Psychological plausibility of the ¡¡insert name¿¿ model.

Niels Ruben Reinders

Januari 2011

1 Modeling cognitive sub-processes of analogy making

The goal of this model is to perform analogical reasoning in the four-word task and not analogical reason-
ing in general. However, this model accounts for several cognitive phenomena, often neglected by other
models. One norm for psychological plausibility could be that all sub-processes of analogical reasoning
are incorporated in the model. Analogical reasoning is divided in several subprocesses: 1.) Representa-
tion building 2.) retrieval of one case given another 3.) mapping between two cases (between ’base’ and
’target’) in working memory 4.) transfer of knowledge/relations and 5.) evaluating the matching of trans-
ferred knowledge/relation on the target domain. As of yet, all sub-processes have been modeled, but no
model has been proposed that includes every sub-process, although many attempts have been made (Koki-
nov & French, 2003). The following sections will emphasize on how the . . . ........ model incorporates the
sub-processes of analogical reasoning.
1. Representation building
Building a representation of the input stimuli (i.e. a word in the 4 word pair) is a sub-process of analogy
that is absent in most former models of analogy. Some models like ANALOGY , Copycat, Tabletop of
Metacat (Kokinov & French, 2003) do build some representation from relatively unprocessed information.
The . . . .............. model has a word or pairs of words as input. Because it uses pre-processed input, this
model falls somewhat short in modelling the representation building process. The only (but effectively
increasing performance of LRA (Turney, 2006)), processing of the input is the inclusion of alternates for
words in the word pair. This way the input word is represented by its meaning instead of just a sequence
of letters, adding to the performance and psychological plausibility of the model.
2. Retrieval
Retrieval from long term memory for analogical thinking is a frequently studied process. In our model,
the retrieval step would be aquiring the relations, and ranking them according to the frequency of occurence
(saliency). Many factors can affect the retrieved content of base and target from long term memory. A
frequently studied factor is being primed to concepts, categories or relations. Being primed to a concept
(Spellman et al., 2001), categories (Green et al., 2006), or even relations (Blanchette & Dunbar, 2002)
increases the likelihood that elements related those concepts, categories or relations are retrieved from
the memory during analogy tasks. This phenomena indicates that retrieval is a very selective, heuristic
search though LTM, unlike an exhaustive search through the whole LTM. It is even empirically supported
in problem solving analogy experiments that not the whole LTM is searched (Gick & Holyoak, 1980). This
model adopts the LRA to retrieve the relation between two words to use as a base and target for determining
the analogy between word pairs. Being an extension of LSA, this is still a computationally demanding
process. However, the dynamicity that is incorporated in the process does allow priming towards relations
from base (R of first word pair) in the retrieval of the target relations (R of second word pair). ***Continue
when model is finished*** This is not possible with the LRA from (Turney, 2006) and likely resembles
human cognitive processes more closely (Green et al., 2006; Spellman et al., 2001; Blanchette & Dunbar,
2002). Another factor, which no model (including ours) accounts for, is the attention given to a certain
element of the base or target. i.e. emotional value of certain elements likely affects the saliency of the
retrieved elements. For example: Imagine a subject who just lost her parents in a car accident and is
presented with a story telling how: ’Johns partner was killed in a plain crash because of a bird in the
engine’ and a story about ’How Sophies mother died riding a bike into a moose’. When asked to name
the analogies in order of saliency, the individual might score the relation, ’losing loved ones’ more salient
than the relation of, ’traffic accident because of a animal’. This example uses trauma, but ethical stance,
morals, or beliefs could affect the saliency of relations as well. Unfortunately, presence of this phenomena
is hard to support empirically, but is likely present in retrieval during analogy making. However, all models
based on LSA/RSA are frequency based, meaning that the saliency of elements is solely dependent on the
number of (co-)occurrences in LTM (the corpora of text used). The saliency of elements in human LTM
unlikely depend solely on frequency of occurrence (Landauer, 1999). We do not claim that LRA is not
psychological plausible in the sense that the process does not mimick any cognitive process, but merely
that it clearly misses aspects of it. There is one significant advantage of LRA, or more specifically, the
use of large corpora of text to model LTM. The choice of text included as corpora can resemble individual
differences in background. E.g., when solving an African analogy task, the use of corpus of text from
African origin could prove more effective than the use of a corpus of English origin (taken that they are
translated to the same language). This would mimick cultural differences in humans. Another option would
be changing the topics included in the corpus. Focusing the content of the corpus on a specific researchfield
might mimick the task performance of researchers in that field better than that of researchers in other fields.
This would mimick differences in knowledge background, that have been observed in analogical reasoning
(Novick, 1988). These examples (especially the latter one) are more easily to support empirically. To
sum up, our model has increased psychological plausibility of the retrieval process because it allows for
individual differences (corpora content) and models the priming phenomena to some extend (dynamicity).
However, it still does not model heuristic retrieval from LTM, or the emotional value of elements retrieved
from LTM.
3.Mapping
This process is known as the core of analogy making. This process determines what elements of the
base matches those of the target. The difficulty to model this lies in the many ways that elements can be
mapped from base to target. When presented with an analogy task involving a man in a blue shirt giving a
present to a women in a red shirt, next to a doctor in a red coat giving a pill to a patient in a blue shirt, differ-
ent element may be mapped. We might engage in a mapping of the blue color and red color of clothing, a
seemingly superficial mapping of similar elements between the two domains, but appropriate nonetheless.
Or we could map the similar giver-receiver relationships between the two domains the, ignoring the similar
element of the color of clothes. Since analogies can be made in minutes of even seconds, mapping vast
amount of elements unlikely takes place in humans, it seems only the most salient elements are mapped
during analogy making. Just as in the retrieval process, the attention given to certain elements (like with
emotional value) affects what elements seem most salient and will firstly or only be mapped between base
and target. Interestingly, patients of psychological disorders can give more or less attention towards certain
types of elements, like attributable or relational elements, during analogy making (Simpson & Done, 2004;
Morsanyi & Holyoak, 2010; Reed, 1996). The................model shows reasonable psychological plausibil-
ity in light of these observations from cognitive science. Some individual variation in choosing the most
salient element for mapping can me modeled by changing the content of corpora of text used, since that
could affect the distribution and ranking of elements. With respect to the mapping of only the most salient
elements, our model has a great advantage, by utilizing the distributional properties of a vector based model.
Turney’s LRA handeles 4000 relations per analogy Turney (2006), a more psychologically plausible model
should map more selectively, to reduce working memory load. Our model includes a dynamic mapping
process. This is less computationally demanding, maps only the most salient relations from the matrix, but
does take in account more than those by increasing the saliency from elements that are similar between
base and target. Unfortunately, this process has the same limitation of the LRA in that salience is deter-
mined solely by frequency and not emotional value. Also, it does not account for another factor present
in cognitive processes that does seem to make mapping less computationally demanding, namely, higher
order predicates (or relations) (Gentner, 1983). Relations are not processed as a collection of independent
facts, like in the LRA vector space, but rather as connected or interrelated relations (Gentner, 1983). A
computational consequence is that the number of relations do not linearly increase computational demand,
like in our model. Gentner goes as fas as to say that higher order relations also increase the saliency of the
relation. For more detailed explanations and examples, read Gentner (1983). In short, the . . . ........... model
increases psychological plausibility in the mapping process because the incorporated dynamicity models
some sort of priming towards elements used in analogy. Also, the selection of most salient elements for

2
mapping reduce computational demands and increase psychological plausibility. However, several aspects
of priming and processing elements seem far from cognitive processes still.
4. Transfer
The transfer of elements (or ’relations’ in the current model) facilitates the learning from analogies.
This process takes place during mapping. Learning takes place when new knowledge is inserted into the
target domain. The model of relational similarity (Turney, 2006) can not capture this process, but the
addition of dynamicity seems to do so. In our model of the four word task, the dynamic mapping process
allows the word pairs to be both base and target, allowing to insert/change the knowledge about both by
changing the saliency of the relations present in a word pair. This process constitutes a critical assumption:
all possible relations in base and target are somewhat present in LTM, just not actively processed. There
is no process that tries to match a relation from the base that is not yet present in the vector of the target.
***continue when model is finished***.
5.Evaluation
The evaluation determines the likelihood that the mapped elements (or the transferred elements) from
the base are applicable (or ’match’) the target. This process is implicitly taking place in mapping and
transfer. In our model the match determined by the presence of equal mapped relations in the target.
***continue when model is finished***.

2 Filters
In the LRA used there are many parameters set, to keep the computational demands limited or reduce noise.
Parameters include the number of synonyms to create alternate wordpairs (num-sin), number of most fre-
quent alternate word pairs used for LRA (num-filter), maximun number of words per phrase between the
words of the wordpair (max-phrase), maximum number of different phrases (or ’patterns) included in the
matrix (max-patterns). In order to asses the psychological plausibility of the model, it is important to ask
what might be the psychological correlate of these parameters? For a direct comparison of parameters
with human cognitive processes, crucial knowledge about the computations in human cognition lacks. But
considering the distinct nature of a computational model, even that might not suffice. Also, the parameters
are not based on human performance which would speak against their psychological plausibility. However,
most parameters of the model filter out data that is not important for further computation, and it is a given
that human cognitive processes, heuristic in nature, don not use all information available for computation
either. It seems such parameters might be as important to humans as to computers to reach faster perfor-
mance. It must be noted however, that the flexibility of such parameters in humans to match situational
demands is unique and never modeled since computers don’t have changing situational demands. Another
parameter setting in humans (independent of situational demands) would be cognitive capacity, like pro-
cess speed or working memory capacity (Holyoak et al., 1999). Such parameters could explain variation
in analogical thinking capacity in humans. Besides cognitive capacity, individual differences in analogical
reasoning strategies could be a product of other parameter setting that arise from genes, education, culture
(Nisbett & Masuda, 2003; Ji et al., 2000) or mental state (Gutchess et al., 2010). So whether the parameters
are psychologically plausible in that they reflect human cognitive capacity remains debatable, but there are
parameters in human cognition, so the inclusion of parameters itself, seems psychologically plausible.

References
Blanchette, I., & Dunbar, K. (2002). Representational change and analogy: How analogical inferences alter
target representations. Journal of Experimental Psychology-Learning Memory and Cognition, 28(4),
672–685.

Gentner, D. (1983). Structure-mapping - a theoretical framework for analogy. Cognitive Science, 7(2),
155–170.

Gick, M. L., & Holyoak, K. J. (1980). Analogical problem-solving. Cognitive psychology, 12(3), 306–355.

3
Green, A. E., Fugelsang, J. A., & Dunbar, K. N. (2006). Automatic activation of categorical and abstract
analogical relations in analogical reasoning. Memory and cognition, 34(7), 1414–1421.
Gutchess, A. H., Hedden, T., Ketay, S., Aron, A., & Gabrieli, J. D. E. (2010). Neural differences in
the processing of semantic relationships across cultures. Social Cognitive and Affective Neuroscience,
5(2-3), 254–263.
Holyoak, K. J., Waltz, J. A., Tohill, J. M., Lau, A., & Grewal, S. K. (1999). Restricting working-memory
capacity impairs relational mapping. MAHWAH; 10 INDUSTRIAL AVE, MAHWAH, NJ 07430 USA:
LAWRENCE ERLBAUM ASSOC PUBL.

Ji, L. J., Peng, K. P., & Nisbett, R. E. (2000). Culture, control, and perception of relationships in the
environment. Journal of personality and social psychology, 78(5), 943–955.

Kokinov, B., & French, R. M. (2003). 0rw1s34rfesdcfkexd09rt0computational mod-

els 1rw1s34rfesdcfkexd09rt0of 0rw1s34rfesdcfkexd09rt0analogy1rw1s34rfesdcfkexd09rt0-
0rw1s34rfesdcfkexd09rt0making1rw1s34rfesdcfkexd09rt0. 1, 113–118.
Landauer, T. K. (1999). Latent semantic analysis: A theory of the psychology of language and mind.
Discourse Processes, 27(3), 303–310.
Morsanyi, K., & Holyoak, K. J. (2010). Analogical reasoning ability in autistic and typically developing
children. Developmental Science, 13(4), 578–587.
Nisbett, R. E., & Masuda, T. (2003). Culture and point of view. Proceedings of the National Academy of
Sciences of the United States of America, 100(19), 11163–11170.
Novick, L. R. (1988). Analogical transfer, problem similarity, and expertise. Journal of Experimental
Psychology-Learning Memory and Cognition, 14(3), 510–520.
Reed, T. (1996). Analogical reasoning in subjects with autism, retardation, and normal development.
Journal of Developmental and Physical Disabilities, 8(1), 61–76.
Simpson, J., & Done, D. J. (2004). Analogical reasoning in schizophrenic delusions. European Psychiatry,
19(6), 344–348.
Spellman, B. A., Holyoak, K. J., & Morrison, R. G. (2001). Analogical priming via semantic relations.
Memory and cognition, 29(3), 383–393.

Turney, P. D. (2006). Similarity of semantic relations. Computational Linguistics, 32(3), 379–416.