, as Figure 3demonstrates, the same geographic entities, such as
points, can refer to different concepts such as objects, events, and field measurements. The arbitrari-
ness of representing such concepts with geo data types indicates that core concepts add an indepen-
dent but relevant piece of information to a given data type. Take the example of a ‘forest’class in a
landcover data set (Figure 5(b)). The fact that this data set is a polygon vector tessellation does not
tell us much on how we can use it in analysis. In particular, it does not tell us that every polygon
really represents a homogeneous spot within a spatial field of landcover values, and that therefore
every location within a given polygon has the same landcover value. This way of representing a
field was called coverage in Scheider et al. (2016), an example is the left map in Figure 4.Adifferent
INTERNATIONAL JOURNAL OF DIGITAL EARTH 7
example of a vector based field representation is a contour map, as shown in the middle of Figure 4.
However, a tessellated polygon data set may also be a representation of a tiled collection of spatial
objects, such as in municipal statistics. For example, the right map in Figure 4shows zip code
regions. This way of representing objects was called lattice in Scheider et al. (2016).
15
Since munici-
palities are conceived as objects and not fields, their measured qualities, such as average elevation in a
municipality, are valid only for the entire object, and not for any of its parts.
It is this distinction which largely influences how the dataset can be meaningfully analysed (Schei-
der and Tomko 2016; Scheider et al. 2016; Scheider, Ballatore, and Lemmens 2019). For example,
vector overlay can be applied only to coverages, because it involves copying the measured quality
for arbitrary parts within a given polygon. This is true also for raster overlay, where cell values
are simply passed down to intersected parts of cells. Intersecting lattices, in contrast, requires
areal interpolation instead (De Smith, Goodchild, and Longley 2007). In a similar way, point interp-
olation is applicable if and only if a point data set represents a field, and not a collection of objects or
events (cf. Figure 3) (Scheider et al. 2016). This is opposed to density and distance computations,
which require identifiable, countable and bounded spatial entities.
As these examples illustrate, there are plenty of reasons to assume that core concepts of spatial
information, together with levels of measurement (Chrisman 2002) and related semantic distinctions
(Scheider and Huisjes 2019), are indispensable in order to assess how a given data source might be
transformed into meaningful answers. Our task in the future is therefore to (1) settle on a definite set
of semantic types which capture the diverse ways how core concepts are represented within a given
geodata type and (2) to find ways to scale up the semantic annotations across various geodata
Figure 3. Point maps representing buildings as objects (left), war-time events (middle), and temperature field measurements
(right). Data from the Amsterdam data portal (https://maps.amsterdam.nl).
Figure 4. From left to right, polygon maps representing fields as coverages (land use types), contours (road traffic noise contour),
and objects as lattices (zip code regions). Data from the Amsterdam data portal (https://maps.amsterdam.nl).
8S. SCHEIDER ET AL.
sources. Regarding the first problem, we have recently made a suggestion for a corresponding
OWL
16
-based ontology pattern in Scheider et al. (n.d.). For the second task, machine learning (Schei-
der and Huisjes 2019) or crowd sourcing might be used (Khan et al. 2016).
3.3. The role of core concepts in posing geo-analytical questions
Core concepts also play an important role in formulating and interpreting questions. And similar to
the data annotation task, there is often ambiguity in how to interpret a given question in terms of
core concepts. In the following, we go through a range of example questions, highlighting different
possible interpretations.
A question about spatial distance or spatial density usually implies spatial boundaries (to deter-
mine distances) and countability (to determine densities). Both are supplied by spatial objects. For
example, such objects are parks and trees as in the following example:
How
(Field)
far is the next
(Object)
park in
(Object)
Amsterdam?
How
(Field)
densely are
(Object)
trees located in
(Object)
Amsterdam?
Distance to something and density of something, in turn, are interpreted here as spatial fields.
While spatial distance itself is a quantified relation between locations, ‘distance to something’pro-
jects this relation to a spatial field by fixing its range. Such a field is one way how we measure
exposure in practice. Note that in all these cases, Amsterdam plays the role of another object
whose region delimits the extent.
However, note that the same distance term might also be interpreted in terms of a spatial network.
In this case, it is measured between pairs of objects, e.g. between any address and the next park
object, and then projected to this address object:
Figure 5. Why core concepts add essential information to geodata types. (a) Concepts, such as fields, objects and networks, can
always be represented by both geodata types, Vector or Raster. Source: Heywood, Cornelius, and Carver (2011). (b) The difference
between representing a field (coverage) or an object (lattice) in terms of a vector tessellation. Land cover is an example for a cover-
age, and average elevation (or any other statistical aggregation) is an example for a lattice. The attribute located at the cross is
determinable for the coverage, but not for the lattice (Scheider et al. 2016).
INTERNATIONAL JOURNAL OF DIGITAL EARTH 9
How
(Network)(Object)
far is the next
(Object)
park in
(Object)
Amsterdam?
Questions may also directly query networks between two objects. The following question implies
a network measuring runner flows on pairs of places (the latter understood as kinds of objects):
(Network)
How many runners run from
(Object)
University campus
(Network)
to
(Object)
downtown?
Fields allow us to probe values at an arbitrary location within their extents (‘How far away is the
next park from here?’). If we interpret exposure to parks in terms of a distance field, then such
locations may be supplied, e.g. by some spatial event (such as Tom’s run), allowing us to ask:
(Field)
How much exposed to
(Object)
parks is
(Event)
Tom’s run?
Tom’s run may likewise be interpreted as a series of object locations used to do the probing:
(Field)
How much exposed to
(Object)
parks is
(Object)
Tom’s run?
Furthermore, fields can be summarized into objects. In the following example, the term ‘green’is
interpreted as some homogeneous patch inside a field of land-use. This sub-field is summarized into
the object of the Amsterdam municipality, where it constitutes a new object quality:
(Object)
How much is Amsterdam covered with
(Field)
green?
Finally, in contrast to fields, objects can be spatially queried and counted using other objects:
How many
(Object)
trees exist in
(Object)
Amsterdam?
What these examples illustrate is not only that core concepts help decompose questions into
(sub-) questions (e.g. the question about Tom’s exposure can be decomposed into questions
about a distance or density field), but also that parts of speech and goal concepts within questions
can be interpreted as core concept transformations (e.g. the transformation of a land-use field
(‘green’) into an object quality). Furthermore, even though such interpretation often allows for
more than one decomposition, the resulting possibilities are constrained.
Every knowledge-based QA system needs a grammar that makes use of constraints to account
for the space of possibilities to formulate questions (Diefenbach et al. 2018). We suggest that
core concept transformations may provide the conceptual basis for a geo-analytical grammar,
where a question’s intent is composed of possible transformations. This grammar can be used
in order to let users formulate questions which can then be automatically translated into mean-
ingful queries over corresponding transformation workflows.Asinthecaseofdataannotations,
the inherent ambiguity can be handled by allowing parts of speech to be parsed in terms of
different concepts.
3.4. The role of core concepts in synthesizing answer workflows
In a certain sense, core concepts describe the origin of spatial information (Scheider et al. 2016). Cor-
respondingly, they also provide necessary constraints for the applicability of functions to given infor-
mation sources towards some geocomputational goal. For example, if we know that a data source
represents a layer of objects, then this implies that it can be counted and thus spatial density can
be computed. This idea can be exploited for loosely specifying and synthesizing GIS workflows
(Lamprecht et al. 2010; Kasalica and Lamprecht 2018). To keep GIS workflow construction (Kind
2014) computationally manageable and to assure a sufficient workflow quality, automated program
synthesis requires semantic constraints for a function’s inputs and outputs which go beyond the cur-
rent geodata types (Hofer et al. 2017). We suggest high-quality workflow synthesis may become
possible using the semantic constraints that come with core concepts. For example, the two workfl-
ows in Figure 6answer our question from Section 1. They were generated based on knowing that,
similar to vector overlay, the ArcGIS tool ‘Polygon to Raster’requires coverages as input (land-
10 S. SCHEIDER ET AL.
use data), while ‘Focal Density’requires object representations (a map of trees). Note that both
workflows generate field representations which are probed by Tom’s GPS track.
The corresponding workflow specifications, consisting of input and output geodata types inter-
preted into core concepts, are given here. Note how the two answer workflow sequences satisfy
these specifications,
Definition 3.1: How much is Tom exposed to green while running through Amsterdam?
in:ObjectVector, EventVector ^out:EventVector
resulting in the following workflow suggestion using core concept datatypes and ArcGIS operator
names:
Workflow 1: ObjectVector, EventVector - Focal Density - FieldRaster - Extract Values to Points -
EventVector
Note that the term ‘green’was now interpreted into objects. Alternatively, one can interpret this
term into a field representation, namely a certain land-use class (Coverage). This would look as
follows:
Definition 3.2: How much is Tom exposed to green while running through Amsterdam?
in:Coverage, EventVector ^out:EventVector
Workflow 2: Coverage, EventVector - Polygon to Raster - FieldRaster - Boolean Reclassify -
Existence Raster - Euclidean Distance - FieldRaster - Extract Values to Points - EventVector
Note while the difference in workflows reflects the ambiguity of semantic interpretation, some
constraints are preserved. For example, the event vector used for probing the exposure field is, in
both cases, origin as well as the goal of the workflow.
Future work should investigate the quality of such core concept based workflow generation
algorithms. A first step has been made in Scheider et al. (n.d.), where workflows similar to the
ones described here could be automatically generated based on core concept data types, specifying
the goal and start concepts and using an extensive set of operational signatures of GIS functions
in OWL.
4. Discussion and conclusion
We have argued that geo-analytical QA, that is, question-answering with a GIS, has a large
potential for data science, yet seems a computational problem very different from ordinary
Figure 6. Exploiting core concepts for workflow composition to generate answers. Green implies field representations, red implies
object representations, and blue implies event representations.
INTERNATIONAL JOURNAL OF DIGITAL EARTH 11
question-answering. While current approaches to geographic QA mainly rely on spatial queries
on factoid knowledge bases, the main difference lies in the fact that geo-analytical knowledge
bases usually do not contain answers but only references to analytical resources. Compared
with a standard QA setting, geo-analytical QA therefore requires accounting for the creativity
of analysis, and for assessing the potential of data sources and tools to answer a given question
(indirect QA). Its relevance lies in spatial questions occurring in all data sciences, while data
scientists are less and less able to learn the variety of functions and data that would allow
them to answer their questions. It should also be noted that indirect QA is not a problem unique
to the geo-spatial domain. For example, statistics face similar problems of overwhelming variety
of tools and data. However, the scope of our study is limited to geo-analytical problems. Given
this scope, we have further argued and illustrated with examples why core concepts are essential
to handle this task. First, they provide many of the needed semantic constraints that capture the
analytic potential of tools and data sources beyond current data types. Second, they are essential
in interpreting and posing spatial questions, in the sense that core concepts are used to construct
and fill the semantic roles in a query. And third, the constraints implied by core concepts can be
exploited for workflow construction in order to compute possible answers, along the lines of
Kasalica and Lamprecht (2018). In order to realize such a geo-analytical QA system, future
work should focus on describing the analytic potential of tools and data. Semantic typing is
needed to capture core concepts across different data representations (Scheider, Ballatore, and
Lemmens 2019). Empirical research is necessary to identify the roles of core concepts in spatial
question patterns. A corresponding grammar would allow us to translate a spatial question into a
query. To investigate the feasibility of answer computation, workflow synthesis methods need to
be tested on annotated resources, and their answering potential needs to be measured by match-
ing queries to workflows using a transformation language based on core concepts. The theoretical
and technical implications of this research may be of relevance not only to the practical appli-
cation of GIS but also to GIS education and training. Any solution to the geo-analytical QA pro-
blem may be turned into a handbook or a manual for GIS analysts. Such manual would assist in
developing necessary expert skills for decomposing geographic questions in terms of core con-
cepts and for mapping them to geodata sets and tools to formulate answer workflows.
Notes
1. Including ArcGIS, PostGIS, QGIS, ERDAS, Grass GIS, R, etc.
2. https://pypy.org/
3. https://cran.r-project.org/
4. Though ‘distance’is usually defined operationally, we consider it nevertheless a concept. ‘Distance’can be rep-
resented both in terms of data and operations and can be even rendered vague in language.
5. https://www.wikidata.org
6. While these kinds of spatial questions are dominating current search engines such as Bing, cf. Hamzei et al.
(2019), they are usually not of much interest for geographic studies.
7. https://lod-cloud.net/
8. https://www.w3.org/RDF/
9. http://dbpedia.org/
10. https://www.mpi-inf.mpg.de/departments/databases-and-information-systems/research/yago-naga/yago/
11. https://wordnet.princeton.edu/
12. In computer science, an abstract data type (ADT) is an abstraction of a type of data in terms of its behavior, i.e.
in terms of the applicability of operations to the data (Liskov and Zilles 1974).
13. http://spatial.ucsb.edu/core-concepts-of-spatial-information/
14. Kuhn (2012) distinguishes content concepts from quality concepts (resolution and accuracy) and base concepts,
such as location.
15. Note that the term lattice as used here does not refer to the corresponding mathematical concept. It rather goes
back to the notion in spatial statistics.
16. https://www.w3.org/TR/owl2-overview/
12 S. SCHEIDER ET AL.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Funding
This work was supported by the European Research Council (ERC) under the European Union’s Horizon 2020
research and innovation programme (grant agreement no. 803498 (QuAnGIS)).
References
Allen, David W2016.GIS tutorial 2: spatial analysis workbook. Esri Press.
Ballatore, Andrea, Simon Scheider, and Rob Lemmens.. 2018.“Patterns of Consumption and Connectedness in GIS
Web Sources.”In The Annual International Conference on Geographic Information Science, 129–148. Springer.
Bao, Junwei, Nan Duan, Ming Zhou, and Tiejun Zhao.. 2014.“Knowledge-Based Question Answering as Machine
Translation.”In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers), Vol. 1, 967–976.
Cai, Guoray, Hongmei Wang, Alan M. MacEachren, and Sven Fuhrmann. 2005.“Natural Conversational Interfaces to
Geospatial Databases.”Transactions in GIS 9 (2): 199–221.
Câmara, Gilberto, U Freitas, and Marco Antônio Casanova. 1995.“Fields and Objects Algebras for GIS Operations.”In
Proceedings of III Brazilian Symposium on GIS, 407–424.
Chen, Wei. 2014.“Developing a Framework for Geographic Question Answering Systems Using GIS, Natural
Language Processing, Machine Learning, and Ontologies.”PhD diss., Ohio State University.
Chrisman, Nicholas. 2002.Exploring Geographic Information Systems. 2nd ed. New York: Wiley.
De Smith, Michael John, Michael F Goodchild, and Paul Longley. 2007.Geospatial Analysis: A Comprehensive Guide to
Principles, Techniques and Software Tools. Leicester, UK: Troubador Publishing Ltd.
Diefenbach, Dennis, Vanessa Lopez, Kamal Singh, and Pierre Maret. 2018.“Core Techniques of Question Answering
Systems over Knowledge Bases: A Survey.”Knowledge and Information Systems 55 (3): 529–569.
Einstein, Albert. 1934.“On the Method of Theoretical Physics.”Philosophy of Science 1 (2): 163–169.
Galton, Antony. 2004.“Fields and Objects in Space, Time, and Space-Time.”Spatial Cognition and Computation 4 (1):
39–68.
Gao, Song, and Michael F. Goodchild. 2013.“Asking Spatial Questions to Identify GIS Functionality.”In 2013 Fourth
International Conference on Computing for Geospatial Research and Application, 106–110. IEEE.
Gupta, Poonam, and Vishal Gupta. 2012.“A Survey of Text Question Answering Techniques.”International Journal of
Computer Applications 53 (4): 0975–8887.
Hamzei, Ehsan, Haonan Li Maria Vasardani, Timothy Baldwin, Stephan Winter, and Martin Tomko.. 2019.“Place
Questions and Human-Generated Answers: A Data Analysis Approach.”In The Annual International
Conference on Geographic Information Science,3–19. Springer.
Heywood, Ian, Sarah Cornelius, and Steve Carver. 2011.An Introduction to Geographical Information Systems. 4th ed.
Harlow, UK: Pearson Prentice Hall.
Hofer, Barbara, Stephan Mäs, Johannes Brauner, and Lars Bernard. 2017.“Towards a Knowledge Base to Support
Geoprocessing Workflow Development.”International Journal of Geographical Information Science 31 (4): 694–
716.
Höffner, Konrad, Jens Lehmann, and Ricardo Usbeck. 2016.“CubeQA–Question Answering on RDF Data Cubes.”In
International Semantic Web Conference, 325–340. Springer.
Höffner, Konrad, Sebastian Walter, Edgard Marx, Ricardo Usbeck, Jens Lehmann, and Axel-Cyrille Ngonga Ngomo.
2017.“Survey on Challenges of Question Answering in the Semantic Web.”Semantic Web 8 (6): 895–920.
Janowicz, Krzysztof, Simon Scheider, Todd Pehle, and Glen Hart. 2012.“Geospatial Semantics and Linked
Spatiotemporal Data–Past, Present, and Future.”Semantic Web 3 (4): 321–332.
Kasalica, Vedran, and Anna-Lena Lamprecht. 2018.“Automated Composition of ScientificWorkflows: A Case Study
on Geographic Data Manipulation.”In 2018 IEEE 14th International Conference on e-Science (e-Science), 362–363.
IEEE.
Khan, Vassilis Javed, Gurjot Dhillon, Maarten Piso, and Kimberly Schelle.. 2016.“Crowdsourcing User and Design
Research.”In Collaboration in creative design, 121–148. Springer.
Kind, Josephine. 2014.“Creation of Topographic Maps.”In Process Design for Natural Scientists, 229–238. Springer.
Kitchin, Rob. 2013.“Big Data and Human Geography: Opportunities, Challenges and Risks.”Dialogues in Human
Geography 3 (3): 262–267.
Kolomiyets, Oleksandr, and Marie-Francine Moens. 2011.“A Survey on Question Answering Technology from an
Information Retrieval Perspective.”Information Sciences 181 (24): 5412–5434.
INTERNATIONAL JOURNAL OF DIGITAL EARTH 13
Kraak, Menno-Jan, and Ferdinand Jan Ormeling. 2013.Cartography: Visualization of Spatial Data. Harlow, UK:
Routledge.
Kuhn, Werner. 2012.“Core Concepts of Spatial Information for Transdisciplinary Research.”International Journal of
Geographical Information Science 26 (12): 2267–2276.
Kuhn, Werner, and Andrea Ballatore. 2015.“Designing a Language for Spatial Computing.”In AGILE 2015, 309–326.
Springer.
Lamprecht, Anna-Lena, Stefan Naujokat, Tiziana Margaria, and Bernhard Steffen.. 2010.“Synthesis-Based Loose
Programming.”In 2010 Seventh International Conference on the Quality of Information and Communications
Technology, 262–267. IEEE.
Laurent, Dominique, Patrick Séguéla, and Sophie Nègre. 2006.“QA Better than IR?”In Proceedings of the Workshop on
Multilingual Question Answering,1–8. Association for Computational Linguistics.
Lin, Jimmy J. 2002.“The Web as a Resource for Question Answering: Perspectives and Challenges.”In Proceedings of
the Third International Conference on Language Resources and Evaluation (LREC-2002), Canary Islands, Spain, 1–8.
Liskov, Barbara, and Stephen Zilles.. 1974.“Programming with Abstract Data Types.”In ACM Sigplan Notices, Vol. 9
(4), 50–59. ACM.
Mai, Gengchen, Bo Yan, Krzysztof Janowicz, and Rui Zhu.. 2020.“Relaxing Unanswerable Geographic Questions
Using a Spatially Explicit Knowledge Graph Embedding Model.”In Geospatial Technologies for Local and
Regional Development. Springer.
Mitchell, Andy. 2012.Modeling Suitability, Movement, and Interaction. Redlands, CA: Esri Press.
Mitra, Arindam, Peter Clark, Oyvind Tafjord, and Chitta Baral.. 2019.“Declarative Question Answering over
Knowledge Bases containing Natural Language Text with Answer Set Programming.”arXiv preprint
arXiv:1905.00198.
Ofoghi, Bahadorreza, John Yearwood, and Liping Ma. 2008.“The Impact of Semantic Class Identification and
Semantic Role Labeling on Natural Language Answer Extraction.”In European Conference on Information
Retrieval, 430–437. Springer.
O’Looney, John. 2000.Beyond Maps: GIS and Decision Making in Local Government. Redlands, CA: ESRI, Inc.
Pulla, Venkata S. K., Chandra S Jammi, Prashant Tiwari, Minas Gjoka, and Athina Markopoulou.. 2013.“QuestCrowd:
A Location-Based Question Answering System with Participation Incentives.”In 2013 IEEE Conference on
Computer Communications Workshops (INFOCOM WKSHPS),75–76. IEEE.
Richardson, Douglas B., Nora D. Volkow, Mei-Po Kwan, Robert M. Kaplan, Michael F. Goodchild, and Robert T.
Croyle. 2013.“Spatial Turn in Health Research.”Science 339 (6126): 1390–1392.
Sawant, Uma, Saurabh Garg, Soumen Chakrabarti, and Ganesh Ramakrishnan. 2019.“Neural Architecture for
Question Answering Using a Knowledge Graph and Web Corpus.”Information Retrieval Journal 22 (3-4): 324–349.
Scheider, Simon, Andrea Ballatore, and Rob Lemmens. 2019.“Finding and Sharing GIS Methods Based on the
Questions They Answer.”International Journal of Digital Earth 12 (5): 594–613.
Scheider, Simon, Benedikt Gräler, Edzer Pebesma, and Christoph Stasch. 2016.“Modeling Spatiotemporal Information
Generation.”International Journal of Geographical Information Science 30 (10): 1980–2008.
Scheider, Simon, and Mark D. Huisjes. 2019.“Distinguishing Extensive and Intensive Properties for Meaningful
Geocomputation and Mapping.”International Journal of Geographical Information Science 33 (1): 28–54.
Scheider, Simon, Rogier Meerlo, Vedran Kasalica, and Anna-Lena Lamprecht.. n.d.“Ontology of Core Concept Data
Types for Answering Geo-Analytical Questions.”http://josis.org/index.php/josis/article/viewArticle/555.
Scheider, Simon, Frank O. Ostermann, and Benjamin Adams. 2017.“Why Good Data Analysts Need to Be Critical
Synthesists. Determining the Role of Semantics in Data Analysis.”Future Generation Computer Systems 72: 11–22.
Scheider, Simon, and Martin Tomko. 2016.“Knowing Whether Spatio-Temporal Analysis Procedures are Applicable
to Datasets.”In FOIS,67–80.
Shah, Asad Ali, Sri Devi Ravana, Suraya Hamid, and Maizatul Akmar Ismail. 2019.“Accuracy Evaluation of Methods
and Techniques in Web-Based Question Answering Systems: A Survey.”Knowledge and Information Systems 58
(3): 611–650.
Simmons, Robert F. 1970.“Natural Language Question-Answering Systems: 1969.”Communications of the ACM 13
(1): 15–30.
Sinton, David. 1978.“The Inherent Structure of Information as a Constraint to Analysis: Mapped Thematic Data as a
Case Study.”Harvard Papers on Geographic Information Systems.
Vahedi, Behzad, Werner Kuhn, and Andrea Ballatore. 2016.“Question-Based Spatial Computing–A Case Study.”In
Geospatial Data in a Changing World,37–50. Springer.
Wang, Mengqiu. 2006.“A Survey of Answer Extraction Techniques in Factoid Question Answering.”Computational
Linguistics 1 (1): 1–14.
Zhang, Zhiwei, Lingling Zhang, Hao Zhang, Weizhuo He, Zequn Sun, Gong Cheng, Qizhi Liu, Xinyu Dai, and
Yuzhong Qu. 2018.“Towards Answering Geography Questions in Gaokao: A Hybrid Approach.”In China
Conference on Knowledge Graph and Semantic Computing,1–13. Springer.
14 S. SCHEIDER ET AL.
