• No results found

Foraging trough semantic memory : a replication with dutch verbal fluency data

N/A
N/A
Protected

Academic year: 2021

Share "Foraging trough semantic memory : a replication with dutch verbal fluency data"

Copied!
32
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

Foraging Trough Semantic Memory: A Replication with Dutch Verbal Fluency Data

Leonie Poelstra, 5870518

(2)

Abstract

Semantic memory is the subsystem of human memory that stores conceptual information. Verbal fluency tasks with semantic cues are widely used in neuropsychological studies to measure the performance of semantic memory. Recently, researchers found that the retrieval process from semantic memory on a verbal fluency task resembles animals’

foraging behavior (Hills et al. 2012). When an animal forages for food, it switches between patches of food in their environment. The current study investigated whether the same search process is found in humans searching for animal names in semantic memory. Participants (N = 323) filled in a two- minute online verbal fluency test. We studied whether participants switched from a local similarity based cue (e.g., "cat") to a global frequency based cue (e.g., "animals") to find a new cluster of items in semantic memory. Next we looked whether switches from local to global cues occurred when a local patch was almost depleted, and if this switching occurred when the local patch intake rate dropped to the mean overall intake rate. Results supported a clustered structure of words in semantic memory. Participants switched from a local to a global cue when the local patch was almost depleted and switched from a local patch to global search when the intake rate dropped because of patch depletion. The study was a replication of the study by Hills et al. (2012) and showed that the search in semantic memory resembles the search of animals foraging for food in their environment.

(3)

Foraging Trough Semantic Memory: A Replication with Dutch Verbal Fluency Data Introduction

Semantic memory is a subsystem of human memory which stores factual and conceptual knowledge. Contrary to episodic memory, which stores life events, semantic memory stores (abstract) concepts not linked to time and/or place (Goni et al., 2011). These concepts are stored permanently in long-term storage (LTS), and forgetting is a result of retrieval failure (Raaijmakers & Shiffrin, 1981). To investigate the effect of neurological disorders, researchers and clinical practitioners widely use the verbal fluency task to investigate the performance of semantic memory (Raoux, Amieva, Le Goff, Letenneur, & Dartigues, 2008; Goni et al., 2011). In this task, participants are asked to give as many responses as possible within a certain time- interval (e.g., 60 s), where recalled items belong to a specified category such as "animals" or "vegetables" (Troyer, Moscovitch, & Winocur, 1997). Previous studies using semantic fluency tasks show a two-stage retrieval process: 1) clustering, in which produced items form subcategories, and 2) switching, which is the transition from one subcategory to another (Troyer et al., 1997). The patch- like structure of memory recall from semantic categories has been found in several studies, Gruenewald Lockhead (1980) refer to these semantic categories as "semantic fields".

In the past, researchers investigated whether people search for cognitive resources analogous to animals foraging for food. Food resources are distributed in spatial patches, e.g. berries on separate bushes, where in human memory cognitive resources are words in semantic fields (W. Bousfield & Sedgewick, 1944; Raaijmakers & Shiffrin, 1981; Romney, Brewer, & Batchelder, 1993). When humans search for items in memory, we produce clusters of related items and transition between clusters (W. Bousfield & Sedgewick, 1944; Hills, Jones, & Todd, 2012; Raaijmakers & Shiffrin, 1981; Troyer et al., 1997). According to the classic model of optimal foraging theory (Charnov, 1976), the forager leaves a patch when the item return rate becomes lower than the average item return rate. This results in an optimal foraging strategy in which the overall item return rate is optimized. Hills et al.

(4)

(2012) found evidence for an optimal foraging strategy in participants performing a semantic verbal fluency task. The authors found that the performance of participants in terms of how many items they recalled depended both on the moment of the transition between semantic fields and the speed with which these transitions occur.

Following Hills et al. (2012), the current study compares models of semantic memory to the classic model of optimal foraging according to Charnov (1976), but now using Dutch verbal fluency data. Hills et al. (2012) used two sources to represent the semantic space searched in the verbal fluency task: 1) hand-coded categorizations from Troyer et al. (1997) and 2) lexical semantic representations from a corpus- based semantic space model called BEAGLE (bound encoding of the aggregate language environment) (Jones &

Mewhort, 2007). To compare results with the English-language based BEAGLE similarity, we used a third source to represent semantic space: 3) snaut, which gives the semantic distances between dutch words based on a Dutch language corpus(Mandera, Keuleers, & Brysbaert, 2017). Using the above stated framework, we can investigate whether the semantic search process is defined by both local exploitation of semantic fields and global switching between semantic fields, and whether the optimal foraging model predicts participant’s global transitions.

Models of Semantic Memory

In the literature many examples can be found of a patch- like semantic memory structure: both in lexical decision tasks as previous studies of free recall from natural categories (e.g. "animals" or "vegetables"), researchers found clustered recall of related items (W. Bousfield & Sedgewick, 1944; Johnson, Johnson, & Mark, 1951). Groups of words (semantic fields) that are semantically similar, tend to be produced together by participants (W. A. Bousfield & Barclay, 1950; Gruenewald & Lockhead, 1980; Howard, Jing, Addis, & Kahana, 2007; Romney et al., 1993). Different models of semantic memory retrieval have incorporated a cognitive foraging process that switches between local and

(5)

global cues, where local cues produce clusters of words and global cues produce transitions between clusters of words (Gronlund & Shiffrin, 1986; Metcalfe & Murdock, 1981).

A common memory model of associative search is the SAM (Search of Associative Memory) model (Raaijmakers & Shiffrin, 1980). In SAM, cues probe memory which leads to the activation and retrieval of objects in memory. The activated cues make up the "memory probe", which changes during memory retrieval after new objects are retrieved. SAM proposes a quantitative ratio rule that determines the probabilities of sampling each object in memory, given any number of cues in the probe (Raaijmakers & Shiffrin, 1981). The probe consists of a global retrieval cue (e.g. "animals") and after item recovery the probe is modified to include the most recently recovered animal name as cue (e.g. "cat"). This new activation leads to a higher probability of retrieving items that are similar to "cat", like "dog" and "mouse". The retrieval of items that are semantically similar can be seen as local exploitation. When a local patch of items is depleted, for example farm animals: CHICKEN - HORSE - COW - SHEEP - GOAT, an individual will switch back to the global cue "animals" and start with a new associated animal name in a new local patch (e.g. "snake", category = reptiles). In this manner, there is both local exploitation and global transitioning in semantic memory. Cue- dependence is essential in SAM for the retrieval of animal names from memory. The strength of associative relationships between probe cues and so- called "memory images" is the basis for retrieval from memory. These relationships are described in a matrix of retrieval strengths from each possible cue to each possible object in memory (Raaijmakers & Shiffrin, 1981).

Another theory of cognitive patch- like structure in semantic memory is the cluster-switching hypothesis formulated by Troyer et al. (1997). Researchers in the past found that words are produced in semantic subcategories and that participants switch from subcategory to another (Troyer et al., 1997; Troyer, Moscovitch, Winocur, Leach, & Freedman, 1998). Instead of a formal association between cues and words, the authors defined categories in which clusters of items belong. For example, "chicken" belongs in the

(6)

category "birds" and "farm animals". Hills et al. (2012) used the Troyer categories to define patch- boundaries. They found that the search through semantic memory can be described as a switch- process between local and global exploration of semantic clusters in the same way that animals forage for food. They found evidence for the local patches in memory, patch depletion before switching to global information and optimal timing of patch

departures. Last, they found that participants whose strategy was better described by the optimal foraging model produced more items.

Optimal Foraging Theory

Stephens Krebs (1987) proposed foraging theory, which states that an animal’s search for food is optimal considering the conditions of the environment. These conditions are comparable with the structure of semantic memory: density of food within a food patch (local cues) and the distribution of food patches in a region (global cues) (Davelaar, 2015). A well- known model that describes an animal’s foraging behaviour is Charnov’s model of marginal value theorem (1976). This model is described as a “patch” model, which premises that (a) food is distributed in patches over a larger region of patches, (b) food depletes when an animal is foraging and (c) the goal of the animal is to maximize the gain per unit of foraging time (Davelaar, 2015). The model is a trade-off between the exploitation of a current food patch and leaving this patch to search for richer food

patches. The marginal value theorem states that an animal leaves a patch when the return rate within the patch becomes smaller than the long-term average rate of return. In formal terms, the model is stated as follows:

R = g(tw) tw+ t0B

(1) R represents the average food intake over all patches, where tw is the time that an

animal spends foraging within a patch, tB is the average time an animal travels between

(7)

this function, the model predicts that an animal spends an optimal amount of time in a local patch twhich leads to the maximization of R∗:

R= g0(t∗) (2)

The animal switches from local exploitation of a patch to global exploration of patches when the gain of food g0(t)

is equal to the overall food intake R. In the search of

items in memory, individuals should leave a memory patch when the recall- rate of items drops to the overall recall- rate of items.

Current Study

The optimal foraging model can be applied to semantic memory when a) individuals search between semantic fields in memory and b) predefined categories of items are present (Hills et al., 2012). Hills et al. (2012) found evidence for the patch- like cognitive structure of memory and found that individuals recall items in line with the marginal value theorem. This study replicates their study using Dutch verbal fluency data and participants.

The first hypothesis states that people switch from a local similarity based cue to a global frequency based cue to find a new patch. When a new patch is found the memory probe contains the local cue again. For example: DOG-CAT-HAMSTER-HORSE is a sequence wherein a transition from the local cue HAMSTER is made to the frequency cue where HORSE is retrieved, which does not share the same local cue as HAMSTER

(category = pets) but a different cue (e.g., category = farm animals). If people switch from a local similarity based cue to a global frequency cue, we can test whether transitions in the optimal foraging model occur in accordance with the marginal value theorem. When a local patch is depleted a transition should occur, the return rate of a person’s responses becomes slower than the overall return rate, which leads to a switch of categories and R is maximized (see equation 1). The optimal time for these switches is then when the intake rate of a local patch drops to the mean global intake rate for all patches (see equation 2).

(8)

Methods Participants

Participants were 323 first year psychology students gathered during the test week sessions at the University of Amsterdam. Each participant had signed an informed consent before the start of the test sessions week. This informed consent had a standard clause and applied to all tests in the test sessions. The study was approved by the psychology ethics committee of the University of Amsterdam. After deletion of invalid responses, 312 subjects’ responses were included in the final analyses.

Materials

The verbal fluency task was taken online on a computer. The test was taken on the computer where participants had to respond to a given stimulus by naming as many as animals as they could think of within two minutes. Each response was coded with a unique response ID, participants were coded with a participant ID. Response times were recorded in milliseconds.

Data encoding

In the study of Hills et al. (2012) a total of 369 unique animals were produced in 5187 valid animal entries. The current study finds approximately the same number of unique animal entries. The original data set consisted of 8938 responses on the verbal fluency test and 397 unique animal responses. The mean number of valid responses over all participants was 26 (SD = 10). For the analysis including the BEAGLE similarity, 8182 valid responses were included and 332 unique animals were taken into analyses. The difference in unique animals was caused because of the use of the English BEAGLE similarity matrix by Hills et al. (2012), in which some recalled Dutch animals were not

(9)

present. These animal names had to be converted to a similar animal (name) (e.g.

"aasgier" -> "gier"). Analyses with snaut distance consisted of 8188 valid responses and 348 unique animals. Some animal names had to be converted because the snaut was not able to find the animal names in semantic space. For more detailed information about the animal conversion, see the Appendix.

Modeling Semantic Space

To model the search process a representation of structural memory is needed: the semantic space. Three methods of representing the semantic memory structure have been used, successively: hand- coded categories, the BEAGLE similarity representation and snaut distances. Hand-coded categories were partly derived from the Troyer et al. (1997) and Hills et al. (2012) categories and extended with other categories (e.g., a dutch birds category "Nederlandse vogelsoorten"). The exhaustive list of categories and animals based on the dutch verbal fluency data set can be found in Appendix, examples of categories are "pets", "African animals", "farm animals", "insects", "reptiles and amphibians" and

"Australian animals". Some animals belonged to multiple categories, for example the animal "cat" belongs to both "pets" and "feline". Animal names were first assigned to categories which related to their habitat. Next animal names were assigned to species and finally to subcategories of larger categories (e.g., "rivier- en meerdieren" within

"waterdieren"). After flagging these items which belonged to multiple categories, final categories were assigned by hand.

To replicate results of the English animal verbal fluency task, the animal cosine similarity matrix of Hills et al. (2012) was used. This matrix is based on the bound

encoding of the aggregate language environment (BEAGLE) model which builds a semantic space representation of meaning and word order by statistical dependencies in language (Jones & Mewhort, 2007). Hills et al. (2012) used a version of the model which only takes into account the contextual information and not the word order to model the semantic

(10)

space of words. Words are represented by high-dimensional holographic vectors. The first time a word is encountered, it is assigned a random environmental vector, ei. These vectors

are sampled from a Gaussian distribution with µ = 0 and σ = 1/D where D the dimensionality of the vector holds. Every unique word has an environmental vector that does not change and is constant after the first encounter with the unique word. This vector holds the physical characteristics of words (e.g. spelling). The memory vector, mi, changes

each time the word is encountered in a piece of text. Each time a word is encountered new information about the word in environmental vector ei is stored in mi, which represents the

internal memory for contexts in which the environmental vectors have been encountered (Jones & Mewhort, 2007). When a corpus has been learned, the word’s memory

representation is then a vector pattern reflecting the word’s history of co-occurrence with other words. Words that frequently co-occur in the corpus will develop similar vector patterns (e.g., "cat" and "mouse"). The final similarity matrix used by Hills et al. (2012) and the current study is the vector cosine similarity between two word vectors. In order to use the similarity matrix, names of English animals were translated to Dutch animals and manually checked on correct translation. To match all Dutch animals with the English animals it was necessary to combine certain animals into one animal (e.g., "huiskat" becomes "kat"). BEAGLE was trained on a 400-million-word Wikipedia corpus (Willits, D’Mello, Duran, & Olney, 2007), and its memory representations were used to compute the pairwise cosine similarity matrix for a list of 771 animals. Only the 332 animals that were named in the study and matched to an English animal label were manually checked on translation.

Another similarity matrix was produced with snaut, a software package that measures the semantic distance between words in Dutch or English

http://meshugga.ugent.be/snaut// (Mandera et al., 2017). The corpus on which the model was trained is derived from the SONAR-500 corpus and a corpus of movie subtitles. The SONAR-500 consists of 500 million words of contemporary Dutch and includes a wide

(11)

variety of text/ document types. It provides a balanced sample of Dutch words based on textual materials from different sources such as books, magazines, newspapers and Internet based sources (e.g., Wikipedia and websites) (Mandera et al., 2017). The model has been tested in their paper by Mandera et al. (in press.) The Continuous Bag of Words Model (CBWM) is part of the family prediction models. Instead of explicitly representing the words and their context in a large matrix, the model incorporates a narrow window that holds relationships while mining trough the corpus. During the process, the weights of the network are updated and the model learns to predict the word given the context of words. The underlying principle of the model is implicitly learning how to predict one event (a word in a text corpus) from associated events. An additional advantage of the model (compared to the BEAGLE) is that it is trained using a stochastic gradient descent. This means that the model can be trained with only one target context pairing available for each update of the weights, instead of "storing" all co-occurrence information in a memory vector as is the case with the count models (Mandera et al., 2017).

Modeling the Search Process

To simulate the search process Hills et al. (2012) relied on the memory retrieval framework described in SAM theory (Raaijmakers & Shiffrin, 1981) and ACT- R

architectures described in (Anderson, 1993). The model assumes that recall from memory is achieved by probing cognitive semantic retrieval structures with a specific cue set: the memory probe. The formal definition of the retrieval process is stated as follows:

P (Ii | Q1, Q2, ..., QM) =

ΠMj=1S(Qi, Ii)βi

ΣN

k=1ΠMj=1S(Qi, Ik)βi

(3) Recall is assumed to be achieved by probing retrieval structures in memory with a set of cues in the memory probe M . I is the item that is recovered from semantic memory. The overall probability P (Ii | Q1, Q2, ..., QM) of the retrieval for an animal name I is

(12)

where S(Q, I) represents the semantic similarity between the cue Q and animal name I, divided by the product of the individual strengths for the total number N of possible animal names in a category. β is the weight for each individual cue. The maximum

likelihood method is used to estimate β for each participant’s sequence of animal names for two cue types: global and local. The global cue refers to the retrieval strength S(Q, I) for each animal name based on that animal names’ frequency of occurrence in the BEAGLE Wikipedia corpus. The local cue (previous word cue) generates a retrieval strength S(Q, I) as the similarity between the previous generated animal name and the current animal name I produced by the participant.

The static and dynamic models defined by Hills et al. (2012) were examined using either one or both of two possible cues: the frequency (global) and/or the previous item cue (local). Static models do not switch from local to global cue and ignore a patch- like cognitive structure. Dynamic models switch from local to global cues and therefore "forage" between patches. In the static models, the memory probe consists of the same set of cues during the entire retrieval interval, whereas in the case of dynamic models, switching occurs from the previous- animal name cue (local) to the frequency cue (global) and back to the local cue when a new patch is explored (Hills et al., 2012). A sequence of

DOG–CAT–HAMSTER–HORSE transitions from a local cue (HAMSTER) to a frequency cue (HORSE), which is not semantically similar to HAMSTER, i.e. a global cue. For model comparison the retrieval strengths for the global cues are based on the BEAGLE model of Hills et al. (2012). In further analyses, we also compare results with snaut distances. The random models have equal retrieval weights for every animal name in the search space. We then used our extended Troyer et al. (1997) categorization scheme to determine where local-to-global transitions occurred in our participants’ item sequences.

Model comparison results are the mean Bayesian Information Criterion (BIC) over participants for each model. The random models specify equal retrieval probabilities for all items in the semantic space. The BIC penalizes for the number of parameters that have to

(13)

be estimated in each model, lower BIC values indicate a better model fit. The results show the improvement of each static or dynamic model with either local, global or both local and global cues.

Results

We first check whether the structure in memory indeed resembles a patch- like structure, before testing the hypotheses. If the search process in semantic memory is analogous to the search process in space, then recalling an animal name from semantic space increases the probability that a semantically similar animal name will be retrieved next. Subsequently retrieved animal names that are close together in semantic space would be more similar than items that are further apart. The result of this check is illustrated in Figure 1. The figure gives the five beagle similarities and snaut distances averaged over all retrieved animal names for the five preceding animal names of a particular animal name. The figure tell us that the first preceding animal name (P -1) has a significantly higher similarity (or smaller distance) to the animal name than the second, third and so on. Both results with semantic spaces are supported by the significance test of ANOVA which gave a result of F(4,1502) = 66.4 with p < 0.001 for the BEAGLE similarity and F(4,1502) = 170.1 with p < 0.001 for the snaut distance.

The patch- like structure is also supported by the model improvements found in Table 1. Both static and dynamic (global and/or local cue) models show a mean

improvement over the random models with equal retrieval weights for every item in the search space. The random weights model never shows a mean improved fit over the static and dynamic (global and/or local cue) models. For most participants, the static and dynamic (global/local cue) models show an improved fit and for very few participants, the models show equal fit as the random weights model (difference in BIC ' 0 ).

Further analyses show that for all participants (N = 312 the static and dynamic (global and/or local cue) models have an improved fit over the random model, except for

(14)

the local and global cue static model, for which in N = 4 participants the models showed an equal fit. If we compare the static and dynamic models in order of the table, we find that for 74% the static local one cue model has an improved fit over the static global one cue model. In 68% of participants the combined local and global static model shows an improved fit over the static local one cue model. Next the combined cue dynamic model based on the Troyer categories shows an improvement in 55% of participants, and the similarity drop model formulated by Hills et al. (2012) shows an improvement in 66% of participants. The similarity drop model shows an improvement in 99% of participants compared to the first static one global cue model. Together these results indicate strong evidence for models that take into account local and global cues instead of random weights. This supports the theory of a patch- like structure of semantic memory.

Hills et al. (2012) stated that items are more similar before a patch transition than after the transition. The similarity between the last item and the new item is lower than the mean pairwise similarity over all item pairs, thus similarity between animal names drops after a patch transition. They formulated another model based on this hypothesis: the similarity drop model. Instead of using hand- coded categories to mark transitions in patches, these transitions are identified by a drop in similarity between two successive items. For example if we have a sequence of items like DOG, CAT, HORSE and COW , and S(DOG, CAT) > S(CAT, HORSE) and S(CAT, HORSE) < S(HORSE, COW), the transition point would be between CAT and HORSE) (Hills et al. 2012). Figure 2 shows the mean ratio of pairwise similarity and distance between five successive items to

participants’ mean pairwise similarity and distance over all item pairs. For the BEAGLE similarity, it is shown that the ratio drops below 1, indicating that the similarity drops between the last item of a patch and a new item of another patch. With snaut distances, this is shown by a greater distance which leads to a ratio that exceeds 1. Another

observation is that the second word after the transition (Position 2) seems to have greater similarity with the previous item (Position 1) than the last two items in a patch (Position

(15)

-2 and -1). This shows that when a new patch is entered, items that are more similar are recalled after each other, and that when a patch is almost depleted, they are less similar.

Hypothesis 1: The dynamic model that makes local–global cue transitions will outperform the static model.

To answer the question whether individuals switch from a local similarity based cue to a global frequency cue is tested by comparing the dynamic model with the static model (see Table 1). The dynamic model with categories defined by the extended Troyer et al. (1997) categories outperforms the static model slightly but not significant p < 0.05. However the similarity drop model does show a significant improvement over the static model with p < 0.01. This supports a dynamic search process, in which individuals switch between patches in their search for items.

Hypothesis 2: Participants switch patches when a local patch is almost depleted.

If we assume that individuals search in accordance with dynamic models, we would expect that they switch patches when a local patch is almost depleted. This means that items recalled just before a patch switch are less similar to all other items in the search space than items right after a patch switch. An example is the following sequence of animals: DOG - CAT - HAMSTER - MOUSE - FERRET, where FERRET is presumably less similar to other animals in search space than the previous animals. Hills et al. (2012) call a word’s semantic proximity to other words its residual proximity. They calculated this as the mean similarity (inverse distance) to all other not yet produced words in the search space (e.g., animals), including only not yet retrieved items in calculations. From figure 4 it is shown that the residual proximity before a patch switch defined by the hand coded categories (Position -1) is lower than when a switch has occurred (Position 1). Also, the first two items in a new patch show a higher residual proximity to other items in search

(16)

space, showing that a new patch was entered. The figure illustrates the transition from a local patch to another patch when the former is depleted.

Hypothesis 3: The optimal time to transition is when the local intake rate drops to the mean global intake rate.

Based on the previous results, a dynamic model of memory search seems to describe the search process in semantic memory. According to the marginal value theorem, the optimal time to switch is when the local intake rate drops to the mean global intake rate (see equation 2). To analyze this we calculate the intake rate to be the ratio between a participant’s interitem response time (IRT) and the overall mean IRT between the

participant’s items. In figure 3 the relative (IRT) between an item and the previous item is demonstrated. The bar above "1" gives the relative IRT for the first word in a patch and from the figure we can see that the ratio between IRT and mean IRT exceeds 1, indicating that the time between items is higher than the overall mean IRT. This shows a switch from local to global when the intake rate becomes slower because of patch depletion, which leads the participant to switch patches. The figure illustrates that the IRT is higher when

switching patch, which demonstrates a local to global transition when the local patch is (almost) depleted. Results of a paired t- test show that the first word after a switch from local to global takes significantly longer than the mean IRT for participants, t(300) = 4.29, p < 0.001. To produce the second word in a patch, participants take significantly less time than the overall mean IRT t(300) = -14.22, p < 0.001. T- tests were performed on the dataset used in the analysis with BEAGLE similarity (N = 312), and some participants (N = 11) did not produce items in the same category and therefore had no matched pair for this analysis. To check results, t- tests were conducted on the data used in snaut analyses, showing again significant results with t(294) = 3.70, p < 0.001 and t(294) = -17.04, p < 0.001. More support for an optimal intake rate is illustrated in the figure, which shows the relative interitem retrieval time (IRT) between a word and the previous word divided by

(17)

the mean IRT in participants. The figure shows that the ratio item IRT over mean IRT of previous item exceeds 1, which demonstrates a local to global transition.

Hills et al. (2012) found that participants who deviated more from the optimal time to leave a patch produced fewer items on the verbal fluency task than individuals with an optimal foraging strategy. They looked whether the last IRT in a patch is almost equal to the overall mean IRT in participants. If the last IRT is considerably smaller this indicates that a local patch is left too soon. On the contrary, if the last IRT before a patch switch is greater, individuals stayed too long in a local patch. To test this hypothesis, a regression was performed to predict the number of items participants produced as a function of the absolute difference between the mean last item IRT and the overall mean IRT. To make the results more interpretable, milliseconds were converted to seconds in the analysis. Again, results showed a significant negative relationship with β = -4.9, t(305) = -5.78 and p < 0.001 for the data with the beagle similarity and β = -5.3, t(305) = -4.09 and p < 0.001 for snaut. In figure 5 the negative relationship is given for both results on the data with beagle similarity and snaut distance, both show the same relation between the absolute difference and the number of words produced by participants.

Discussion

The current study investigated whether the search in semantic memory resembles the search of animals foraging for food in their environment. This search process is described by a process in which individuals switch from local patches (semantic fields) to global exploration of semantic memory (Hills et al., 2012). According to the marginal value theorem, animals switch from a local patch to global exploration when a patch is almost depleted, and the intake rate within this patch drops to the average overall intake rate (Charnov, 1976).

Support for a semantic memory distributed in patches was shown in the model comparison. There was no improvement in the model with the hand- coded Troyer et al.

(18)

(1997) animal categories, but the similarity drop model based on similarities between animal names instead of assigned categories did show improvement. The reason for this difference could have been that the way animals were categorized was not optimal. Hills et al. (2012) assigned patch membership in comparison to animal names recalled by

participants. In this study animal categories were in the first place assigned by semantic meaning only. In a next study, it would be better if the animal categories were assigned relative to the other animal names a participant recalled. The context in which an animal is recalled is very varying, for example if we look at the two following sequences, we can see that the patch category is different if we would take into account the context of previous recalled animal names: CAT - DOG - MOUSE - HAMSTER; RAT - MOUSE - HAMSTER - CAT. In the latter sequence, the transition of patch occurs with cat (which is not a rodent), while in the previous sequence, there seems no transition because all animals are domestic. Therefore the animal name HAMSTER belongs to different categories,

depending on the context. If we would assign category membership in order of the

responses, it could be easier to detect whether animal names belong to the same category. In the current study, there were many hand- coded boundaries between animal names and therefore the patch- structure of memory was harder to detect.

After testing for a patch- distributed semantic memory, tests to assess an optimal foraging strategy in participants showed indeed that participants search memory in line with the marginal value theorem. Individuals switch from patch when the return rate becomes slower, which leads to patch switches when items in a local patch are depleted.

The current study replicated the study of Hills et al. (2012) with Dutch verbal fluency data. Looking at the results of both studies, we can conclude that a patch-distributed structure of semantic memory seems in line with how individuals search for items. These findings are important for several reasons: 1) they support theories of neural networks, showing that items are clustered in long- term memory and people transition from one cluster to another (Hills et al., 2012), 2) they illustrate the flexibility of the search

(19)

process, explained by associative theories (e.g. SAM) (Raaijmakers & Shiffrin, 1981) and 3) they support an optimal foraging strategy in individuals, suggesting that individuals switch from patch when the number of recalled items drops (Charnov, 1976).

In a next study it would be nice to compare the models with a Dutch based semantic space. In this analysis, the English animal names were translated which might lead to different similarities. For example, if we train the similarities on English corpora, it is possible that this influences the semantic similarities between animal names (e.g., "it rains cats and dogs" could occur more often which leads to a stronger similarity between these to words).

Previous studies in Alzheimer patients support the results, showing that these patients have trouble recalling from memory and producing less items than healthy individuals (Raoux et al., 2008). The authors found that this is due to the fact that patients switch less and therefore produce fewer items, indicating an executive function problem instead of a semantic problem.

(20)

Table 1

Mean BIC for static and dynamic models with either local/global cues or both cues over all participants. For beta, median estimates over all participants are shown. Standard

deviations are shown in parentheses.

Model β BIC

Mean BIC of all random (equal weights) models 170.1 (63.4) Local or global cue static models

Global 7.1 (2.9) 147.3 (57.7)

Local 3.5 (1.2) 143.0 (56.3)

Local and global cue static models 140.9 (55.2)

Global cue 5.1 (3.3)

Local cue 2.5 (2.0)

Local and global cue dynamic models

Categories 140.2 (55.2)

Global cue 5.9 (3.2)

Local cue 3.9 (3.3)

Similarity drop model (Hills et al. 2012) 138.3 (54.1)

Global cue 5.3 (3.2)

Local cue 3.8 (2.1)

(21)

References

Anderson, J. R. (1993). Rules of the mind. Hillsdale, NJ: Erlbaum.

Bousfield, W., & Sedgewick, C. (1944). An analysis of sequences ofrestricted associative responses. J Gen Psychol, 30 , 149-165.

Bousfield, W. A., & Barclay, W. D. (1950). The relationship between order and frequency of occurrence of restricted associative responses. Journal of experimental psychology, 40 (5), 643.

Charnov, E. L. (1976). Optimal foraging: attack strategy of a mantid. The American Naturalist, 110 (971), 141-151.

Davelaar, E. J. (2015). Semantic search in the remote associates test. Topics in cognitive science, 7 (3), 494-512.

Goni, J., Arrondo, G., Sepulcre, J., Martincorena, I., de Mendizábal, N. V.,

Corominas-Murtra, B., B., . . . Villoslada, P. (2011). The semantic organization of the animal category: evidence from semantic verbal fluency and network theory.

(Cognitive processing, 12(2), 183-196)

Gronlund, S. D., & Shiffrin, R. M. (1986). Retrieval strategies in recall of natural categories and categorized lists. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12 (4), 550.

Gruenewald, P., & Lockhead, G. (1980). The free recall of categoryexamples. J Exp Psychol Hum Learn Mem, 6 (3), 225-240.

Hills, T. T., Jones, M. N., & Todd, P. M. (2012). Optimal foraging in semantic memory. Psychological Bulletin, 119 (2), 431-440.

Howard, M. W., Jing, B., Addis, K. M., & Kahana, M. J. (2007). Semantic structure and episodic memory. Handbook of latent semantic analysis, 121-142.

Johnson, D. M., Johnson, R. C., & Mark, A. L. (1951). A mathematical analysis of verbal fluency. The Journal of General Psychology, 44 (1), 121-128.

(22)

in a composite holographic lexicon. Psychological review, 114 (1), 1.

Mandera, P., Keuleers, E., & Brysbaert, M. (2017). Explaining human performance in psycholinguistic tasks with models of semantic similarity based on prediction and counting: A review and empirical validation. Journal of Memory and Language, 92 , 57-78.

Metcalfe, J., & Murdock, B. B. (1981). An encoding and retrieval model of single-trial free recall. Journal of Verbal Learning and Verbal Behavior , 20 (2), 161-189.

Raaijmakers, J. G., & Shiffrin, R. M. (1980). Sam: A theory of probabilistic search of associative memory. Psychology of Learning and Motivation, 14 , 207-262.

Raaijmakers, J. G., & Shiffrin, R. M. (1981). Search of associative memory. Psychological review, 88 (2), 93.

Raoux, N., Amieva, H., Le Goff, S. H., Letenneur, L., & Dartigues, J. . (2008). Clustering and switching processes in semantic verbal fluency in the course of alzheimer’s disease subjects: Results from the paquid longitudinal study.

(Cortex, 44(9), 1188-96)

Romney, A. K., Brewer, D. D., & Batchelder, W. H. (1993). Predicting clustering from semantic structure. Psychological Science, 4 (1), 28-34.

Troyer, A. K., Moscovitch, M., & Winocur, G. (1997). Clustering and switching as two components of verbal fluency: Evidence from younger and older healthy adults. Neuropsychology, 11 (138), 894-4105. doi: 10.1037/–146. 0.11.1.138

Troyer, A. K., Moscovitch, M., Winocur, G., Leach, L., & Freedman, M. (1998). Clustering and switching on verbal fluency tests in alzheimer’s and parkinson’s disease. Journal of the International Neuropsychological Society, 4 (2), 137-143.

Willits, J. A., D’Mello, S. K., Duran, N. D., & Olney, A. . (2007). January). distributional statistics and thematic role relationships. In Proceedings of the Cognitive Science Society (Vol., 29 , 29.

(23)

Figure 1 . BEAGLE similarity and snaut distance for an animal name’s preceding animal names in a category for participants. Animal names that are produced earlier are less similar (have greater distance) to the animal name than the animal names produced later (Position -1). The error bars represent the standard error of the mean.

(24)

Figure 2 . Mean ratio of pairwise similarity and distance between successive retrieved animal names to a participant’s mean pairwise similarity or distance over all item pairs. For the snaut, the ratio for the animal name in Position 1 exceeds 1, since the distance is greater between the name before than after the path switch.

(25)

Figure 3: The relative interitem retrieval time (IRT) of a word and the mean IRT of

participants. The bar above "1" gives the IRT between the first word in a category and the previous last word in a patch divided by the mean IRT. The figure demonstrates a local to global transition since the IRT after a transition exceeds 1, which is equal to the mean IRT for a participant over the entire task. Error bars are standard error of the mean.

(26)

Figure 4 . Residual proximity of animal name in relation to their position before or after a patch transition. Included are only animal names not yet retrieved in the computation of an animal name’s residual proximity value. Error bars are standard error of the mean.

(27)

Figure 5 . Absolute difference between the mean last-item IRT across patches and the mean IRT over the entire task. BEAGLE similarity in crosses and straight line, snaut distance in points and dotted line.

(28)

Appendix

Participants recalled the following categories with Dutch animal names:

- Huisdieren (domestic animals): Cavia, hamster, fret, guppy, hond, husky, kat, kanarie, konijn, labrador, muis, papegaai, parkiet, rat, vis, vogel, wandelende tak (16 animals).

- Afrikaanse en jungle dieren (african and jungle animals): Aap, aardvarken, antilope, baviaan, bizon, bonobo, brulaap, buffel, cheetah, chimpansee, dikdik, gazelle, gier, giraffe, gnoe, gorilla, hyena, impala, jaguar, kameel, leeuw, lemur, lijger, luiaard, luipaard, lynx, makaak, neushoorn, nijlpaard, oerang oetan, okapi, olifant, panda, panter, piranha, poema, ringstaartmaki, rode franjeaap, schorpioen, slingeraap, steenbok, stekelvarken, stokstaartje, struisvogel, tasmaanse duivel, tijger, vleermuis, zebra, zwijn (49 animals).

- Bosdieren (forest animals): Bruine beer, das, eekhoorn, egel, ekster, eland, emoe, haas, hert, mol, ree, rendier, stinkdier, uil, vos, wild zwijn, wasbeer, wolf, zwarte beer (19 animals).

- Boerderijdieren (farm animals): Big, ezel, geit, haan, kalf, kalkoen, kip, koe, kuiken, lam, paard, pony, rund, schaap, varken, veulen (16 animals).

- Insecten (insects): Bij, duizendpoot, horzel, insect, kakkerlak, kever, krekel, larf, libelle, lieveheersbeestje, meelworm, mier, mot, mug, naaktslak, pissebed, regenworm, rups, slak, spin, sprinkhaan, stofmijt, teek, termieten, vlieg, vlinder, vlo, wandelende tak, wesp, worm (30 animals).

- Reptielen en amfibieën (reptiles and amphibians): Adder, alligator, amfibie, anaconda, boa constrictor, cicade, cobra, gekko, hagedis, kameleon, kikker, kikkervisje, komodovaraan, krokodil, leguaan, pad, ratelslang, reptiel, salamander, slang, zeeslang (21 animals).

- Australische dieren (australian animals): Buidelrat, dingo, kangoeroe, koala, vogelbekdier, wombat (6 animals).

(29)

zwarte beer (8 animals).

- Nederlandse vogelsoorten (dutch bird types): Bergeend, buizerd, duif, eend, ekster, fazant, gans, havik, ijsvogel, kalkoen, kauw, kieviet, koekoek, kraai, leeuwerik, lepelaar, mees, meeuw, merel, mus, nachtegaal, ooievaar, patrijs, pauw, raaf, reiger, roodborstje, specht, spreeuw, uil, valk, vlaamse gaai, vogel, zeemeeuw, zwaan, zwalum (36 animals).

- Zuid- en Noordpooldieren (arctic animals): Dolfijn, ijsbeer, orka, pinguïn, poolvos, walrus, zeehond, zeekoe, zeeleeuw, zeeolifant (10 animals).

- Katachtigen (felines): Cheetah, civet, jaguar, kat, leeuw, lijger, luipaard, lynx, panter, poema, poes, siamees, sneeuwluipaard, tijger (14 animals).

- Hondachtigen (canines): Bulldog, chihuahua, dalmatiër, dingo, golden retriever, hond, husky, hyena, labrador, poedel, tekkel, wilde hond, wolf (13 animals).

- Knaagdieren (rodents): Bever, capibara, cavia, chinchilla, eekhoorn, gerbil, haas, hamster, konijn, marmot, molrat, muis, otter, rat, relmuis, spitsmuis, veldmuis (17 animals).

- Insectivoren (insectivores): Aardvarken, egel, gordeldier, miereneter, mol, tapir, vleermuis, vliegende eekhoorn (8 animals).

- Hoefdieren (ungulates): Antilope, dikdik, eland, emoe, ezel, gazelle, geit, hert, kalf, schaap, wild zwijn (9 animals).

- Dieren voor bont (fur animals): Alpaca, lam, lama, schaap (4 animals). - Waterdieren (water animals): Aal, aalscholver, alg, baars, barracuda, beloega, bever, blauwe vinvis, bloedzuiger, bruinvis, bultrug, capibara, clownvis, dolfijn, forel, gamba, garnaal, guppy, haai, hamerhaai, haring, inktvis, karper, katvis, kogelvis,

koraalduivel, krab, kreeft, kwal, maanvis, makreel, mantrarog, marmot, mossel, nijlpaard, octopus, oester, orka, otter, pijlstaartrog, piranha, schelpdieren, schildpad, snoek, spons, steenvis, stekelbaars, tilapia, tonijn, vis, walrus, walvis, witte haai, zalm, zee anemoon, zee-egel, zeebaars, zeehond, zeekoe, zeeleeuw, zeeolifant, zeeotter, zeepaardje, zeeslang,

(30)

- Meer- en rivierdieren (lake and river animals): Aal, aalscholver, baars, bever, forel, karper, paling, snoek, tilapia, vis, zalm (11 animals).

- Primaten (primates): Aap, baviaan, bonobo, brulaap, chimpansee,

doodshoofdaapje, gibbon, gorilla, lemur, makaak, maki, oerang oetan, ringstaartmaki, rode franjeaap, slingeraap (15 animals).

- Vogels (birds): Adelaar, ara, bergeend, buizerd, duif, eend, ekster, fazant, flamingo, gans, gier, havik, ijsvogel, kaketoe, kalkoen, kanarie, kauw, kieviet, kip, koekoek, koolmees, kraai, leeuwerik, lepelaar, meerkoet, mees, meeuw, merel, mus, nachtegaal, ooievaar,

papegaai, parkiet, patrijs, pauw, pelikaan, raaf, reiger, roodborstje, specht, sperwer, spreeuw, toekan, uil, valk, vlaamse gaai, vogel, zee-arend, zeemeeuw, zwaan, zwaluw (51 animals).

- Zuid- Amerikaanse dieren (South- American animals): Alpaca, chinchilla, gordeldier, lama, luiaard, miereneter, poema (7 animals).

- Buffelachtigen (buffaloids): Bizon, buffel, koe, os, rund, waterbuffel, yak (7 animals). - Marterachtigen (mustelids): Bunzing, fret, hermelijn, nerts, vos, wezel (6 animals). - Lastdieren (pack animals): Buffel, dromedaris, ezel, kameel, lama, os, paard, pony, waterbufel, yak (10 animals).

- Geleedpotigen (anthropods): Krab, kreeft, schorpioen, spin, tarantula, teek, zwarte spin (7 animals).

Schelp- en schaaldieren (shellfish): Mossel, oester, schelpdieren (3 animals).

Animal names that had to be converted in each category were converted to already existing animal names in each category:

- Huisdieren (domestic animals): Dwerghamster -> hamster, goudvis -> vis, huiskat -> kat, poes -> kat.

- Afrikaanse en jungle dieren: Aasgier -> gier, jachtluipaard -> cheetah, japanse makaak -> makaak, kruisspin -> spin, maki -> lemur, neusaap -> aap, rode brulaap ->

(31)

brulaap, rode panda -> panda, sabeltandtijger -> tijger, sfynx -> kat.

- Bosdieren (forest animals): Boskonijn -> konijn, brilbeer -> beer, edelhert -> hert, everzwijn -> wild zwijn.

- Boerderijdieren (farm animals): Ram -> schaap.

- Insecten (insects): Ameise -> mier, eendagsvlieg -> vlieg, fruitvlieg -> vlieg,

hommel -> bij, hooiwagen -> spin, langpootmug -> mug, langpootspin -> spin, mestkever -> kever, oorkruiper -> insect, ringworm -> worm, rode mier -> mier, schrijvertje -> kever, tor -> kever, veenmol -> krekel, waterjuffer -> libelle, zilvervisje -> insect.

- Reptielen en amfibieën (reptiles and amphibians): Kraaghagedis -> hagedis, nijlkrokodil -> krokodil, ringslang -> slang.

- Australische dieren (australian animals): Baardagaam -> hagedis, blobvis -> vis, buideldier -> kangoeroe.

- Beren (bears): Brilbeer -> beer, rode panda -> panda.

- Nederlandse vogelsoorten (dutch bird types): Gaai -> vlaamse gaai, grutto -> vogel, houtduif -> duif, houtuil -> uil, koolmees -> mees, meerkoet -> eend, nijlgans -> gans, postduif -> duif, sperwer -> vogel, steenuil -> uil (36 animals).

- Katachtigen (felines): Europese korthaar -> kat, jachtluipaard -> cheetah, sabeltandtijger -> tijger, sfynx -> kat.

- "Hondachtigen" (canines): Herdershond -> hond, hulphond -> hond, maltezer -> hond.

- Knaagdieren (rodents): Beverrat -> bever, boskonijn -> konijn, dwerghamster -> hamster.

- Hoefdieren (ungulates): Bok -> geit, edelhert -> hert, everwzijn -> wild zwijn, ram -> schaap.

- Waterdieren (water animals): Beverrat -> bever, blobvis -> vis, coelacanth -> vis, diepzeevis -> vis, goudvis -> vis, kabeljauw -> vis, kokkel -> schelpdieren, kooikarper -> karper, langoustine -> kreeft, neusvis -> vis, paling -> aal, pistoolgarnaal -> garnaal,

(32)

platvis -> bot, potvis -> walvis, rog -> mantrarog -> roodbaars -> baars, schaaldieren -> schelpdieren, schol -> bot, sidderaal -> snoek, sliptong -> bot, tijgerhaai -> haai,

venusschelp -> schelpdieren, vinvis -> blauwe vinvis, vliegende vis -> vis, weekdier -> schelpdieren, zeeduivel -> vis, zeekomkommer -> zee anemoon.

- Meer- en rivierdieren (lake and river animals): Beverrat -> bever, kooikarper -> karper, sidderaal -> snoek, witvis -> vis.

- Primaten (primates): Japanse makaak -> makaak, neusaap -> aap, rode brulaap -> brulaap.

- Vogels (birds): Aasgier -> gier, arend -> adelaar, gaai -> vlaamse gaai, houtduif -> duif, houtuil -> uil, nijlgans -> gans, paradijsvogel -> vogel, postduif -> duif, roofvogel -> vogel, steenuil -> uil, xenops -> nachtegaal, zebravink -> vink.

- Buffelachtigen (buffaloids): Schotse hooglander -> rund, stier -> rund.

- Geleedpotigen (anthropods): Blauwe spin -> spin, hooiwagen -> spin, kruisspin -> spin, rode spin -> spin, vogelspin -> tarantula.

Schelp- en schaaldieren (shellfish): Kokkel -> schelpdieren, langoustine -> kreeft, schaaldieren -> schelpdieren, venusschelp -> schelpdieren.

Animal names that had to be deleted in each category:

- Insecten (insects): Eencelligen, organisme, pantoffeldiertje, schimmels. - Waterdieren (water animals): Amoebe, plankton

Referenties

GERELATEERDE DOCUMENTEN

In mijn opinie is het behandelen van klachten ook meer een taak voor een speciaal daartoe uitgeruste klachtencommissie, hetzij door de overheid, hetzij door het tuchtcollege of een

Vanwege het bestorten van de voor oever is deze aanvoer de laatste jaren wel gereduceerd, maar er spoelen nog wel degelijk fossielen aan, schelpen zowel als haaientanden.. Van

We call this problem Compressive System Identification (CSI). CSI is beneficial in applications when only a limited data set is available. Moreover, CSI can help solve the issue

Keywords: memory retrieval, query theory, incumbency advantage, information processing, political decision making..

“Bramhall, John (bap.. Long Shadows: truth, lies and history. Politicians and pamphleteers: propaganda during the English civil wars and interregnum. “Enacting

Moreover, the focus was mainly on the Stockholm Declaration and the Task Force on International Cooperation on Holocaust Education, Remembrance and Research and

By varying the shape of the field around artificial flowers that had the same charge, they showed that bees preferred visiting flowers with fields in concentric rings like

I don't have time, in the middle of a conversation, for them to search their memory bank for what a protein is made of or for them to go off and look up the answer and come back