Stimulating and Simulating Creativity with Dr Inventor
*
Diarmuid P. O’Donoghue, *
Yalemisew Abgaz, *
Donny Hurley, §Francesco Ronzano, §Horacio
Saggion *Department of Computer Science, Maynooth University, Ireland. §Universitat Pompeu Fabra, Barcelona, Spain
diarmuid.odonoghue@nuim.ie
Abstract
Dr Inventor is a system that is at once, a computational
model of creative thinking and also a tool to ignite the
creativity process among its users. Dr Inventor uncovers
creative bisociations between semi-structured documents
like academic papers, patent applications and
psychology materials, by adopting a “big data” perspective
to discover creative comparisons. The Dr Inventor
system is described focusing on the transformation of
this textual information into the graph-structure required
by the creative cognitive model. Results are described
using data from both psychological test materials
and published research papers. The operation of Dr
Inventor for both focused creativity and open ended
creativity is also outlined.
Introduction
This paper describes the Dr Inventor project that is both a
creativity support tool while its internal operation means
that is also functions as a model of creative discovery. One
of the core artifacts processed by Dr Inventor to boost scientific
creativity is represented by Research Objects (RO)
(Belhajjame et al., 2012), which are creative academic outputs
including academic publications, patent applications
and related data. Dr Inventor aims to actively explore creative
bisociations (Koestler, 1964) between these Research
Objects using a cognitively inspired model of creative
thinking. This paper adopts a big data perspective on Research
Objects attempting to uncover latent creative comparisons
that might lie undiscovered within its dataset. Dr
Inventor directly addresses two of Honavar’s (2014) facets
of computationally mediated scientific discovery: firstly
the development of computational representations and secondly,
computationally augmenting scientific discovery.
This paper is structured as follows. We first present a
case for bisociative and analogy-based creativity, addressing
some issues arising from Boden’s attribution of bisociative
reasoning to a category called “combinatorial
creativity” (Boden, 1998). We then describe the Dr Inventor
model, focusing on the processes that enable it to identify
analogies between its text-based inputs. Next, we outline
some results from text-based sources including human
psychological tests and published research papers, illustrating
its operation as both a tool for focused creativity and
also for open ended creativity. Finally a summary and
some concluding remarks are made.
Analogical Reasoning and Creativity
The model of bisociative reasoning developed in this paper
is built primarily on a computational model of analogical
reasoning, which is extended to include additional background
information. While computational treatments of
analogy originally focused on the analogy per se, recent
attention has focused more on situated models addressing
topics like Ravens Progressive Matrices (Kunda,
McGreggor and Goel, 2013). The analogy process provides
a unique perspective from which to view computational
creativity, lying at the crossroad of research in areas including
cognitive science (Gick and Holyoak, 1980), developmental
psychology (Rattermann and Gentner, 1998),
computer science (Ramscar and Yarlett, 2003;
O'Donoghue, Bohan and Keane, 2006; O'Donoghue and
Keane, 2012) and neuroscience (Green et al., 2010) . Research
in these areas often constrain one another and offer
the possibility of uncovering truly deep insights into the
creative process. This may ultimately lead to formation of
a cohesive multi-perspectival vision of one mode of creativity.

Analogy in Creative Reasoning
Psychological evidence has highlighted people’s ability to
reason using analogical comparisons in the laboratory setting
(Gick and Holyoak, 1980). Subjects are typically presented
with two analogous stories and are required to develop
the latent analogy as a key to solving a problem in
one of those stories. Later in this paper we shall demonstrate
Dr Inventor’s ability to take the texts used in these
psychology tests and develop the same analogies as observed
in (many) human participants in these trials.
 The use of analogy has also been described in a “real
world” scenario. Blanchette and Dunbar (2001) recorded
and described the use of real-world analogies during laboratory
meetings of molecular biologists and immunologists.
They examined 16 different meetings in a number of different
laboratories. They identified over 99 analogical
comparisons and scientists typically used anything from 3
Proceedings of the Sixth International Conference on Computational Creativity June 2015 220
to 15 analogies in a one-hour meeting. The majority of the
analogies discovered were between biological and immunological
information – the so called “within-domains”
analogies. However , the authors noted that scientists used
more “between-domains” analogies (involving semantically
distant source domains such as literature or engineering),
when the goal involved a creative task such as formulating
an hypothesis.
Goldschmidt et al, (2011) and others have highlighted
that “problem fixation” often frustrates peoples efforts to
think creatively. That is, people experience difficulties in
seeing new uses for existing information. The authors argue
that to overcome this fixation and to promote creative
thinking, that people be presented with semantically distant
comparisons for a given problem. Research by Bowden et
al (2005) and others has highlighted that insight occurs
when problem solvers suddenly see a connection that previously
eluded them. One possible mechanism of supporting
insight is the discovery or a creative bisociations, like
analogies and blends (Fauconnier and Turner, 1998).
Analogy and Transformational Creativity
Margaret Boden (Boden, 1990) offers three well-known
levels of creativity, with increasingly impressive impact at
the levels of improbable, exploratory or transformational
creativity. Boden argues that analogy is effectively the
lowest form of creativity (improbable); however we argue
that when analogical reasoning is seen within the context
of a cohesive system of human reasoning the picture is less
clear. If the inferences mandated by an analogy contradict
some fundamental axiomatic belief, especially beliefs with
that numbers of associated deductions and inferences, then
resolving this contradiction might well involve the “shock
and amazement” associated with Boden's highest level of
transformational creativity. It appears that analogies may in
fact, drive creativity at any of Boden’s levels of creativity.
Our creativity model is domain independent and does not
include a pragmatic component or domain context. So, as
our model does not use domain-specific knowledge, arguably
it cannot be easily cast as one of improbable, exploratory
or transformational creativity in Boden’s terms.
Creativity Producers and Consumers
Creativity is generally seen from the perspective of the
creator. But, Dr Inventor needs to make a distinction between
itself and its users who are consumers of its creative
outputs. O’Donoghue and Keane (2012) made the point
that a creative process may present a creative comparison
so as to highlight the latent similarities, perhaps using terminology
that highlights this commonality. However, discovering
such creative comparisons will generally have to
combat these differences in order to discover that commonality
ab initio.
 When they encounter a creative artefact, the interested
consumer should also experience an episode of creativity,
once they engage properly with the artefact. The process of
engaging with a creative artefact should empower the consumer,
ultimately leading them to a new conceptual space
akin to that of the creator. If the artefact doesn't cause this
reaction, then its creative impact is greatly lessened and
may be considered less creative. So, a truly creative output
is not merely a recorded by-product of the creative experience
of its creator, but it must also engender creativity
within those consumers that engage properly with it. To
achieve this, creative artefacts must have communicative
potential and arguably, multiple creative artefacts may be
necessary to clarify a new conceptual space - or to convince
an unwilling consumer.
 We call secondary creativity the act of engaging with a
creative artefact so as to transform ones conceptual space,
with primary creativity being the initial creative episode.
We believe that secondary creativity is also essential for
truly creative artefacts, helping wide adoption of this new
perspective. Dr Inventor is concerned with both finding
creative bisociations and with presenting these outputs to
its users. It will use both ontology and visual analytics to
support this secondary creativity.
Dr Inventor
Dr Inventor is a computationally creativity system that can
both model scientific creativity and can also use its outputs
to stimulate creative thinking within its users. It is as concerned
with the process of creativity as it is with the products
that arise from these processes (Stojanov and
Indurkhya, 2012). Dr Inventor is built on a cognitively
inspired model of human bisociative reasoning, based on
analogical comparisons and the counterpart projection of
conceptual blends (Fauconnier and Turner, 1998; Veale,
O'Donoghue and Keane, 2000). CrossBee (Jursic et al.,
2012) looked at exploring scientific papers, its focus lay in
finding bridging terms between them. The focus of Dr Inventor
is on finding and extending systemic similarities for
creative purposes.
 This paper focuses primarily on three of the four spaces
of conceptual blending, namely the two input descriptions
and the generic space. The dotted lines in Figure 1 indicate
the correspondences between these inputs, derived with the
help of Gentner’s structure mapping theory (1983). Dr Inventor’s
3-space model identifies a generic space containCounterpart

mapping Input
ROS-1
Generic and
Ontology
Input
ROS-2
Output ROS
Figure 1: Conceptual spaces used by the bisociative model of
Dr Inventor including the analogically founded mapping between
the inputs
Proceedings of the Sixth International Conference on Computational Creativity June 2015 221
ing the ontological similarity between paired relations from
the Input 1 and Input 2. Dr Inventor thus identifies the generic
space corresponding to the aligned items from the
bisociation. This generic space also enables Dr Inventor to
monitor the semantic congruity within a bisociation, to
uncover comparisons more in fitting with the users’ needs.
Finally, the output space represents the new interpretation
of one of those inputs. As each “target” maybe reinterpreted
by multiple sources, and because that target
may also act as a source for some other Research Object
Skeleton (ROS), each newly created ROS is stored separately.
For simplicity, this paper generally uses the terms
source and target, unless specific point about the Blend is
being made.
 This means that a new ROS may act to later inspire subsequent
creativity. Thus, Dr Inventor can potentially operate
as a “Self-sustaining” creativity model as of described
in O’Donoghue et al (2014). One of the chief obstacles
hindering Dr Inventor in achieving this self-sustaining creativity
lies in the quality of the new ROS and a sufficiently
diverse knowledge base from which to progress.
 The core data artefacts used by Dr Inventor are Research
Objects (Belhajjame et al., 2012), which are research outputs
including publications, patents, data, software
(O'Donoghue et al., 2014b), social network information
and other resources. Dealing with such heterogeneous data
sources, characterized by consistent amounts of information
to integrate and process, big data approaches and
technologies are essential in order to enable the computational
approaches to creativity in Dr Inventor. This paper
focuses on the textual contents of RO, particularly of publications
and patents. These documents are first subject to a
number of processing activities to properly mine their contents
in order to generate inputs that are useful to Dr Inventor's
analogy-based model.
 From each RO Dr Inventor generates a graph-based representation
called the Research Object Skeleton (ROS)
representing the key concepts and relationships extracted
from that RO. Dr Inventor identifies similarities between
these ROS with a view to extending these similarities and
uncovering creative possibilities.
Dr Inventor Model
The overall Dr Inventor model contains components that
deal with document summarization, information extraction;
ontology learning, matching and personal recommendation;
ROS generation, assessment, similarity and analogy/blending;
validation, mapping, retrieval and finally visual
analytics. The discussion in this paper will focus on the
ROS generation, analogy/blending model and the creativity
assessment components.
Mining Textual Contents to Populate ROS
In Dr Inventor, Research Object Skeletons (ROSs) are built
by mining the contents aggregated by the corresponding
Research Objects (ROs). To populate a ROS, Dr Inventor
mainly relies on the extraction of information from the
textual contents of a RO. To analyze these contents, Dr
Inventor integrates a Natural Language Processing Pipeline
(DRI-NLP pipeline) that aggregates and customizes several
Information Extraction (Piskorski and Yangarber, 2013)
and Text Summarization (Saggion, 2014) approaches and
tools.
 Since scientific publications constitute one of the main
kinds of textual documents included in a RO, DRI-NLP
pipeline has been properly structured to support the analysis
of research papers. The great majority of papers are
currently available as PDF files. As a consequence, the
conversion of PDF into plain text constitutes an essential
prerequisite to properly perform any further text analysis.
To this purpose, DRI-NLP pipeline relies on PDFX
(Constantin, Pettifer and Voronkov, 2013) that converts a
PDF document of a scientific publication to a semistructured
text (XML). The plain text output of PDFX is
thus processed so as to identify sentences by means of a
custom rule-based sentence splitter. Each sentence is processed
by means of the MATE dependency parser (Bohnet,
2010) to extract dependency relations which are represented
in a dependency tree. DRI-NLP pipeline dependency
parser has been customized in order to properly deal with
several peculiarities of scientific publications, including
the presence of inline citations. In particular, inline citation
markers like “(AuthorA et al.)” or “(1)” are excluded from
the dependency tree if they have no syntactic functions in
the sentence where they are present. Dr Inventor is focused
on the discipline of computer graphics as its test-bed, thus
a particular challenge has been dealing with the many
mathematical expressions in these papers and allowing
their treatment separately from the main body of the text.
Besides dependency parsing, DRI-NLP pipeline enables
the creation extractive summaries of papers by ranking
their sentences by relevance (Saggion, 2014).
 As result of dependency parsing, each word of a sentence
is characterized by its Part-Of-Speech (POS) (noun,
verb, adjective, etc.) and dependency relations (subject,
object, verb chain, modifier of nominal, etc.). The linguistic
information extracted from each publication can be
condensed in the tables: the Syntactic dependency and the
POS tag table. In particular, Figure 2 focuses on the analysis
of a specific sentence taken from the abstract of a paper.
 While Dr Inventor is focused on the test-bed of computer
graphics publications, it remains a general model capable
of dealing with arbitrary text inputs. This paper also
uses data derived from psychology text materials and work
is ongoing using the texts of patent applications.
Figure 2: Processing PDF papers by Dr Inventor Natural Language
Processing Pipeline
Proceedings of the Sixth International Conference on Computational Creativity June 2015 222
ROS Generation
The next task for Dr Inventor is to generate a ROS from
the results of the parsing process. The representation we
chose for these graphs is sufficiently general to represent
different types of RO. Since we want to structure objects
and their inter-relationships this information is stored as a
graph, aimed at supporting the later structure mapping process
(Gentner, 1983).
 Each ROS is constructed as an attributed relational
graph (ARG), which is a directed graph where nodes and
edges may contain additional properties like labels, categories
and numeric values. If required, we can store additional
identifying information (e.g. Author, Affiliation, etc.)
within the graph, but this information is not required for
the analogy process per se.
 The primary information in a ROS is the concept nodes
(nouns) and the relationships (verbs) between them. Concept
nodes are not linked directly to one another but are
connected with relation nodes. To generate the ROS we
use the general structure “subject” – “verb” – object” - as
required by SMT. These triples arise from the dependency
and POS tables as the input to ROS generation. Early testing
has shown that taking triples directly from the dependency
table typically leaves many of them incomplete, leaving
ROS without the necessary structure to support identification
of creative inter-domain correspondences. Therefore,
Dr Inventor performs a deeper exploration of the tables
in order to generate more useable ROS structures.
 By constructing a dependency graph from the tables and
applying a set of heuristics to the graphs, a more complete
set of triples is generated. The heuristics involve combining
some of the nodes and tracing through the graph finding
pairs for each verb.
 Figure 3 depicts two ROSs generated for the “Zerdia”
and “Karla” stories (Table 1) used in human psychological
studies (Gentner and Landers, 1985). They were generated
by the text mining and ROS generation techniques discussed
earlier, but some manual post-editing was performed
to identify co-referencing concepts nodes in the
ROS. In the “Zerdia” story the word “it” is used twice, but
the ROS were edited so one instance was replaced by the
referent “Zerdia” and another by “Gagrach”. In the “Karla”
story the word “she” refers to “Karla” and “he/him” refers
to “hunter”. While these co-referents were resolved manually
work is underway in the text pipeline to automatically
resolve these referents.
 Dr Inventor explicitly represents higher-order (causal
relations connect first-order relations) relations within a
ROS. A distinct set of nodes represents the higher-order
relations, these connecting the first-order (and potentially
other high-order) relations. However, ROS generated from
within our Research Objects corpus show that high-order
(causal) relations are rarely explicitly identified. As we
shall see, this influenced our choice of mapping algorithm.
Karla the Hawk:
Karla, an old hawk, lived at the top of a tall oak tree.
One afternoon, she saw a hunter on the ground with a
bow and some crude arrows that had no feathers. The
hunter took aim and shot at the hawk but missed. Karla
knew the hunter wanted her feathers so she glided down
to the hunter and offered to give him a few. The hunter
was so grateful that he pledged never to shoot at a hawk
again. He went off and shot a deer instead.
 Zerdia True Analogy:
Once there was a small country called Zerdia that
learned to make the world’s smartest computer. One
day Zerdia was attacked by its warlike neighbor,
Gagrach. But the missiles were badly aimed and the
attack failed. The Zerdian government realized that
Gagrach wanted Zerdian computers so it offered to sell
some of its computers to the country. The government of
Gagrach was very pleased. It promised never to attack
Zerdia again
Table 1: Textual contents of “Karla” and “Zerdia”
Figure 3: ROS for the “Karla” and “Zerdia” analogy used in human studies and Dr Inventor
Proceedings of the Sixth International Conference on Computational Creativity June 2015 223
Graph Storage
Dr Inventor uses the Neo4j graph database to store its
ROSs. Neo4j has as its core structures; nodes, relationships
between them and properties on both, this being the same
structure as the ARG. Additional information such as the
SentenceID or SectionTitle for each triple can also be
stored in the Neo4j database. This can be useful when we
want to map only between particular sections (e.g. Abstract
or Conclusion) and also to reference back to the original
sentences from which the triples were extracted.
Data Differences
Previous analogy models like SME (Forbus, Ferguson and
Gentner, 1994), IAM (Keane and Bradshaw, 1988) and
Kilaza (O'Donoghue and Keane, 2012) used hand coded
data. The ROS generated above differs from the earlier
hand-coded data in at least two significant respects. First
ROS contain very few high-order relations, which are
heavily used by mapping models mentioned above. Dr
Inventor does not focus on the hierarchical structure of
hand-coded data, using instead some lower level topological
structure. Secondly (as mentioned in (O'Donoghue and
Keane, 2012)) hand coded data often simplifies the comparison
process by using relations that highlight the latent
similarity. Dr Inventor must uncover and identify the hidden
similarity even in the absence of such lexical cues.
Dr Inventors Creativity Engine
This paper focuses on the creativity engine that lies at the
heart of Dr Inventor. Thus, we focus on creative analogybased
comparisons and show a number of features of Dr
Inventor that specifically attempt to support the identification
and generation of these creative analogies.
Creative Analogies
A number of properties appear to be shared amongst many
creative analogical comparisons (O'Donoghue and Keane,
2012) and these facets are used to generate novel and potentially
useful analogies and blends. Firstly the source
(domain) of inspiration is typically semantically different
from the given target problem. That is, creative sources
tend to be sufficiently different and any similarity is nonobvious
and has not been previously explored in detail.
Secondly, the creative source contains the necessary structural
similarity that is required to generations of viable
analogy with the given problem.
To this end, Dr. Inventor specifically seeks out bisociations
that involve two semantically distant domains, that
form a rich inter-domain mapping and that yield inferences
suggesting something new about one of those domains.
Graph Mapping
To support creative analogies Dr Inventor’s retrieval and
mapping activities makes frequent use of topological features
derived from each ROS. For analogical mapping we
exploit features such as type of the nodes (verb, noun),
types of relationships (subject, object), degree (in-degree,
out-degree) and node rank values calculated by Node Rank
algorithm (Bhattacharya et al., 2012).
We initiate the mapping process by calculating the Node
Rank and by sorting the nodes in a descending order. The
ranking allows us to start the graph matching process from
the most centrally connected and useful node. This will
further be used to serve as a threshold to screen useful
nodes to improve the performance of the mapping process.
The results presented in this paper have been generated
using smaller RO (such as the abstracts of graphics paper
RO), so performance has not been an issue. However this
situation will change when mapping between ROS with
large number of nodes is required.
 The relation (verb) nodes in each ROS are represented
distinctly, with one instance of a relation node for each
verb contained in the RO. Verb nodes are central to the
process of representing the content of the RO, however
their connectedness is limited to a degree of 2 and thereby
affects the resulting Node Rank values. However, multiple
references to the same concept (noun) node will appear in
the ROS as a single concept node – but referenced multiple
times by each distinct relation node. Thus concept nodes
have the greatest direct impact on the Node Rank values as
a single concept node may be linked through many relations
within a ROS. The mapping process avails of this
referential structure when generating the largest graphmapping
between two ROS.
 To identify a pair of mapping nodes from the source
to the target, we used structural similarity score (using the
connectedness of the nodes) and the literal similarity score.
Using structural similarity, we consider two nodes as candidate
mapping nodes if they have a higher similarity
score. Whereas, literal similarity calculates the similarity
coefficient between two words and yields a value between
0 and 1, where 0 indicates no similarity and 1 indicates
complete similarity (synonym). This is achieved by using
the Wu&Palmer (Wu and Palmer, 1994) WordNet-based
similarity metric.
The mapping algorithm firstly selects a pair of nodes
P(sNode, tNode) from the source and the target respectively,
with the highest node ranked nodes being selected first.
In this way, the algorithm focuses on highly connected
nodes within the graph because they contribute most to the
mapping and analogical inference activities. Secondly, the
mapping process checks if the selected pair P(sNode,
tNode) is structurally feasible for analogical mapping. A
structurally feasible pair contains a source node which has
degree (in degree and out degree) greater than or equal to
degree (in degree and the out degree) of the target node
respectively. The comparison ensures the identification of
a sub-graph or an isomorphic graph of the target graph in
the source graph. It further assesses the semantic similarity
of the two nodes using Wu& Palmer. Next, mapping adds
P(sNode, tNode) to the inter-domain map to incrementally
build a mapping sub-graph, if P(sNode, tNode) is feasible.
The mapping stores the pair of mapping nodes along with
their similarity scores.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 224
The mapping process then generates new candidate
mappings by expanding sNode and tNode of P(sNode,
tNode) to their respective connected nodes that are not already
expanded. By following the “subject” or “object”
relationship path it reaches the connected nodes of the
graph, incrementally adding these to the inter-domain
mapping. After the candidate pairs are generated, they are
ranked using their semantic similarity score. Ranking the
candidate pairs will give us a chance to expand pairs with
the highest semantic similarity first.
After including all mappings arising from the initial root
mapping, the process then resumes with the next highest
ranked and unmapped predicates. The algorithm employs
depth first search to expand the nodes in the graph to identify
new mapping pairs. Finally, it selects the mapping that
contains the largest sub-graph and returns the mapping
nodes together with their semantic similarity score.
 We now look at the results produced by generating a
mapping between the “Karla” and “Zerdia” psychology
materials listed above, with the corresponding ROS being
depicted in Figure 3. We note that this simulation of human
analogy process began with the same text materials
that were presented to human subjects. This comparison is
an example to focused creativity, where both the source
and target have been pre-identified.
 The mapping between “Karla” and “Zerdia” gives us 11
mapped nodes between the source and target (Table 2). For
example the noun node “Karla” maps to “Zerdia”, “Feather”
maps to “Computer” and “Hunter” maps to “Gagrach”.
Such a mapping identifies analogous items between the
source and the target and is crucial for transferring new
knowledge form the source to the target. In this specific
example 50% of the nodes in the target ROS are mapped to
the source ROS with an average Wu&Palmer similarity
score of 0.56. The original domains can often include information
that does not participate in the mapping, such as
the (missile be-take-aim attack) in the Zerdia story. However
the absence of this relation from the mapping is not
terribly significant as it is an isolated fragment of information
and does not contribute largely to the main story –
that contributes to the largest connected component of that
ROS.
Mapping Nouns Mapping Verbs
Source Target Source Target
Hunter Gagrach Want Want
Crude World Live Offer_to_sell
Feather Computer Arrow_have Learn_to_make
Want Country Glide_offer_to_give Be_attack
Karla Zerdia Glide Promise_to_attack
Know Call
Table 2: Mapping between “Karla” and “Zerdia”.
Analogies within Graphics Collections
To examine the mapping process, we used 10 papers from
computer graphics domain. The abstracts of the papers
were extracted and were processed using the previously
described steps. Each ROS were mapped to the other 10
ROS including itself. The most basic step is to compare a
ROS against itself. For all the 10 papers Dr Inventor yields
the highest number of mapped nodes when a ROS is compared
with itself – with all or almost all nodes being successfully
mapped. This could be considered as a very basic
step toward the evaluation of the mapping component of
Dr Inventor.
 The mapping of a ROS against the remaining 9 ROS
identifies pairs that have the highest mapping nodes and
pairs that have the lowest mapping nodes. The most analogous
papers are those with large number of mapped nodes
and highest similarity score. For example, the most similar
non-identical mapping among the 10 papers is between
“Bar-Net_Driven Skinning for Character Animation” and
“Real Time Large Deformation Character Skinning in
Hardware” with 14 mapping nodes and an average
Wu&Plamer similarity score of 0.36. While the semantic
similarity score may appear quite low, this was achieved
from within a small collection of papers. We conducted a
quick manual comparison between the abstracts of these
papers and initial results indicate that these papers can be
considered somewhat analogous to one another as for example,
both papers present different approaches to the
computer graphic topics of “skinning”. This analogy arose
from the desire to identifying the largest mapping with the
strongest semantic similarity from within the 10 papers
however, the next section will discuss a more creative Use
Case scenario.
 The lowest mapping occurs between the papers “Curve
Skeleton Skinning for Human and Creature Character”
and “Pose Space Deformation: A Unified Approach to
Shape Interpolation and Skeleton-Driven Deformation”
with 5 mapping nodes and average similarity of 0.35. The
mapping process, as it is expected, is not symmetrical, i.e.
mapping between (S, T) may not be the same as mapping
between (T, S). For example, saying “a man is like a pig”
is not the same as saying that “a pig is like a man”! However,
in this specific data set, it does not significantly affect
the highest mapping node pairs.
Analogical Inference and Pattern Completion
Once we find a mapping between the two ROS, the next
phase is to generate the resulting inferences by applying a
“pattern completion” process to that mapping. This adds
the newly inferred information to the target ROS to produce
the new interpretation of that concept. In the exploratory
analogy mapping process, the user may be interested
to explore all the candidate nodes once he/she knows the
existence of analogy between the source and the target.
Creativity Support and Evaluation in Dr Inventor
Dr Inventor is focused firstly on operating as a Creativity
Support Tool (CST) and secondly, as a simulation of the
analogy process. Shneiderman (2007) noted that there are
no obvious metrics to quantify for CST's and this problem
lies at the core of creativity assessment and evaluation. The
following two approaches are useful to evaluate the level
of creativity support provided by Dr Inventor.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 225
 Among the functionality being developed for Dr Inventor
is an “Inspire Me” Use Case, enabling users to creatively
re-interpret one of their own papers. This will be
achieved by using the paper as a target and searching the
archive for papers that can act as a creative source domain,
forming a large and semantically well-balanced mapping
by making use of the topological structure of each ROS. Dr
Inventor will identify and present to the user those analogies
offering a collection of novel inferences that highlight
the potential benefits of adopting this creative analogical
comparison. Internal metrics will serve to select the most
promising analogies to present to users, assessing the structural
and semantic foundations of the comparison.
 Implicit feedback on the presented analogies will be
gathered by the user interface, enabling comparative evaluation
of different comparisons by monitoring user engagement.
Explicit user feedback also plays a very important
role in evaluating Dr Inventor, using experts in computer
graphics for evaluation. The Creative Support Index (CSI)
(Cherry and Latulipe, 2014) is a psychometric survey that
will serve to assess the creativity support provided by Dr
Inventor. It is quick and easy to administer and is composed
of two sections; a rating scale section and a pairedfactor
comparison section. It identifies 6 major factors of
creativity, namely: enjoyment, exploration, expressiveness,
immersion, results worth effort and collaboration. Under
each of the factors, CSI asks two questions that are rated
between 0 and 10, where 0 indicates the lowest value and
10 indicate the highest achievement. The paired-factor
comparison section consists of each factor paired against
every other factor for a total of 15 comparisons. As Dr Inventor
will support users with different levels of expertise
(first year PhD students to experienced Professors), this
factor in particular will have to be controlled monitored
during the evaluation process.
 The Creative Achievement Questionnaire (CAQ)
(Carson, Peterson and Higgins, 2005) is a very broad and
general creativity assessment technique. Within the context
of Dr Inventor, achievements appear to be primarily assessed
by qualifying the number of published scientific
papers. But the CAQ provides poor coverage of lower levels
of creative achievement (before publication) that could
guide development of the Dr Inventor project. However,
the CAQ might be useful for the final evaluation of Dr
Inventor.
 We also note that (Jordanous, 2014) identifies five criteria
to support meta-evaluation of computational creativity
per se, as opposed to our current focus on Dr Inventor as a
creativity support tool.
Summary and Conclusions
Dr Inventor is a computationally creative model that acts
as both a simulation of human creative reasoning and also
as a creativity support tool. We described how Dr Inventor
performs text extraction from research publications presented
in pdf format, describing how it addresses many
complications that result from use of the pdf format. The
dependency parser was described, as was the process of
constructing the graph representation used by the core
model. Some peculiarities of the resulting graphs were
noted, particularly the extreme rarity of identifiable higherorder
causal relations. Some implications were noted the
process of identifying the inter-domain correspondence.
The text used from human psychological trials showed the
ability of Dr Inventor to generate comparison using these
same textual materials. Results for other Research Objects
were outlined.
 The mapping and evaluation process uses ontological
information as a preference criterion to choose comparisons
with the greatest potential for creativity on Dr Inventor
users. Ontology also opens the way to re-describe the
original documents, highlight the identified similarities.
While this work is ongoing, it opens the way for early
evaluation of Dr Inventor by comparing the impact upon
users of creatively re-interpreted documents.
Acknowledgements
The research leading to these results has received funding
from the European Union Seventh Framework Programme
([FP7/2007-2013]) under grant agreement no 611383.
<references_biblio/>
References
Belhajjame, K., Corcho, O., Garijo, D., Zhao, J., Missier,
P., Newman, D., Bechhofer, S., García, E., Gómez-Pérez,
J.M., Palma, R., Soiland-Reyes, S., Verdes-Montenegro,
L., Roure, D.D. and Goble, C. (2012) 'Workflow-centric
research objects: First class citizens in scholarly discourse',
9th Extended Semantic Web Conference, Hersonissos,
Greece.
Bhattacharya, P., Iliofotou, M., Neamtiu, I. and Faloutsos,
M. (2012) 'Graph-based Analysis and Prediction for
Software Evolution', in Proceedings of the 34th
International Conference on Software Engineering,
Piscataway, NJ, USA: IEEE Press.
Blanchette, I. and Dunbar, K. (2001) 'Analogy use in
naturalistic settings: The influence of audience, emotion,
and goals', Memory & Cognition, vol. 29, no. 5, pp. 730-
735.
Boden, M. (1990) The Creative Mind, Weidenfeld and
Nicolson.
Boden, M. (1998) 'Computer Models of Creativity', in
Sternberg, R.J. Handbook of Creativity, Cambridge
University Press.
Bohnet, B. (2010) 'Very high accuracy and fast
dependency parsing is not a contradiction.', Proc. 23rd
International Conference on Computational Linguistics,
89–97.
Bowden, E.M., Jung-Beeman, M., Fleck, J. and Kounios, J.
(2005) 'New approaches to demystifying insight', Trends in
cognitive sciences, vol. 9, no. 7, pp. 322-328.
Brown, T.L. (2003) Making Truth: Metaphor in Science,
University of Illinois Press.
Carson, S., Peterson, J.B. and Higgins, D.M. (2005)
'Reliability, Validity, and Factor Structure of the Creative
Proceedings of the Sixth International Conference on Computational Creativity June 2015 226
Achievement Questionnaire', Creativity Research Journal,
vol. 17, no. 1, pp. 37–50.
Cherry, E. and Latulipe, C. (2014) 'Quantifying the
Creativity Support of Digital Tools through the Creativity
Support Index', ACM Transactions on Computer-Human
Interaction, vol. 21, no. 4.
Constantin, A., Pettifer, S. and Voronkov, A. (2013)
'PDFX: fully-automated PDF-to-XML conversion of
scientific literature', Proceedings of the 2013 ACM
symposium on Document engineering, 177–180.
Fauconnier, G. and Turner, M. (1998) 'Conceptual
integration networks', Cognitive science, vol. 22, no. 2, pp.
133--187.
Forbus, K.D., Ferguson, R.W. and Gentner, D. (1994)
'Incremental structure-mapping', Proc 16th Conference of
the Cognitive Science Society, 313–318.
Gentner, D. (1983) 'Structure-mapping: A theoretical
framework for analogy', Cognitive Science, vol. 7, pp. 155-
170.
Gentner, D. and Landers, R. (1985) 'Analogical reminding:
A good match is hard to find', Proceedings of the
International Conference on Systems, Man, and
Cybernetics., Tucson, AZ.
Gick, M.L. and Holyoak, K.J. (1980) 'Analogical problem
solving', Cognitive psychology, vol. 12, no. 3, pp. 306-355.
Goldschmidt, G. (2011) 'Avoiding Design Fixation:
Transformation and Abstraction in Mapping from Source
to Target', The Journal of Creative Behavior, vol. 45, no. 2,
pp. 92-100.
Green, A.E., Kraemer, D., Fugelsang, J.A., Gray, J.R. and
Dunbar, K.N. (2010) 'Connecting long distance: semantic
distance in analogical reasoning modulates frontopolar
cortex activity', Cerebral Cortex, vol. 20, no. 1, pp. 70--76.
Honavar, V.G. (2014) 'The Promise and Potential of Big
Data: A Case for Discovery Informatics', Review of Policy
Research, vol. 31, no. 4, pp. 326-330.
Jordanous, A. (2014) 'Stepping Back to Progress Forwards:
Setting Standards for Meta-Evaluation of Computational
Creativity,', Proc ICCC, 129-136.
Jursic, M., Cestnik, B., Urbancic, T. and Larvac, N. (2012)
'Cross-domain Literature Mining', Proc ICCC, pp 33-40.
Keane, M.T. and Bradshaw, M. (1988) 'The Incremental
Analogical machine', in Sleeman, D. (ed.) 3rd European
Working Session on Machine Learning, Kaufmann, CA,
USA.
Keane, M.T., Ledgeway, T. and Duff, S. (1994)
'Constraints on Analogical Mapping: A Comparison of
Three Models', Cognitive Science, vol. 18, no. 3, pp. 387-
438.
Koestler, A. (1964) The Act of Creation, Penguin Books.
Kuhn, T.S. (1962) The structure of scientific revolutions,
University of Chicago press.
Kunda, M., McGreggor, k. and Goel, A.k. (2013) 'A
computational model for solving problems from the
Raven’s Progressive Matrices intelligence test using iconic
visual representations', Cognitive Systems Research, vol.
22-23, pp. 47-66.
O'Donoghue, D.P., Bohan, A. and Keane, M.T. (2006)
'Seeing Things: Inventive Reasoning with Geometric
Analogies and Topographic Maps', New Generation
Computing, 24, (3) (special issue on Computational
Creativity), pp. 267-288.
O'Donoghue, D.P. and Keane, M.T. (2012) 'A Creative
Analogy Machine: Results and Challenges', 4th
International Conference on Computational Creativity
(ICCC), UCD, Dublin, Ireland, 17-24.
O'Donoghue, D.P., Monahan, R., Grijincu, D., Pitu, M.,
Halim, F., Rahman, F., Abgaz, Y. and Hurley, D. (2014b)
'Creating Formal Specifications with Analogical
Reasoning', PICS-Publication Series of the Institute of
Cognitive Science.
O'Donoghue, D.P., Power, J., O'Briain, S., Dong, F.,
Mooney, A., Hurley, D., Abgaz, Y. and Markham, C.
(2014) 'Can a Computationally Creative System Create
Itself? Creative Artefacts and Creative Processes',
nternational Conference on Computational Creativity
(ICCC), Ljubljana, Slovenia.
Piskorski, J. and Yangarber, R. (2013) 'Information
extraction: Past, present and future', Multi-source,
multilingual information extraction and summarization,
pp. 23-49.
Ramscar, M. and Yarlett, D. (2003) 'Semantic grounding in
models of analogy: an environmental approach', Cognitive
Science 27, pp. 41-71.
Rattermann, M.J. and Gentner, D. (1998) 'More evidence
for a relational shift in the development of analogy:
Children's performance on a causal-mapping task',
Cognitive Development, vol. 13, no. 4, pp. 453-478.
Saggion, H. (2014) 'Creating Summarization Systems with
SUMMA.', Proceedings of Language Rescources and
Evaluation Conference.
Shneiderman, B. (2007) 'Creativity support tools:
Accelerating discovery and innovation', Communications
of the ACM, vol. 50, no. 12, pp. 20-32.
Silvia, P.J., Wigert, B., Reiter-Palmon, R. and Kaufman,
J.C. (2012) ' Assessing Creativity with Self-Report Scales:
A Review and Empirical Evaluation.', Psychology of
Aesthetics, Creativity, and the Arts, vol. 6, no. 1, pp. 19-34.
Stojanov, S. and Indurkhya, B. (2012) 'Perceptual
Similarity and Analogy in Creativity and Cognitive
Development.', SAMAI Workshop at ECAI, Montpelier
France, 19-24.
Veale, T., O'Donoghue, D.P. and Keane, M.T. (2000)
'Computation and Blending', Cognitive Linguistics 11
(3/4), pp. 253-281.
Wu, Z. and Palmer, M. (1994) 'Verbs Semantics and
Lexical Selection', in Proceedings of the 32nd Annual
Meeting on Association for Computational Linguistics, Las
Cruces, New Mexico: Association for Computational
Linguistics.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 227