A Computational Model of Analogical Reasoning in Dementia Care
Konstantinos Zachos and Neil Maiden
Centre for Creativity in Professional Practice
City University London
Northampton Square, London EC1V 0HB, UK
{k.zachos, N.A.M.Maiden}@city.ac.uk
Abstract
This paper reports a practical application of a computational
model of analogical reasoning to a pressing social
problem, which is to improve the care of older people
with dementia. Underpinning the support for carers for
people with dementia is a computational model of analogical
reasoning that retrieves information about cases
from analogical problem domains. The model implements
structure-mapping theory adapted to match
source and target domains expressed in unstructured
natural language. The model is implemented as a computational
service invoked by a mobile app used by carers
during their care shifts.
Dementia Care and Creativity
Dementia is a condition related to ageing. After the age of
65 the proportion of people with dementia doubles for every
5 years of age so that one fifth of people over the age of
85 are affected (Alzheimers Society 2010). This equates to
a current total of 750,000 people in the UK with dementia,
a figure projected to double by 2051 when it is predicted to
affect a third of the population either as a sufferer, relative
or carer (Wimo and Prince 2010). Dementia care is often
delivered in residential homes. In the UK, for example, two
in three of all home residents have some form of dementia
(e.g. Wimo and Prince 2010), and delivering the required
care to them poses complex and diverse problems carers
that new software technologies have the potential to overcome.
However, this potential is still to be tapped.
The prevailing paradigm in dementia care is personcentered
care. This paradigm seeks an individualized approach
that recognizes the uniqueness of each resident and
understanding the world from the perspective of the person
with dementia (Brooker 2007). It can offer an important
role for creative problem solving that produces novel and
useful outcomes (Sternberg 1999), i.e. care activities that
both recognize a sense of uniqueness and are new to the
care of the resident and/or carer. However, there is little
explicit use of creative problem solving in dementia care,
let alone with the benefits that technology can provide.
Therefore, the objective of our research was to enable more
creative problem solving in dementia care through new
software technologies.
This paper reports two computational services developed
to support carers to manage challenging behaviors in person-centered
dementia care – a computational analogical
matching service that retrieves similar challenging behavior
cases in less-constrained domains, and a second service
that automatically generates creativity prompts based on
the computed analogical mappings. Both are delivered to
carers through a mobile software app. The next two sections
summarize results from one pre-design study that
motivates the role of analogical matching in managing
challenging behavior in dementia care then describe the
two computational creativity services.
A Pre-Design Study
Creative problem solving is not new to care work. Osborn
(1965) reported that creative problem solving courses were
introduced in nursing and occupational therapy programs
in the 1960s. Le Storti et al. (1999) developed a program
that fostered the personal creative development of student
nurses, challenging them to use creativity techniques to
solve nursing problems. This required a shift in nursing
education from task- to role-orientation and established a
higher level of nursing practice – a level that treated nurses
as creative members of health care teams. There have been
calls for creative approaches to be used in the care of people
with dementia. Successful creative problem solving
was recognized to counteract the negative and stressful
effects that are a frequent outcome of caring for people
with dementia (Help the Aged, 2007). Several current dementia
care learning initiatives can be considered creative
in their approaches. These include the adoption of training
courses in which care staff are put physically into residents’
shoes, and exercises to encourage participants to
experience life mentally through the eyes of someone with
dementia (Brooker 2007). Caring for people with late stage
dementia is recognized to require more creative approaches,
and a common theme is the need to deliver care specific
to each individual’s behavioral patterns and habits.
To discover the types of dementia care problem more
amenable to this model of creative problem solving, we
observed care work and interviews with carers at one UK
residential home revealed different roles for creative problem
solving in dementia care. One of these roles was to
Proceedings of the Fourth International Conference on Computational Creativity 2013 48
reduce the instances of challenging behavior in residents.
Challenging behavior defined as “culturally abnormal behavior(s)
of such an intensity, frequency or duration that
the physical safety of the person or others is likely to be
placed in serious jeopardy, or behavior which is likely to
seriously limit use of, or result in the person being denied
access to, ordinary community facilities” (Bromley and
Emerson 1995). Examples include the refusal of food or
medication, and verbal aggression.
Interviews with carers revealed that creative problem
solving has the potential to generate possible solutions to
reduce instances of challenging behavior. For example, if a
resident is uncooperative with carers when taking medication,
one means to reduce it might be to have a carer wear a
doctor’s coat when giving the medication. The means is
creative because it can be useful, novel to the resident if
not applied to him before, and novel to the care team who
have not applied it before. Therefore, with carers in the
pilot home, we explored the potential of different creativity
techniques to reduce challenging behavior.
During one half-day workshop with 6 carers we explored
the effectiveness and potential of different creativity
techniques to manage a fictional challenging behavior.
During a three-stage process the carers were presented with
the fictional resident and challenging behavior, generated
ideas to reduce the behavior, then prepared to implement
these ideas. They used different creativity techniques, presented
to them as practical problem solving techniques, to
reduce the fictional challenging behavior. The carers
demonstrated the greatest potential and appetite for the
other exploratory creativity technique, called Other Worlds
(Innovation Story 2002). During the workshop, the carers
sought to generate ideas to reduce the challenging behavior
in four different, less constrained domains - social life,
research, word of mouth and different cultures. These ideas
were then transferred to the care domain to explore their
effectiveness in it. Other Worlds was judged to be the most
effective as well as the most interesting to carers. It created
more ideas than any of the other techniques, and two of the
ideas from the session were deemed sufficiently useful to
implement in the pilot home immediately. Carers singled
out the technique because, unlike others, it purposefully
transferred knowledge and ideas via similarity-based reasoning
from sources outside of the immediate problem
spaces – the resident, residential home and dementia care
domain.
The Carer App
To implement Other Worlds in care work we decided to
develop a mobile software app, called Carer, which carers
can use during their work. In the place of human facilitation,
the software retrieves then guides carers to explore
other worlds that are retrieved by the app, and in place of
face-to-face communication, the software was to support
asynchronous communication between carers who would
digitally share information about care ideas and practices
via the software.
The Carer app accesses a digital repository to retrieve
natural language descriptions of cases of good care practice
in XML based on the structure of dementia care case studies
reported by the Social Care Institute for Excellence
(Owen and Meyer 2009) as well as challenging behavior
cases in non-care domains such as teen parenting, student
mentoring and prison life. Each case has two main parts of
up to 150 words of prose each – the situation encountered
and the care plan enhancement applied – and is attributed
to one class of domain to which the case belongs. The current
version of the repository contains 115 case descriptions.
Figure 1. The Carer mobile app showing how carers describe
challenging behaviors (on the left-hand side) and a detailed de- scription of one of these cases (on the right-hand side)
Carer app automatically retrieves the previous cases
using different services in response to natural language
entries typed and/or spoken by a carer into the app. One
supports case-based reasoning with literally similar cases
based on information retrieval techniques, similar to strategies
applied to people with chronic diseases (Houts et al.
1996). A second supports the other worlds technique more
generally by automatically generating different domains
such as traveling or cooking in which to generate care plan
enhancements to a current situation without the constraints
of the care domain (Innovation Company 2002). The user
is encouraged to think about how to solve the aggression
situation in the school playground. A simple flick of the
screen will generate a different other world, such as parachuting
from an aircraft. A third service automatically
generates creativity prompts from retrieved case content.
Lastly, the Carer app invokes AnTiQue, an analogical
reasoning discovery service that matches the description of
a challenging behavior situation to descriptions in the repository
of challenging behavior cases in non-care domains.
To do this, the service implements a computational
analogical reasoning algorithm based on the StructureProceedings
of the Fourth International Conference on Computational Creativity 2013 49
Mapping Theory (Gentner 1983; Falkenhainer et al. 1989)
with natural language parsing techniques and a domainindependent
verb lexicon called VerbNet (Kipper et al.
2000). A carer can then record new ideas resulting from
creative thinking in audio form, then reflect on them by
playing them back to change them, generate further ideas,
compose them into a care plan and share the plan with other
carers. Some of these features are depicted in Figure 1.
The right-hand side of Figure 1 shows one retrieved analogical
case description – Managing a disrespectful child –
as it is presented to a carer using the app. The Carer app is
described at length in Maiden (2012). The next section
describes two of the computational creativity services – the
analogical reasoning discovery service and the creativity
prompt generation service.
The Analogical Reasoning Discovery Service
This service (called AnTiQue) matches a description of
challenging behavior in dementia care to descriptions of
challenging behavior problems and resolutions in other
domains, for example good policing practices to manage
disorderly revelers and good teaching practices to manage
disruptive children. AnTiQue’s design seeks to solve 2
research problems: (i) match incomplete and ambiguous
natural language descriptions of challenging behaviour in
dementia care and challenging behaviour problems and
resolutions in other domains using different lexical terms;
(ii) compute complex analogical matches between descriptions
without a priori classification of the described domains.
Analogical service retrieval can increase the number of
cases that are useful to the care staff by retrieving descriptions
of cases solved successfully in other domains, for
example good policing practices to manage disorderly
revelers and good teaching practices to manage disruptive
children. The problem and solution description of each
case might have aspects that, through analogical reasoning,
can trigger discovery of new ideas on the current challenging
behaviour. For example, a description of good policing
practice to manage disorderly revellers can provide analogical
insights with which to manage challenging behaviour
in dementia care. AnTiQue seeks to leverage these new
sources of knowledge in dementia care.
Analogical retrieval in AnTiQue uses a similarity model
called the Structure Mapping Theory (SMT) (Gentner
1983) which seeks to transfer a network of related facts
rather than unrelated one (Gentner 1983) from a source to a
target domain. To enable structure-matching AnTiQue
transforms natural language queries and case descriptions
into predicates that express prepositional networks of
nodes (objects) and edges (predicate values). Attributional
predicates state properties of objects in the form PredicateValue(Object)
such as asleep(resident) and absent(relative).
Relational predicates express relations between
objects as PredicateValue (Object1,Object2) such as
abuse(resident, care-staff) and remain(resident,room).
According to the SMT, a literal similarity is a comparison
in which attributional and relational predicates can both be
mapped from a source to a target. In contrast an analogy is
a comparison in which relational predicates but few or no
attributional predicates can be mapped. Therefore AnTiQue
retrieves cases with high match scores for relational
predicates and low match scores for attributional predicates,
for example a match with the predicate
abuse(detainee,police-officer) but no match with the predicate
drunk(detainee).
AnTiQue
NLP Parser
Predicate Parser
Predicate Expansion
Predicate Matcher
Similarity
Query
Parsed
Sentences
Predicates
XQ Registry ueries
Services
Expanded
Predicates
Matched Services
VerbNet
Predicate
Rules & Heuristics
WordNet,
SSParser,
Stanford Parser
Verb Classes
WordNet,
Dependency
Thesaurus
Figure 2. Internal structure of AnTiQue
Figure 2 depicts AnTiQue’s 5 components. When invoked
the service first divides query and case problem description
text into sentences, then part-of-speech tagged,
shallow parsed to identify sentence constituents and
chunked in noun phrases. It then applies 21 syntax structure
rules and 7 lexical extraction heuristics to identify
predicates and extract lexical content in each sentence.
Natural language sentences are presented as predicates in
the form PredicateValue(Object1, Object2). The service
then expands each query predicate with additional predicate
values that have similar meaning according to verb
classes found in VerbNet to increase the likelihood of a
match with predicates describing each case. For example
the predicate value abuse is in the same verb class as attack.
The service then matches all expanded predicates to a
similar set of predicates that describe the problem description
of each case in the repository. This is achieved using
XQuery text- searching functions to discover an initial set
of cases that satisfy global search constraints. Finally it
applies semantic and dependency-based similarity
measures to refine the candidate case study set. The service
returns an ordered set of analogical cases based on the
match score with the query.
The components use WordNet, VerbNet, and the Dependency
Thesaurus to compute attributional and relational
similarities. WordNet is a lexical database inspired by psycholinguistic
theories of human lexical memory (Miller
Proceedings of the Fourth International Conference on Computational Creativity 2013 50
1993). Its word senses and definitions provide the data
with which to disambiguate terms in queries and case problem
descriptions. Its semantic relations link terms to other
terms with similar meanings with which to make service
queries more complete. For example a service query with
the term car is expanded with other terms with similar
meaning, such as automobile and vehicle, to increase
matches with web service descriptions.
VerbNet (Kipper et al. 2000) is a domain independent
verb lexicon. It organizes terms into verb classes that refine
Levin classes (Levin 1993) and add sub-classes to achieve
syntactic and semantic coherence among members of a
verb class. AnTiQue uses it to expand query predicate values
with different members from the same verb class. For
example, queries with the verb abuse are expanded with
other verbs with similar meaning such as attack.
The Dependency Thesaurus supports dependency-based
word similarity matching to detect similar words from text
corpora. Lin (1998) used a 64-million word corpus to compute
pair-wise similarities between all of the nouns, verbs,
adjectives and adverbs in the corpus using a similarity
measure. Given an input word the Dependency Thesaurus
can retrieve similar words and group them automatically
into clusters. AnTiQue used the Dependency Thesaurus to
compute the relational similarity between 2 sets of predicates.
In the remainder of this section we demonstrate the AnTiQue
components using text from the following example
challenging behaviour situation:
A resident acts aggressively towards care staff and
the resident verbally abuses other residents at
breakfast. Suspect underlying insecurities to new
people.
Natural Language Processing
This component prepares the structured natural language
(NL) service query for predicate parsing and expansion. In
the first step the text is split into sentences. In the second a
part-of-speech tagging process is applied that marks up the
words in each sentence as corresponding to a particular
lexical category (part-of-speech) using its definition and
context. In the third step the algorithm applies a NL processing
technique called shallow parsing that attempts to
provide some machine understanding of the structure of a
sentence without parsing it fully into a parsed tree form.
The output is a division of the text's sentences into a series
of words that, together, constitute a grammatical unit. In
our example the tagged sentence a resident acts aggressively
towards care staff and the resident verbally abuses
other residents at breakfast is shown in Figure 3. Tags that
follow a word with a forward slash (e.g. driver/NN) correspond
to lexical categories including noun, verb, adjective
and adverb. For example, the NN tag means “noun singular
or mass", DT means “determinant” and VBZ means “verb,
present tense, 3rd person singular”. Tags attached to each
chunk (e.g. [The/DT driver/NN]NP) correspond to phrasal
categories. For instance, the NP tag denotes a “noun
phrase”, VP a “verb phrase”, S a “simple declarative
clause”, PP a “prepositional phrase” and ADVP a “adverb
phrase”.
[A/DT resident/NN]NP [acts/VBZ]VP [aggressively/RB]ADVP
[towards/]PP [care staff/NN]NP.
Figure 3. The sentence a resident acts aggressively towards care
staff after performing part-of-speech tagging and chunking
The component then decomposes each sentence into its
phrasal categories used in the next component to identify
predicates in each sentence structure.
Predicate Parsing
This component automatically identifies predicate structures
within each annotated NL sentence based on syntax
structure rules and lexical extraction heuristics. Syntax
structure rules break down a pre-processed NL sentence
into sequences of phrasal categories where each sequence
contains 2 or more phrasal categories. Lexical extraction
heuristics are applied on each identified sequence of phrasal
categories to extract its lexical content used to generate
one or more predicates.
Firstly the algorithm applies 21 syntax structure rules.
Each rule consists of a phrasal category sequence of the
form Ri ! [Bj], meaning that the rule Ri consists of a
phrasal category sequence B1, B2,…, Bj. For example the
rule R4 ! [NP, VP, S, VP, NP] reads: rule R1 consists of a
NP followed by a VP, a S, a VP, and a NP, where NP, VP
and S mean a noun phrase, a verb phrase and a simple declarative
clause respectively. The method takes a phrasal
category list as input and returns a list containing each discovered
syntax structure rule and its starting point in the
corresponding phrasal category list, e.g. {(R1,3), (R5,1)}.
In our example, the input for the pre-processed sentence
shown in Figure 3 corresponds to a list Input = (NP, VP,
ADVP, PP, NP). Starting from the first list position the
method recursively checks whether there exists a sequence
within the phrasal category list that matches one of the
syntax structure rules. The output after applying the algorithm
on list Input is a list of only one matched syntax
structure rule, i.e. Output = {(R2,1)}.
Secondly the algorithm applies lexical extraction heuristics
on a syntax structure rule-tagged sentence to extract
content words for generating one or more predicates. For
each identified syntax structure rule in a sentence the algorithm:
(1) determines the position of both noun and verb
phrases within the phrasal category sequence; (2) applies
the heuristics to extract the content words (verbs and
nouns) from each phrase category; (3) converts each verb
and noun to its morphological root (e.g. abusing to abuse);
and (4) generates the corresponding predicate p in the form
PredicateValue(Object1, Object2) where PredicateValue is
the verb and Object1 and Object2 the nouns. To illustrate
this the algorithm identified rule R2+ for our example senProceedings
of the Fourth International Conference on Computational Creativity 2013 51
tence in Figure 3. According to one heuristic {R2+} corresponds
to the following phrasal category sequence [NP,
VP, ADVP, PP, NP]. Therefore the algorithm determines
the position of both noun and verb phrases within this sequence,
i.e. noun phrases in {NP,1} and {NP,5} and verb
phrases in {VP,2}. Lexical extraction heuristics are applied
to extract the content words from each phrase category, i.e.
{NP,1} ! resident, {NP,5} ! care staff, {VP,2} ! act.
Returning to our example, the algorithm generates two
predicates for the sentence a resident acts aggressively
towards care staff and the resident verbally abuses other
residents at breakfast, namely act(resident,care_staff) and
abuse(resident,resident).
Predicate Expansion
Predicate expansion and matching are key to the service’s
effectiveness. In AnTiQue queries are expanded using
words with similar meaning. AnTiQue uses ontological
information from VerbNet to extract semantically related
verbs for verbs in each predicate.
 VerbNet classes are organised to ensure syntactic and
semantic coherence among members, for example the verb
abuse as repeatedly treat a victim in a cruel way is one of
24 members of the judgement class. Other members
include attack, assault and insult and 20 other verbs as
potential expansions. Thus VerbNet provides 23 verbs as
potential expansions for the verb abuse. Although classes
group together verbs with similar argument structures, the
meanings of the verbs are not necessarily synonymous. For
instance, the degree of attributional similarity between
abuse and reward is very low, whereas the similarity
between abuse and assault is very high. The service
constrains use of expansion to verb members that achieve a
threshold on the degree of attributional similarity
computed with WordNet-based similarity measurements
(Simpson and Dao 2005). Given 2 sets of text, T1 and T2,
the measurement determines how similar the meaning of
T1 and T2 is scored between 0 and 1. For example, for the
verb abuse, the algorithm computes the degree of
attributional similarity between abuse and each co-member
within the judgement class. In our example verbs such as
attack, assault and insult but not honour and doubt are
used to generate additional predicates in the expanded
query.
Predicate Matching
Coarse-grained Matching The expanded query is fired at
problem descriptions of cases in the repository as an
XQuery. Prior to executing the XQuery we pre-process all
problem descriptions of cases in the registries using the
Natural Language Processing and Predicate Parsing components
and store them locally. The XQuery includes functions
to match each original and expanded predicate value
to equivalent representations of candidate problem descriptions
of cases. The service retrieves an initial set of
matched cases.
Fine-grained Matching The Predicate Matcher applies
semantic and dependency-based similarity measures to
assess the quality of the candidate case set. It computes
relational similarity between the query and each case
retrieved during coarse-grain matching. To compute
relational similarities that indicate analogical matches
between service and query predicate arguments the
Predicate Matcher uses the Dependency Thesaurus to
select web services that are relationally similar to mapped
predicates in the service query.
In our example the case Managing a disrespectful child,
which describes a good childcare practice to manage a disrespectful
child, is one candidate case retrieved during
coarse-grained matching. Figure 4 shows the problem and
solution description of the case.
Name Managing a disrespectful child
Problem An intelligent 13-year-old boy voices opinions that
are hurtful and embarrassing. The child refuses to
consider the views of others and often makes dis- criminatory statements. The parents have removed
his privileges and threatened to take him out of the
school he loves. This approach has not worked. He
now makes hurtful comments to his mother about
her appearance. The child insults neighbours and
guests at their home. He is rude and mimics their
behaviour. The child shows no remorse for his
actions. His mother is at the end of her tether.
Solution The son needs very clear boundaries set. The par- ents are going to set clear rules on acceptable be- haviour. They will state what they are not prepared
to tolerate. They will highlight rude comments in a
firm tone with the boy. He will receive an explana- tion as to why the comments are hurtful. Both par- ents will agree punishments for rule breaking that
are realistic. They will work as a team and follow
through on punishments. The son can then regain
his privileges as rewards for consistent good be- haviours.
Figure 4. A retrieved case describing a good childcare practice to
manage a disrespectful child
The algorithm receives as inputs a pre-processed sentence
list for the query and problem description of the case.
It compares each predicate in the pre-processed query sentence
list Pred(j)Query with each predicate in the preprocessed
problem description sentence list Pred(k)Case to
calculate the relevant match value, where
Pred(j)Query = PredValQuery(Arg1Query; Arg2Query)
and
Pred(k)Case = PredValCase (Arg1Case; Arg2Case).
The following conditions must be met in order to accept
a match between the predicate pair:
1. PredValCase exists in list of expanded predicate values of
PredValQuery;
2. Arg1Query and Arg1Case (or Arg2Query and Arg2Case respectively)
are not the same;
3. Arg1Case (or Arg2Case) exists in the Dependency Thesaurus
result set when using Arg1Query (or Arg2Query) as the
query to the Thesaurus;
4. the resulting attributional similarity value from step 3 is
below a specified threshold.
Proceedings of the Fourth International Conference on Computational Creativity 2013 52
If all conditions are met, PredCase is added to the list of
matched predicates for the current case. If not the algorithm
rejects PredCase and considers the next list item.
AnTiQue queries the Dependency Thesaurus to retrieve
a list of dependent terms. Terms are grouped automatically
according to their dependency-based similarity degree.
Firstly the algorithm checks whether the case predicate
argument exists in this list. If so, it uses the semantic similarity
component to further refine and assess the quality of
the case predicate with regards to relational similarity.
Using this 2-step process AnTiQue returns an ordered
set of analogical cases based on the match score with the
query. In our example consider Pred(j)Query =
abuse(resident,residents) extracted from the sentence the
resident verbally abuses other residents at breakfast, and
the Pred(k)Case = insult(child,neighbours) from the sentence
The child insults neighbours and guests at their home taken
from the description of the Managing a disrespectful
child good childcare practice case in Figure 4. In this example
all conditions for an analogical match are met: the
predicate values abuse and insult are semantically equivalent
whilst the object names resident and child and residents
and neighbours are not the same. According to the
Dependency thesaurus child is similar based on dependencies
to resident, and neighbour is similar based on dependencies
to resident. Finally the attributional similarity value
of resident and child is 0.33, for resident and neighbour
0.25 – both below the specified threshold. As a result the
predicate insult(child,neighbours) is added to the list of
matched predicates for the predicate
abuse(resident,resident).
At the end of each invocation, the service returns an ordered
set of the descriptions of the highest-scoring cases
for the app component to display to the care staff.
The Creativity Trigger Generation Service
Although care staff can generate new resolutions directly
from retrieved case descriptions, formative usability testing
with the app revealed that users were often overwhelmed
by the volume of text describing each case and uncertain
how to start idea generation. Therefore we developed an
automated service that care staff can invoke to generate
creative triggers that extract content from the retrieved
descriptions to conjecture new ideas that care staff can
consider for the resident. Each trigger expresses a single
idea that care staff can use to initiate creative thinking. The
service uses the attributional predicates generated by the
analogical matching discovery service to generate prompts
that encourage analogical transfer of knowledge using the
object-pair mappings identified in each predicate. It has the
form Think about a new idea based on the, followed by
mapped subject and object names in the target domain. To
illustate, referring back to the Managing a disrespectful
child good practice case retrieved from the childcare domain
shown in Figure 1, Figure 5 shows how they are presented
in the Carer mobile app while Figure 6 lists all creativity
prompts that the service generates for the analogical
case.
Figure 5. The Carer mobile app showing creativity prompts generated
for the Managing a disrespectful child case
Think about a new idea based on the boundaries
Think about a new idea based on the clear rules
Think about a new idea based on the acceptable behaviour
Think about a new idea based on the rude comments
Think about a new idea based on the firm tone
Think about a new idea based on the explanation
Think about a new idea based on the comments
Think about a new idea based on the punishment
Think about a new idea based on the rule breaking
Think about a new idea based on the rewards
Think about a new idea based on the privileges
Think about new idea based on the consistent good behaviour
Figure 6. Creativity prompts generated for the Managing a disrespectful
child case
Discovering Novel Ideas
Our design of the Carer app builds on Kerne et al. (2008)’s
notion of human-centered creative cognition, in which information
gathering and idea discovery occur concurrently,
and information search and idea generation reinforce each
other. The computational model of analogical reasoning
searches for and retrieves information from analogical domains,
and the creativity trigger generation service manipulates
this information to support more effective idea
generation from information, however the generation of
new ideas remains a human cognitive activity undertaken
by carers, supported by bespoke features implemented in
the app.
For example, a carer can audio-record a new idea at any
time in response to retrieved analogical cases and/or presented
creativity triggers by pressing the red button visible
in Figures 1 then verbalizing and naming the idea. Recorded
ideas can be selected and ordered to construct a new
care enhancement plan that can be extended with more
Proceedings of the Fourth International Conference on Computational Creativity 2013 53
ideas and comments at any time. The carer can also play
back the audio-recorded ideas and care enhancement plans
to reflect and learn about them, inspired by similar use of
the audio channel in digitally supported creative brainstorming
(van Dijk et al. 2011). Reflection about an idea is
supported with guidance from the app to reflect on why the
idea is needed, what the idea achieved, and how and when
the idea should be implemented. Reflection about a care
enhancement plan is more sophisticated. A carer can dragand-drop
ideas in and out of the plan and into different
sequences in it. Then, during play back of the plan, the app
concatenates the individual idea audio files and plays the
plan as a single recording, allowing the carer to listen to
and reflect on each version of the plan as a different narrated
story. Moreover, s/he can reflect collaboratively with
colleagues using the app to share the plan as e-mail attachments,
thereby enabling asynchronous communication
between carers.
Formative Evaluation of the Carer App
The Carer app was made available for evaluation over
prolonged periods with carers in a residential home. At the
start of the evaluation, 7 nurses and care staff in the residential
home were given an iPod Touch for their individual
use during their care work over a continuous 28-day period.
All 7 carers received face-to-face training in how to use
the device and both apps before the evaluation started. A
half-day workshop was held at the residential home to allow
them to experiment with all of both apps’ features. The
carers were also given training and practice with the 3
forms of Other Worlds creativity technique through practice
and facilitation to demonstrate how it can lead to idea
generation. We deemed this training in the creativity technique
an essential precondition for successful uptake of the
app.
 Even though it only lasted 4 weeks, the reported evaluation
of the Carer app in one residential home provided
valuable data about the use of mobile computing and creativity
techniques in dementia care. Figure 7 depicts the
results.
Residential
cases
Analogical
domain cases
Ideas
generated
Enhancement
plans generated
Totals 27 5 14 10
Figure 7. Situations, ideas and care enhancement plans generated
by care staff using Carer app
The focus group revealed that the nurses and carers implemented
at least one major change to the care of one resident
based on ideas generated using the app.
However, most of this success was not based on the analogical
cases retrieved by the computational model. Whilst
carers using the app did use the analogical matching service,
and the service did retrieve relevant cases from analogical
domains such as childcare and student management,
the carers were unable to map and transfer
knowledge from each of these source domains to the current
dementia-related challenging behavior. The log data
recorded only 5 uses of the analogical reasoning service to
retrieve descriptions of cases of challenging behaviors
from non-care domains. Rather, the carers appeared to use
the case-based reasoning service to retrieve descriptions of
challenging behavior cases from the care domain – the log
data recorded 28 uses of this service, and most of the 114
recorded uses of the creativity prompt generation service
were generated from these same-domain dementia cases.
The focus group revealed that the carers did not use retrieved
non-care domain cases because they were unable to
recognize analogical similarities between them and the
challenging behavior situation. We identified two possible
reasons for this. Firstly, AnTiQue implements an approach
that approximates analogical retrieval, hence there is always
the possibility of computing seemingly “wrong” associations
and retrieve cases that do not have analogical
similarities. Previous evaluations of AnTiQue with regards
to the precision and recall (Zachos & Maiden, 2008) revealed
a recall score of 100% and a precision score of
66,6% highlighting one potential limitation of computing
the attributional similarity using WordNet-based similarity
measures.
Secondly, the results suggests that carers will require
more interactive support based on results generated by the
computational model to support cognitive analogical reasoning,
consistent with previously reported empirical findings
(e.g. Gick 1983). Examples of such increased interactive
support include explicitly reporting each computed
analogical mapping to the carer, use of graphical depictions
of structured knowledge to transfer from the source to
the target domain, and more deliberate analogical support
prompts, for example based on the form A is to B as C is to
D. We are extending Carer app with such features and look
forward to reporting these extensions in the near future.
Related Work
Since the 1980s, the efforts of many Artificial Intelligence
researchers and psychologists have contributed to an
emerging agreement on many issues relating to analogical
reasoning. In various ways and with differing emphases, all
current computational analogical reasoning techniques use
underlying structural information about the sources and the
target domains to derive analogies. However, at the algorithmic
level, they achieve the computation in many different
ways (Keane et al. 1994).
Based on the Structure Mapping Theory (SMT), Gentner
constructed a computer model of this theory called Structure
Mapping Engine (SME) (Gentner 1989). The method
assumes that both target and source situations are represented
using a certain symbolic representation. The SME
also only uses syntactic structures about the two situations
as the main input knowledge — it has no knowledge of any
kind of semantic similarity between various descriptions
and relations in the two situations. All processing is based
on syntactic structural features of the two given representations.

The application of analogical reasoning to software reuse
is not new. For example, Massonet and van
Proceedings of the Fourth International Conference on Computational Creativity 2013 54
Lamsweerde (1997) applied analogy-making techniques to
complete partial requirements specifications using a rich,
well-structured ontology combined with formal assertions.
The method was based on query generalization for completing
specifications. The absence of effective ontologies
and taxonomies would expose the weaknesses of the proposed
approach due to the reliance on ontologies. Pisan
(2000) tried to overcome this weakness by applying the
SME to expand semi-formal specifications. The idea was
to find mappings from specifications for problems similar
to the one in hand and use the mappings to adapt an existing
specification without requiring domain specific
knowledge. The research presented in this paper overcomes
limitations of the above-mentioned approaches by
using additional knowledge bases to extent the mapping
process with semantic similarity measures.
Conclusion and Future Work
This paper reports a practical application of a computational
model of analogical reasoning to a pressing social problem,
which is to improve the care of older people with dementia.
The result is a mobile app that is capable technically
of accepting spoken and typed natural language input
and retrieving analogical domain cases that can be presented
with creativity triggers to support analogical problem
solving.
The evaluation results reported revealed that our model
of creative problem solving in dementia care did not describe
all observed carer behavior, so we are currently repeating
the rollout and evaluation of Carer in other residential
homes to validate this finding. Carer is being extended
with new creativity support features that include
web images that match generated creativity prompts, and
more explicit support for analogical reuse of cases from
non-dementia care domains. We are extending the repository
with new cases that are semantically closer to dementia
care and, therefore, easier to recognize analogical similarities
with.
Acknowledgment
The research reported in this paper is supported by the EUfunded
MIRROR integrated project 257617, 2010-14.
<references_biblio/>
References
Alzheimers Society, 2010. Statistics.
http://www.alzheimers.org.uk/site/scripts/documents_info.p
hp?documentID=341
Wimo A. & Prince M., 2010, ‘World Alzheimer Report 2010: The Global
Economic impact of Dementia,
http://www.alz.co.uk/research/worldreport/
Brooker, D., 2007, ‘Person-centred dementia care: Making Services Better,
Bradford Dementia Group Good Practice Guides. Jessica Kingsley
Publishers London and Philadephia
Osborn A., 1965. The Creative Trend in Education. In: Source Book For
Creative Problem Solving: A Fifty Year Digest of Proven Innovation
Processes. Creative Education Foundation Press, New York.
Le Storti A., J., Cullen P., A., Hanzlik E., M., Michiels, J., M., Piano L.,
A., Lawless Ryan P., Johnson W., 1999. Creative Thinking In Nursing
Education: Preparing for Tomorrow's Challenges. Nursing Outlook. Vol.
47, no. 2 62–66.
Help The Aged, 2007, ‘My Home Life: Quality of Life In Care Homes; A
Review of the Literature London. [online]
http://myhomelifemovement.org/ downloads/mhl_review.pdf [Accessed 5
Jan 2011].
Bromley J. and Emerson E., 1995, ‘Beliefs and Emotional Reactions of
Care Staff working with People with Challenging Behaviour’, Journal of
Intellectual Disability Research 39(4), 341-352
Innovation Company, 2002. ‘Sticky Wisdom (?WhatIf!)’. Capstone Publishing
Company Limited, Chichester.
Owen T. and Meyer J., 2009, ‘Minimizing the use of ‘restraint’ in care
homes: Challenges, dilemmas and positive approaches’, Adult Services
Report 25, Social Care Institute of Excellence, http://www.scie.org.uk/
publications/reports/report25.pdf.
Houts, P.S., Nezubd, A.M., Magut Nezubd, C, Bucherc, J.A., 1996, ‘The
prepared family caregiver: a problem- solving approach to family caregiver
education’, Patient Education and Counselling, 27,1, 63-73.
Gentner D., 1983, 'Structure-Mapping: A Theoretical Framework for
Analogy', Cognitive Science 5, 121-152.
Falkenhainer B., Forbus K.D. & Gentner D., 1989, 'The StructureMapping
Engine: Algorithm and Examples', Artificial Intelligence 41, 1-
63.
Kerne, A., Koh, E., Smith, S. M., Webb, A., Dworaczyk, B., 2008,
‘combinFormation: Mixed-Initiative Composition of Image and Text
Surrogates Promotes Information Discovery’, ACM Transactions on
Information Systems, 27(1), 1-45.
Kipper K., Dang H.T. and Palmer M., 2000, ‘Class-based Construction of
a Verb Lexicon’, Proceedings AAAI/IAAI Conference 2000, 691–696.
Miller K., 1993, ‘Introduction to WordNet: an On-line Lexical Database’
Distributed with WordNet software.
Levin B., 1993, ‘English Verb Classes and Alternations: A Preliminary
Investigation’, University Chicago Press.
Lin D., 1998, ‘Automatic retrieval and clustering of similar words’, In
COLINGACL, 768-774.
Maiden N.A.M., 2012, ‘D5.2 Deliverable: Techniques and Software Apps
for Integrated Creative Problem Solving and Reflective Learning Version
1’, Technical Report. Available at http://www.mirrorproject.eu/showroom-a-publications/deliverables.
Simpson, T. and Dao, T. (2005). Wordnet-based semantic similarity
measurement. codeproject.com/cs/library/semanticsimilaritywordnet.asp
van Dijk J., van der Roest J., van der Lugt R. & Overbeeke K., ‘NOOT: A
Tool for Sharing Moments of Reflection during Creative Meetings’, Proceedings
10th ACM Creativity and Cognition Conference, Atlanta Georgia,
Nov 2011, ACM Press.
Gick M.L., 1989, ‘Two Functions of Diagrams in Problem Solving by
Analogy’, Knowledge Acquisition from Text and Pictures’, ed. M. Handi
& J.R. Levin, Elsevier Publishers B.V. North-Holland, 215-231.
Keane, M.T., Ledgeway, T. & Duff, S. (1994). Constraints on analogical
map- ping: A comparison of three models. Cognitive Science, 18, 387–
438.
Massonet, P. and van Lamsweerde, A. (1997). Analogical reuse of requirements
framework. 3rd IEEE International Symposium on Requirements
Engineering.
Pisan, Y. (2000). Extending requirement specifications using analogy.
22nd International Conference on Software Engineering (ICSE), limerick,
Ireland.
Zachos K. & Maiden N.A.M., 2008, ‘Inventing Requirements from Software:
An Empirical Investigation with Web Services’, Proceedings 16th
IEEE International Conference on Requirements Engineering, IEEE
Computer Society Press, 145-154.
Proceedings of the Fourth International Conference on Computational Creativity 2013 55