=Paper=
{{Paper
|id=Vol-2455/paper5
|storemode=property
|title=Investigating the Dimensions of Spatial Language
|pdfUrl=https://ceur-ws.org/Vol-2455/paper5.pdf
|volume=Vol-2455
|authors=Adam Richard-Bollans,Lucía Gómez Álvarez,Brandon Bennett,Anthony Cohn
|dblpUrl=https://dblp.org/rec/conf/cosit/Richard-Bollans19
}}
==Investigating the Dimensions of Spatial Language==
https://ceur-ws.org/Vol-2455/paper5.pdf
Investigating the Dimensions of Spatial Language
Adam Richard-Bollans
University of Leeds, UK
mm15alrb@leeds.ac.uk
Lucía Gómez Álvarez
University of Leeds, UK
sc14lga@leeds.ac.uk
Brandon Bennett
University of Leeds, UK
b.bennett@leeds.ac.uk
Anthony G. Cohn
University of Leeds, UK
a.g.cohn@leeds.ac.uk
Abstract
Spatial prepositions in the English language can be used to denote a vast array of configurations
which greatly diverge from any typical meaning and there is much discussion regarding how their
semantics are shaped and understood. Though there is general agreement that non-geometric aspects
play a significant role in spatial preposition usage, there is a lack of available data providing insight
into how these extra semantic aspects should be modelled. This paper is aimed at facilitating the
acquisition of data that supports theoretical analysis and helps understand the extent to which
different kinds of features play a role in the semantics of spatial prepositions. We first consider key
features of spatial prepositions given in the literature. We then introduce a framework intended
to facilitate the collection of rich data; including geometric, functional and conventional features.
Finally, we describe a preliminary study, concluding with some insights into the difficulties of
modelling spatial prepositions and gathering meaningful data about them.
2012 ACM Subject Classification Computing methodologies → Natural language processing; Com-
puting methodologies → Lexical semantics
Keywords and phrases Lexical Semantics, Spatial Prepositions, Situated Dialogue
Supplement Material github.com/alrichardbollans/spatial-preposition-annotation-tool-blender
1 Introduction
Spatial prepositions in the English language can be used to denote a vast array of configura-
tions which greatly diverge from any typical meaning and there is much discussion regarding
how their semantics are shaped and understood. Though there is general agreement that
non-geometric aspects play a significant role in spatial preposition usage, there is a lack of
available data providing insight into how these extra semantic aspects should be modelled.
This paper is aimed at facilitating the acquisition of data that supports theoretical
analysis and helps understand the extent to which different kinds of features play a role in
the semantics of spatial prepositions. We first consider key features of spatial prepositions
given in the literature. We then introduce a framework1 intended to facilitate the collection
of rich data; including geometric, functional and conventional features. Finally, we describe
a preliminary study, concluding with some insights into the difficulties of modelling spatial
prepositions and gathering meaningful data about them.
1
github.com/alrichardbollans/spatial-preposition-annotation-tool-blender/ 47
Copyright © 2019 for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)
In Proceedings Speaking of Location 2019: Communicating about Space, Regensburg, Germany, September-2019.
Editors: K. Stock, C.B. Jones and T. Tenbrink (eds.);
Published at http://ceur-ws.org
48 Investigating the Dimensions of Spatial Language
2 Background
Regarding the names of the objects being discussed we use figure (also known as: target,
trajector, referent) to denote the entity whose location is important e.g. ‘the bike next to
the house’ and ground (also known as: reference, landmark, relatum) to denote the entity
used as a reference point in order to locate the figure e.g. ‘the bike next to the house’.
2.1 Semantic Complexity
Initial attempts to understand and model spatial language naturally focused heavily on
geometry. However, as has been recognised in the past couple of decades, spatial constraints
are not enough to fully characterise spatial prepositions [1, 5, 6, 10, 14]. The use of prepositions
is determined by geometric, functional and conventional considerations, as evidenced in
[5, 8, 10].
Talmy [25] introduced and highlighted the importance of ‘force-dynamics’ in language and
cognition, considering the force interactions of objects as a primitive notion that pervades
language through metaphor. This work inspired future researchers to pay more attention to
the force interactions present; most notably in the investigations of [6, 10] which considered
the interactions of geometry and functionality in spatial semantics, in particular highlighting
that the functional control of the ground over the figure strongly influences preposition usage.
Garrod et al. [10] give the well-cited example that a pear may be considered as in a bowl
when it is not even partially contained by the convex hull of the bowl — if it is sat on top
of a pile of other pears in the bowl. We also see examples of this in our collected data. It
has also been shown that the way objects are labelled and conceptualised affects preposition
use. Coventry et al. [5] found that when given exactly the same scene of an object on a
plate/dish, humans will describe the configuration as in when the ‘plate’ is labelled as a dish,
and on when labelled as a plate. It is suggested that this is due to the affordances associated
with the concepts ‘plate’ and ‘dish’.
Following [22], we therefore believe that a full semantic account of spatial prepositions,
particularly for on and in, ought to include distinct functional & conventional components in
order to (1) be closely aligned with human usage and understanding and (2) aid automated
interpretation and generation of spatial expressions.
2.2 Related Work
The work of Platonov & Schubert [21] is closely aligned with the current work. In order to
create and test a computational model of spatial prepositions they also generate 3D scenes
to annotate. The annotations are created via two separate tasks, a ‘Truth-Judgement Task’
and a ‘Description Task’. In both tasks participants are shown screenshots of the scene. In
the truth judgement task participants are asked if a preposition is fitting for two objects,
e.g. ‘Is the cube on the table?’, and respond by selecting an option from ‘Yes’, ‘Rather Yes’,
‘Uncertain’, ‘Rather No’, ‘No’. In the Description Task participants are given an object, by
referring to the object label e.g. ‘Where is Pencil 1?’, and asked to provide a description of
the location of the object. On the whole, despite some minor deficiencies2 , this represents
a valuable groundwork for our framework, which we enrich by supporting tasks in 3D
environments (instead of screenshots) where participants can navigate the scenes and select
objects. We believe that such additional features may be crucial for the integration of certain
2
Many annotations were obtained from blocks world environments or simple uncluttered indoor scenes.
A. Richard-Bollans, L. Gómez Álvarez, B. Bennett, and A. G. Cohn 49
pragmatic factors in different tasks as well as provide more flexibility for the exploration of
borderline geometrical configurations. Moreover, Platonov & Schubert’s framework does not
seem to be intended for reuse by different projects.
There exist some wide-ranging annotated datasets, see [9, 27], however the annotations
are not restricted to spatial language and many of the datasets are image-based so extracting
meaningful features is extremely difficult.
Many experimental studies have been conducted over the past couple of decades into
particular aspects of spatial prepositions; however these are either image-based [2, 7, 15,
17, 19] or in real environments [10, 20, 26] where feature extraction is difficult; or in very
constrained/simple environments [8, 11, 12, 16].
Overall, we find that there is a lack of detailed geometric, functional and contextual
data which hinders the capacity to properly investigate the semantic complexity of spatial
prepositions and provide pragmatic analysis on how they are used to achieve communicative
success.
3 Framework
Our framework1 is built on the 3D modelling software Blender3 . This software allows 3D
scenes to be created, which can then be converted into annotation environments using the
built-in game engine. Scenes are easy to create, simply taking the given scene template and
populating it with objects, see the GitHub repository1 for instructions. See Figure 1 for
a screenshot of a scene from our study (some object labels have been added but were not
visible for participants during the study). To allow for easy selection, objects in the scene
are indivisible entities e.g. a table in the scene can be selected but not a particular table leg.
Figure 1 Example Scene
Once a scene has been created, a python script is run in order to create distinct tasks
from the scene. Two tasks that we used for our preliminary study are described below.
3
https://www.blender.org/
50 Investigating the Dimensions of Spatial Language
3.1 Tasks
The framework initially provides two distinct tasks with which our preliminary study was
conducted — a Selection Task and a Description Task. The Selection Task was designed to
efficiently collect large amounts of data regarding the semantics, with minimal pragmatic
considerations. The Description Task provides more focused data which is intended to aid
pragmatic analysis and test models of figure selection. In both tasks, participants are given
a first person view of a scene which they can navigate using the mouse and keyboard.
In the Selection Task participants are given a preposition on screen and asked to select all
figure-ground pairs in the scene which fit the preposition. Once they have selected all pairs
they believe to be admissible they are shown another preposition and asked to repeat the
process. As the output of each selection is a figure, preposition and ground no post-processing
is required to identify these. In our preliminary study, we limited this task to prepositions
we believe to have a functional component along with those prepositions that seem to act as
a geometric counterpart: ‘in’, ‘inside’, ‘on’, ‘on top of’, ‘against’, ‘over’, ‘under’, ‘above’ and
‘below’.
In the Description Task objects are highlighted and participants are able to type in
a spatial description of the object. In our preliminary study we asked participants to
give descriptions of the object locations using a definite description, in the format figure
+ preposition + ground e.g. ‘the guitar by the bookshelf’. We also allowed the use of
multiple prepositions if the participant deemed it necessary e.g. ‘the cup on the table near
the lamp’. In order to increase the number of annotations containing prepositions with a
functional component, while still allowing participants a choice of natural descriptions, we
asked participants to only use the prepositions in the Selection Task plus ‘to the right of’,
‘to the left of’, ‘in front of’, ‘behind’, ‘near’, ‘next to’, ‘at’.
3.2 Feature Extraction
The use of virtual 3D environments allows for the extraction of a wide range of features that
would not be immediately available in real-world or image based studies.
3.2.1 Geometric Features
Geometric features (distance between objects, object size etc..) are relatively simple to extract.
We made use of and adapted some existing code4 for this purpose, see [21]. The geometric
features extracted are all quantitative rather than qualitative and some simplifications have
been made. For example, we measured containment as the proportion of the bounding box
of the figure that is shared with the bounding box of the ground. This measure may be
improved by refining how the shared volume is calculated (e.g. by using convex hulls rather
than bounding boxes) or by distinguishing separate parts of the ground where overlaps with
the figure are more important (e.g. the containable inside [4] of the ground). However,
calculations involving convex hulls can become computationally expensive and automated
demarcation of such salient parts of objects is a non-trivial task.
4
https://github.com/gplatono/SRP/tree/master/blender_project Date Accessed: 22/11/2018
A. Richard-Bollans, L. Gómez Álvarez, B. Bennett, and A. G. Cohn 51
3.2.2 Functional Features
There are two particular functional notions that appear over and over in the literature:
‘support’ and ‘location control’. It is often considered that ‘on’ expresses the functional notion
of ‘support’ while ‘in’ expresses ‘location control’ [3]. We take ‘support’ to express that
the ground impedes motion due to gravity of the figure, while ‘location control’ expresses
that moving the ground moves the figure. By including these features in our data we
hope to provide significant support for the theories of [6, 10]. Rather than attempting to
formally define these notions, as in [13], following Sloman [23] we quantified these notions via
simulation using Blender’s built-in physics engine. To assess the degree to which a ground
gives support to a given figure, we measure the change in vertical position of the highest
point of the figure when the ground is removed from the scene. We then normalised this
measure by dividing by the height of the ground.
For simplicity, we have used this support measure as a proxy for location control as it is
easier to quantify and calculate from a scene. Further, the notions are conceptually quite
similar and there are findings to suggest that the notions of containment and support are
closely related in humans’ conceptual framework [18].
3.2.3 Conventional Features
In order to assess the degree to which non-contextual features affect preposition use, we
used ConceptNet [24] to gather object properties. ConceptNet is a large-scale relational
knowledge base for commonsense knowledge, taking information from multiple crowd-sourced
projects. Currently, we only considered a single commonsense feature — to discern whether
the ground object is considered a container. To do this, for a given ground, we extracted the
weight of the ‘IsA’ edge between the ground and the concept ‘container’.
4 Preliminary Study
In this section we describe a preliminary study carried out with our framework and discuss
the results. See the GitHub repository1 for the collected data.
We created three separate virtual indoor environments which were each limited to around
60 household objects. Each scene contained a robot and for the Description Task the
participants were asked to describe the objects in such a way that the robot would pick up
the correct object in order to bring it to them. We chose to only highlight objects which were
not uniquely identifiable in the scene by their name i.e. where there were multiple objects of
the same type in the scene.
4.1 Session & Resulting Data
To carry out the study we hosted a session in one of our computing labs and invited
participants from across the university campus to take part5 . We asked participants to
complete either the Selection Task or Description Task in one of our scenes. Participants
were given a brief introduction to the study and instructions on how to complete the given
task along with explanations of key terminology6 . All participants were asked to create
5
Ethics Approval Code: 271016/IM/216. Participants were given the incentive of free pizza.
6
See GitHub repository1 for instructions that were given.
52 Investigating the Dimensions of Spatial Language
Table 1 Selection Task: Total number of selections made for each preposition & the number of
distinct configurations selected
Preposition Above Over Below Under Against In Inside On On top of
Selections 103 15 32 71 53 27 29 208 171
Distinct
83 14 30 65 46 15 15 149 91
Configurations
annotations for ten minutes. 23 native English speakers participated, of whom 13 performed
the Selection Task and 10 performed the Description Task.
For the Selection Task only minimal cleaning of the data was required. We removed one
user that appeared to give selections at random (52 Selections). We also removed annotations
repeated by the same user (153 Selections), annotations in which the same object was selected
for figure and ground (19 Selections) and also annotations where a user had selected a pair
of objects twice but reversing the figure-ground selection (11 Selections). Also, during the
study participants stated that they had wanted to select the room to say there was objects
‘in’ or ‘inside’ it. We therefore removed annotations that we believe were intended for this
purpose (4 Selections). Post-cleaning, we obtained 709 annotations in the Selection Task; see
Table 1 for a breakdown of selections by preposition.
We obtained 245 annotations in the Description Task. We did not clean this data in any
way but singled out ‘simple’ descriptions (consisting of a single figure, ground and preposition)
to make analysis easier. We applied the off-the-shelf NLTK7 Part-Of-Speech (POS) tagger,
which labels each individual word with a POS tag, to find the simple descriptions. We added
some hard-coded rules to deal with prepositional phrases of more than one word which are
present in our data. This left some issues regarding more complex noun phrases e.g. ‘the
book on the square table’ where the parser identifies three nouns — ‘book’, ‘square’ and
‘table’. However, on the whole most of the simple phrases were identified and the phrases
identified as ‘simple’ were indeed simple. There are 114 such ‘simple’ descriptions.
In the following sections we look at potential insights gained from the preliminary study,
focusing on ‘in’ and ‘on’.
4.2 Selection Task
Containment & Containers
Apart from the exception of the bookcase, regardless of the geometric and functional relations,
users annotating our scenes would only select ‘in’ for a pair of objects if the ground was a
container 8 . We believe this sort of insight is potentially useful for systems aiming to generate
realistic natural language expressions and in making pragmatic inferences related to possible
descriptions of an object.
Similarity of ‘in’ & ‘on’
We found that there was significant overlap between ‘in’ and ‘on’. This highlights some of
the complexity of modelling this language; borderline scenarios are common and often more
than one preposition can be used in a given situation. Out of the 15 distinct configurations
7
http://www.nltk.org/ Date Accessed: 05/11/18
8
We call an object a container if there exists an ‘IsA’ edge between it and ‘container’ in ConceptNet.
A. Richard-Bollans, L. Gómez Álvarez, B. Bennett, and A. G. Cohn 53
Table 2 Success of simple models
Proximity Containment
Preposition On On top of In Inside Against In Inside
Correct 25 8 6 1 2 6 1
Incorrect 27 4 1 1 1 1 1
Unmatched 3 2 4 0 0 4 0
that were selected as ‘in’, 12 of them were also selected as ‘on’. We believe there were two
main reasons for this: ambiguity of the ground’s role as a container & the ground being a
container but the relationship between figure and ground being typical for ‘on’. For five of
these configurations it is ambiguous whether the ground is a ‘container’ e.g. books were
labelled as both ‘in’ and ‘on’ the bookcase/shelf. We believe this may partly be explained by
synecdoche i.e. the bookcase is being used to refer to one of its particular shelves which the
book is on. We also believe this is partly due to the fact that ‘in’ is generally used where the
ground is a container, and it is somewhat ambiguous as to whether a bookcase is a container
or not and therefore whether ‘in’ or ‘on’ is suitable. For another five of these configurations
the ground is a container, but the relationship between the figure and ground is near the
typical notion for ‘on’, for example ‘banana’ and ‘bowl’ in Figure 1.
Location Control & Support
A key aspect that we wanted to analyse was the role of functional features. Support appears
important to ‘on’ — the average value of support for ‘on’ was 0.72 compared to an average
for all prepositions of 0.41. However, this measure by itself was not indicative of ‘in’, having
an average value of 0.4. We believe, however, that this measure will be more informative in
a more nuanced model that accounts for the distinct but related meanings, or polysemes, of
‘in’. These results suggest that refinements of our data collection are necessary in order to
make any significant claims.
The Selection Task relied on the thoroughness of participants in selecting all admissible
configurations for each preposition, however we found that our scenes likely contained too
many objects for this to be a reliable outcome for every participant. This hampered our
efforts to provide any significant analysis.
4.3 Description Task
Detailed Semantics & Communicative Success
Many computational models of spatial prepositions achieve some success when only accounting
for relatively simple geometric features. It is evident that in order to locate items when given
a spatial description (given that all the objects in the scene are known), one can usually
achieve a high degree of accuracy without a detailed semantic model. We expect that, when
given a locative expression, these sorts of models would often be sufficient to locate the
figure.
Moreover, though prepositions may encode a large amount of information and the speaker’s
choice of preposition is dependent on many features, some often imply proximity and this
alone can be useful in locating the figure object. We imagine that ‘in’,‘on’,‘by’,‘at’ and
‘against’ expressions can often be successfully decoded by only considering proximity to the
ground.
54 Investigating the Dimensions of Spatial Language
To test this we constructed simple models that would take a definite description as input
and output an object from the scene. Given an expression ‘figure + preposition + ground’
e.g. ‘the book on the table’, the models would first find all possible figure-ground pairs e.g.
all possible books and tables in the scene. The Proximity Model would then output the
figure object where the pair has the shortest distance. The Containment Model would output
the figure object where the pair has the highest degree of containment. See Table 2 for the
results of these models9 .
Note that due to the ambiguous nature of many of the given descriptions (discussed
below), it is not possible to give an accurate measure of the success of these models. These
results, however, indicate that simple strategies can often be successful.
Though in general simple solutions can be useful for the problem of figure selection in
indoor environments, it is clear there are situations where more nuance in the model would be
necessary. It is also not clear that such simple models would be robust enough to be reusable
across different domains and tasks. Moreover, we expect that a more detailed picture is
necessary for the purposes of language generation. However, we need to conduct further
studies to clarify this.
Pragmatics & Overcoming Ambiguity
We hoped that participants would give descriptions that were ambiguous (e.g. ‘the book on
the table’ where there are two instances of a book on a table) that would be unambiguous in
practice due to pragmatic considerations. Such pragmatic considerations may encapsulate a
wide variety of world-knowledge, from relatively shallow features such as object occlusion to
more task-dependent and commonsense factors such as proximity of an object to the speaker,
as well as semantic knowledge, such as the acceptability of potential descriptions.
When analysing the ‘simple’ descriptions given by participants (containing only one
preposition), even as a human, many appear genuinely ambiguous e.g. ‘the notebook on the
table’ in Scene 2 (see Figure 1). This could be as a result of participants not being fully
aware of all the objects in the scene, not having properly read the instructions or that the
aim of the task was not made clear enough. Further, to discern pragmatic strategies used by
participants it is necessary to focus on a specific pragmatic aspect as otherwise the problem
is too varied and complex. As a result, we were unable to provide any meaningful pragmatic
analysis. Further study is necessary and modifications need to be made in order to ensure
the collected data is useful in future.
5 Conclusions
In this paper we have introduced our framework for collecting data on spatial prepositions
and discussed a preliminary study conducted. By creating a framework that allows for
the collection of rich data on spatial prepositions which includes geometric, functional and
conventional features, we intend to begin a process of providing large-scale support for
theoretical claims in the field, which will hopefully help not only to clarify their semantics but
also to develop appropriate strategies to model them. Along these lines, our preliminary study
has given some tentative insights and highlighted aspects of our data collection methodology
that need further refining.
9
Objects were unmatched where a suitable ground was not recognised.
A. Richard-Bollans, L. Gómez Álvarez, B. Bennett, and A. G. Cohn 55
6 Future Work
In order to build upon the current work, based on some of the discussion provided here,
we are extending and improving upon the existing framework and running further studies.
We have split the Selection Task into two separate tasks so that we can provide a more
meaningful and robust analysis. One task shows participants a specific configuration and
asks which prepositions could be used to describe it. This means that we are less reliant on
participants being thorough and also can collect non-instances of a preposition. A second
task shows participants a ground object as well as potential figure objects and asks them to
select the figure object that best fits a given preposition. This will allow us to directly assess
how features influence the typicality of spatial prepositions. The data collected in these
first two tasks will be used to inform a prototype dialogue system that will be implemented
in a description task. This task will take the form of a game, such as the user providing
descriptions to collect objects in a scene for some task where they are given a score based on
their performance. Our prototype system will be used to interpret the user’s description and
if it guesses correctly, the object will disappear from the scene and be added to the users
inventory. Finally, in order to facilitate larger ongoing studies, we have created an updated
web version of our framework which is currently online10 .
References
1 John A Bateman, Joana Hois, Robert Ross, and Thora Tenbrink. A linguistic ontology of
space for natural language processing. Artificial Intelligence, 174(14):1027–1071, 2010.
2 Melissa Bowerman and Erik Pederson. Topological relations picture series. In Space stimuli
kit 1.2, page 51. Max Planck Institute for Psycholinguistics, 1992.
3 Laura Carlson, Emile Van der Zee, and Emile Zee. Functional features in language and space:
insights from perception, categorization, and development, volume 2. Oxford University Press
on Demand, 2005.
4 Anthony G Cohn, David A Randell, and Zhan Cui. Taxonomies of logically defined qualitative
spatial relations. Journal of human-computer studies, 43(5-6):831–846, 1995.
5 Kenny R. Coventry, Richard Carmichael, and Simon C. Garrod. Spatial prepositions, object-
specific function, and task requirements. Journal of Semantics, 11(4):289–309, 1994.
6 Kenny R Coventry and Simon C Garrod. Saying, seeing and acting: The psychological
semantics of spatial prepositions. Psychology Press, 2004.
7 Kenny R. Coventry, Mercè Prat-Sala, and Lynn Richards. The Interplay between Geometry
and Function in the Comprehension of Over, Under, Above, and Below. Journal of Memory
and Language, 44(3):376–398, 2001.
8 Michele I Feist and Derdre Gentner. On plates, bowls, and dishes: Factors in the use of English
IN and ON. In Proc 20th annual meeting of the cognitive science society, pages 345–349, 1998.
9 Francis Ferraro, Nasrin Mostafazadeh, Ting-Hao, Huang, Lucy Vanderwende, Jacob Devlin,
Michel Galley, and Margaret Mitchell. A Survey of Current Datasets for Vision and Language
Research. arXiv:1506.06833 [cs], 2015.
10 Simon Garrod, Gillian Ferrier, and Siobhan Campbell. In and on: investigating the functional
geometry of spatial prepositions. Cognition, 72(2):167–189, September 1999.
11 Dave Golland. Semantics and Pragmatics of Spatial Reference. PhD Thesis, University of
California, Berkeley, USA, 2013.
10
http://adamrichard-bollans.co.uk/spatial_language_project.html
56 Investigating the Dimensions of Spatial Language
12 Sergio Guadarrama, Lorenzo Riano, Dave Golland, Daniel Go, Yangqing Jia, Dan Klein, Pieter
Abbeel, and Trevor Darrell. Grounding spatial relations for human-robot interaction. In Proc
IROS, pages 1640–1647. IEEE, 2013.
13 Maria M. Hedblom, Oliver Kutz, Till Mossakowski, and Fabian Neuhaus. Between Contact
and Support: Introducing a Logic for Image Schemas and Directed Movement. In Proc IAAI,
volume 10640, pages 256–268. Springer, 2017.
14 Annette Herskovits. Language and spatial cognition. Cambridge University Press, 1987.
15 John Kelleher, Colm Sloan, and Brian Mac Namee. An investigation into the semantics of
English topological prepositions. Cognitive Processing, 10(S2):233–236, September 2009.
16 John D. Kelleher and Fintan J. Costello. Applying computational models of spatial prepositions
to visually situated dialog. Computational Linguistics, 35(2):271–306, 2009.
17 John D. Kelleher, Geert-Jan M. Kruijff, and Fintan J. Costello. Proximity in context: an
empirically grounded computational model of proximity for processing topological spatial
expressions. In Proc ACL, pages 745–752. Association for Computational Linguistics, 2006.
18 Jean M Mandler. How to build a baby: II. Conceptual primitives. Psychological review,
99(4):587, 1992.
19 David M. Mark and Max J. Egenhofer. Topology of prototypical spatial relations between
lines and regions in English and Spanish. In Proc Auto Carto, pages 245–254, 1995.
20 Reinhard Moratz and Thora Tenbrink. Spatial reference in linguistic human-robot interaction:
Iterative, empirically supported development of a model of projective relations. Spatial
cognition and computation, 6(1):63–107, 2006.
21 Georgiy Platonov and Lenhart Schubert. Computational Models for Spatial Prepositions. In
Proc 1st International Workshop on Spatial Language Understanding, pages 21–30, 2018.
22 Adam Richard-Bollans. Towards a Cognitive Model of the Semantics of Spatial Prepositions.
In ESSLLI Student Session Proceedings. Springer, 2018.
23 Aaron Sloman. Aiming for More Realistic Vision Systems. Technical Report COSY-TR-0603,
University of Birmingham, 2006.
24 Robert Speer and Catherine Havasi. Representing General Relational Knowledge in ConceptNet
5. In LREC, pages 3679–3686, 2012.
25 Leonard Talmy. Force dynamics in language and cognition. Cognitive science, 12(1):49–100,
1988.
26 Thora Tenbrink, Elena Andonova, Gesa Schole, and Kenny R. Coventry. Communicative
Success in Spatial Dialogue: The Impact of Functional Features and Dialogue Strategies.
Language and Speech, 60(2):318–329, 2017.
27 Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton van den Hengel.
Visual Question Answering: A Survey of Methods and Datasets. arXiv:1607.05910 [cs], 2016.