Word Learning in Context Thesis Presented in Partial ... - OhioLINK ETD

2 downloads 155 Views 354KB Size Report
... a certain word- learning context, could also help young children to learn new words. ..... ambiguous by blurring it in Photoshop 6.0 (see Figure 1). There was ...
Word Learning in Context

Thesis

Presented in Partial Fulfillment of the Requirements for the Degree Master of Arts in the Graduate School of The Ohio State University By Xin Yao, B.S. Graduate Program in Psychology

The Ohio State University 2010

Thesis Committee: Vladimir Sloutsky, Advisor John Opfer Laura Wagner

Copyright by Xin Yao 2010

Abstract

Word learning is a difficult problem. On the one hand, young children could learn new words with few positive examples quickly and easily. On the other hand, word learning itself involves many complicated cognitive processes. For example, one needs to map a new label to an object and generalize the new label from one object to the entire category of the object. Some researchers have suggested that word learning involves hypotheses testing: when young children learn new words, they rely on a priori constrains that could help them limit the space of hypotheses. According to this view, the constraints allow young word learners not to consider some of the possible hypotheses; whereas others are ruled out some others are ruled out on the basis of positive examples given to the word learner. Other researchers have further developed this idea. Instead of ruling in or ruling out the hypotheses, young word learners are updating these possibilities by using a Bayesian approach to hypotheses testing. However, the five experiments in the current study suggest that hypotheses testing, including the Bayesian approach, is not necessary. In particular, hypotheses testing could not account for all the reported findings. Instead, semantic word association, which could provide young word learners a certain wordlearning context, could also help young children to learn new words.

ii

This document is dedicated to my family.

iii

Acknowledgments

I would like to express my deepest gratitude to my adviser, Dr. Vladimir Sloutsky. His guidance and support, as well as limitless brilliant ideas, made this research possible. I thank Dr. John Opfer for his brilliance and support. I also thank Laura Wagner for her brilliance and support. I thank all members of our research associate team, past and present, especially Abby Claflin, Marie Valentine-Elam, Cynthia Walter, and Rachel Zufall for helping to collect the data presented in this thesis. I thank all members of the Cognitive Development lab, especially Dr. Chris Robinson, Dr. Jennifer Kaminski, and Sravani Vinapamula. I thank my parents for loving me and supporting me. This thesis is simply impossible without their love and support. I thank my friends Bo Guan, Yun Tang, Nuo Xi, Miao Luo, Hao Wu, Wei Deng, Xin Wang, Haozhi Xiong, and Xiaoshu Xu, , for all the emotional support, entertainment, and caring they provided in the past years. They let me own a happy graduate life in the United States. iv

Vita

2006................................................................B.S. Psychology, East China Normal University 2006-2007 .....................................................Graduate Teaching Associate, Department of Psychology, The Ohio State University 2007-2008 .....................................................Graduate Research Associate, Cognitive Development Lab, The Ohio State University 2008 to present ..............................................Graduate Teaching Associate, Department of Psychology, The Ohio State University

Publications

Sloutsky, V. M., & Yao, X. (2008). Learning words from context. In B. C. Love, K. McRae, & V. M. Sloutsky (Eds.), Proceedings of the 30th Annual Conference of the Cognitive Science Society (pp. 64-70). Austin, TX: Cognitive Science Society.

Fields of Study Major Field:Psychology v

Table of Contents

Abstract ............................................................................................................................... ii Acknowledgments.............................................................................................................. iv Vita ...................................................................................................................................... v Table of Contents ............................................................................................................... vi List of Figures ................................................................................................................... vii Chapter 1: Introduction ...................................................................................................... 1 Chapter 2: Experiments...................................................................................................... 9 Chapter 3: General Discussion.......................................................................................... 29 References ..................................................................................................................... 33 Appendix A: Questions used in induction task in Experiment 1 ...................................... 35 Appendix B: Associative Strength of the word “animal” for both adults and children (Taxonomic List) .............................................................................................................. 36 Appendix C: Associative Strength of the word “animal” for both adults and children (Associative List) .............................................................................................................. 37 Appendix D: Associative Strength of the word “animal” for both adults and children (Control List) .................................................................................................................... 38 vi

List of Figures

Figure 1.Stimuli used in the study phase of Experiment 1 ............................................... 10 Figure 2. Stimuli used in label extension task in Experiment 1........................................ 11 Figure 3. Proportion of label extension responses by age group and condition in Experiment 1. .................................................................................................................... 14 Figure 4. Proportions of induction responses by age group and condition in Experiment 1. ........................................................................................................................................... 15 Figure 5.Proportions of label extension responses by age group and condition in Experiment 2. .................................................................................................................... 21 Figure 6.Stimuli used in lexical extension task in Experiment 5 ...................................... 28

vii

Chapter 1: Introduction

Inductive generalization is an interesting problem in human cognition, especially in the field of children’s conceptual development. How do young children acquire categories? How are categories represented? And do these representations change in the course of development and learning? These are not trivial issues in the study of conceptual development. There are variety of forms of inductive generalization, including word learning, category learning, and induction of properties. Young children and even infants show the ability to generalize across a variety of tasks (Gelman, 1990; Balaban & Waxman, 1997, Welder & Graham, 2001, but see Sloutsky& Fisher, 2004). For example, in Gelman’s (1990) induction tasks, young children were given two test items and a target item, and were informed about properties of the test items. Children were then asked to choose which property the target item had. In the task, the target item looked like one of the test items, whereas it shared the label with another test item. It was found that young children were more likely to generalize the property in question from the test item that shared the label with the target than from the one that looked alike (Gelman, 1990). In categorization tasks, the word facilitation effect is also found in 12-month infants (Balaban & Waxman, 1997). In these studies, infants were familiarized to several objects from the same group (for example, a rabbit). The object was presented either paired with 1

a word phrase (“look at the the rabbit”) or with a tone. In the test phase, a novel object from the same group (a novel rabbit) and an object from another group (a pig) were presented simultaneously. It was found that infants show novelty preference in the word condition but not in the tone condition However, although these phenomena have been reported, the mechanism underlying these effects remains unclear. For example, even learning a name for object category involves complex cognitive processes and represents a difficult induction problem. It could be traced back to the classic problem of the indeterminacy of translation, which was initially formulated by Quine (1960). Consider Quine's famous example of the "Gavagai" problem. How can we know that “Gavagai” means “rabbit?” There are in fact an unlimited number of possibilities for a potential word-world mapping. The word might refer to all rabbits, all mammals, all animals, the individual rabbit, all furry things, running things, or the ears of the rabbit. Despite these unlimited possibilities, infants and toddlers can acquire new words at a rapid rate (Booth, Waxman, & Huang et al, 2005). How do young children solve this problem? There are several possibilities. Some researchers suggest that there are some a priori constraints that enable infants to quickly map the linguistic labels to objects and generalize properties from members to the category (see Murphy, 2004 for a review). These researchers argue that although the input from the environment is unconstrained, there are some innate constraints (e.g., assumptions and biases) that enable children to focus on the key features only (Murphy, 2004). For example, a young word learner may prefer interpreting a novel word as referring to a whole object rather than to a salient part of the object (Markman & 2

Wachtel, 1988); individual objects are assumed to be organized in categories (Gelman, 2004); and there is preference to the novel label denoting basic level objects (Murphy, 2004). These assumptions and biases reduce the number of possible word-object mapping, which makes it possible for young children to solve the mapping problem. However, this position has a number of limitations. For example, if young children prefer to extend novel labels to the basic level objects, what is the mechanism for them to learn words for superordinate or subordinate level categories? It is also not sufficient to explain how a single or few positive examples enable young children to infer a word’s meaning (see Xu & Tenenbaum, 2007 for a review), even if children are testing hypothesis from a limited hypotheses space. For example, suppose that a young word learner sees three grey rabbits and each one is referred to as “Gavagai.” None of the hypotheses are eliminated in this situation: it is still possible that Gavagai refers to grey rabbits, to rabbits or to animals with four legs. To rule out the latter two hypotheses, negative examples are necessary, for instance, when children see a white rabbit or a duck and are told this is not Gavagai. Thus, positive examples alone do not help rule out other hypotheses of word-object mapping. One approach that can potentially address these difficulties is the Bayesian inference framework (Xu & Tenenbaum, 2007). The Bayesian approach is committed to the idea that young children learn new words by testing hypotheses. However, the hypotheses about word meaning are not ruled in or out but are evaluated by probability, which is calculated according to the Bayes rule. According to this approach (Xu & Tenenbaum, 2007), young word learners evaluate the posterior probability, the probability of the target hypothesis being true given several positive examples. The posterior probabilities 3

are proportional to the product of prior probabilities p(h) and the likelihoods p(X|h). Priors could reflect the assumptions or constraints young word learners have. For instance, basic level categories will have higher priors than superordinate or subordinate categories if young word learners indeed have such preference (see Murphy 2004, for a review), to denoting the novel label to basic level categories. The likelihood is the probability of getting the current examples under each considered hypothesis. Therefore, hypotheses will be assigned higher probability to be true as the number of consistent examples increases (Xu & Tenenbaum, 2007). For example, in mentioned above “Gavagai” problem, there are several possible hypotheses of the meaning of “Gavagai.” It could be Hypothesis A (i.e., Gavagai means mammal), Hypothesis B (i.e., Gavagai means rabbit) or many other hypotheses. When Gavagai is introduced once, the likelihood of Hypothesis A will be proportional to 1/m, where m is the total number of the mammals, assuming it is randomly sampled. Accordingly, the likelihood of Hypothesis B will be proportional to 1/n, where n is the total number of the rabbits. As there are more mammals than rabbits in real life, if the word “Gavagai” is introduced multiple times (k times), and each time it is introduced with a rabbit, the likelihood of Hypothesis A being true, Gavagai referring to all the mammals, will be (1/m)k, while the likelihood of Hypothesis B being true, Gavagai referring to all the rabbits, will be (1/n)k. Therefore, the probability, or the posterior, of Gavagai referring to all the mammals or to all the animals will be much lower than the probability of Gavagai referring only to rabbits. Thus, young word learners will be more likely to infer that Gavagai refers to rabbits rather than to mammals. To summarize, the Bayesian approach suggests that 4

when young children learn new words, they are testing hypotheses by updating the posterior of each hypothesis. The posterior is determined by (a) the prior, (e.g., young child’s bias or preference), and (b) the likelihood of a given hypothesis. The novel label is more likely to refer to the category with higher posterior probability. As no hypotheses would be ruled in or ruled out, word learners should be able to keep the information in mind and update it every time when they receive new examples. This approach could answer the question of how young children could learn new word very quickly by just being given a few positive examples. However, this approach does not address several important problems. First, it suggested that word learners have structured taxonomic knowledge (Xu & Tenenbaum, 2007), and this information is critical for determining the likelihood of each hypothesis. However, it is not clear how young children acquire this knowledge prior to learning new words. Second, it is not clear how young children can organize and store all the information about hypotheses, their probabilities and keep updating them when new information becomes available. These processes require enormous memory capacity and it is not clear whether young children do have this capacity. Here we propose a different idea about how young children may learn new words. Both hypotheses testing approach and the Bayesian inference framework are based on the assumption that there is a priori knowledge that guides young children’s word learning. However, studies also show that young word learners are sensitive to statistical information in the input without a priori knowledge (Saffran, Aslin, & Newport, 1996; Colunga & Smith, 2005; Yu & Smith, 2007; French et al., 2004). For instance, in Saffran 5

et al.’s (1996) study, 8-month-old infants can successfully extract word boundary information on the basis of the sequential statistics of speech. In their study, infants were exposed to a continuous speech consisting of four three-syllable words. After only two minutes of exposure, they exhibited discrimination between familiar and novel threesyllable strings. It is argued that it is the transitional probability between syllables (high probability within words and low probability between words) that helps infants to discover the word boundaries. Furthermore, sensitivity to statistical regularity was also found in young children’s learning names for new objects (Smith et al., 2002). It was found that the first 300 nouns that young children learn are correlated with the shape (Samuelson & Smith, 1999). Accordingly, researchers found that young children showed “shape bias” in naming objects and extending the labels to the objects sharing the same shape (Smith et al., 2002). The argument is that because most objects young children learn are categorized by shapes, shape becomes a good cue for generalizing a newly learned word to new instances. Moreover, when researchers trained young children to generalize object names by shape, they found that acquisition of new object names outside the laboratory was also accelerated (Smith et al., 2002). Thus, these empirical results suggest that children could extract statistical information from the input, which could guide them in learning the language. Therefore, it is possible that reasoning is not necessary in learning new words while word learning could be achieved by statistical information. In the semantic context, the statistical information will be the semantic association, which is reflected as the association between words. There is evidence indicating that early in development the 6

highly associated words are those having high transitional probabilities (e.g., dog and bark) (Brown & Berko, 1960), whereas taxonomically related words (e.g., cat and dog) do not show high association until later in development (Brown & Berko, 1960). Therefore, the context in which a novel label is introduced can trigger semantic associations, and, as a result, it could provide extra guidance for young word learners to infer the word meaning. For example, in the sentence “look, gavagai has big eyes,” the word “eyes” could activate certain animacy properties and guide children’s word learning. This mechanism differs from the Bayesian approach. First, it does not require a priori assumptions or prior taxonomic knowledge -- semantic associations are acquired directly from the linguistic input. Moreover, because associations between taxonomically-related words do not develop until later development, it is possible that young word learners are guided by semantic associations. For example, the context where the words “furry” and “feeding” are presented could activate animacy features because these words are associated with the word “animal.” However, taxonomicallyrelated words “cat” and “dog” may not initially guide word learning (even thought they both belong to the category of animals) because they are not associated with the word “animal.” Second, word learners are not testing hypotheses, instead word learning is guided by existing semantic associations. Thus, when linguistic input they receive is inconsistent, hypothesis testing may fail, whereas word learning may still succeed. These possibilities were tested in the five reported experiments. Experiment 1 gave us the baseline inductive generalization performance of young children and was to examine whether they could make generalization in the absence of sufficient perceptual 7

information. It was followed by four subsequent experiments, which further investigated the mechanism underlying word learning. The primary goal of these experiments was to examine whether young children’s word learning could be guided by automatic association activated by semantic context and whether the mechanisms were the same across development. In these experiments, participants were presented with a modified word-learning task, in which a novel word was presented in a list of familiar words. Participants were then asked to extend this novel word to either animals or artifacts. Another critical issue to be tested was whether the mechanism of word learning remains the same across development.

8

Chapter 2: Experiments

Experiment 1 Experiment 1 intended to establish the baseline of how young children and adults learn words from semantic context. The experiment had a between-subject design and it had three phases: word learning, label extension, and induction. The word-learning phase was presented as a memory task. It was expected that even young children can successfully use semantic context provided in the word learning phase to learn the novel word. Methods Participants Participants were 30 (13 girls and 17 boys, M = 53.1 months, SD=3.6 months) fouryear-old children recruited from day-care centers located in middle class suburbs of Columbus, Ohio and 30 undergraduates (9 women and 21 men) from the Ohio State University participated in the current experiment for a course credit. In each age group, there were two between-subject conditions, with 15 participants in each condition. Materials and Procedure There are two phases in the experiment: study and test phases. Study Phase: Participants were presented with a word learning task. The materials included an 8cm by 13cm picture of a woman holding an item which was made 9

ambiguous by blurring it in Photoshop 6.0 (see Figure 1). There was also a new count noun a “fep”(or a “dax”) introduced in the study phase. There were two between-subject conditions: (a) baseline and (b) animate object. In the animacy condition, the participants were presented with the picture and told that “This woman is feeding a fep (or a dax),” while in the baseline condition, the participants were told that “This woman is playing with a fep (or a dax).” It was expected that in the animacy condition, participants would infer that the new word refers to an animate object, whereas in the baseline condition, they word could refer either to an animate or inanimate object.

Figure 1.Stimuli used in the study phase of Experiment 1

10

Test Phase: Test phase consisted of two tasks, label extension and induction. In the label extension task, participants were presented four trials with four 3cm by 3cm pictures in each trial. On each trial, there was a familiar animal, a novel animal, a familiar artifact, and a novel artifact (See Figure 2). The positions of the four items were pseudorandom and the four trials were randomized across participants. Participants were asked to choose the one that they thought was a dax (or a fep).

Figure 2. Stimuli used in label extension task in Experiment 1

11

In the induction task, participants were asked eight questions about the properties of a fep or a dax. Four of these properties were biological (e.g., “Does a fep/dax need water?”), whereas the remaining properties were related to artifacts (e.g., “Can a fep/dax be fixed?”). These were forced-choice yes or no questions (See questions in Appendix A) and the order of the questions was randomized. For young children, there was an additional familiarity check at the end of the experiment so as to make sure that children knew the familiar items shown in the label extension task. There were a total of eight familiarity check trials and on each trial participants were presented with a 10cm by10cm picture. The eight pictures consisted of the familiar items, the familiar animal and the familiar artifact, on each trial of the label extension task and children were asked to name the picture. The order of the pictures was randomized. If children could say the name of the item correctly, it was assumed that they knew the item. Children were interviewed individually by a female experimenter in a quiet room in their day-care centers. Undergraduate students were tested in the testing rooms on campus. For all participants, the experiment was administered using Superlab 2.0 software. No feedback was given in any of the tasks. Results and Discussion The results discussed in all the experiments are based on proportion that participants inferred the novel word (fep or dax) denotes an animal but not an artifact.

12

Label Extension Task: As shown in Figure 3, the main effect of condition was significant. Both adults and young children were more likely to map the novel label to the animate items in the animacy condition (for young children, M = 85%, SD = .21, for adults, M = 80%, SD = .36), where participants were told that “this woman was feeding a dax,” than in the baseline condition (for young children, M = 47%, SD = .41, for adults, M = 8.3%, SD = .26), where participants were told that “this women was playing with a dax,” F (1, 56) = 44.644, p < .01. The interaction between age group and the condition was also significant, F (1, 56) = 4.10, p < .048. More specifically, young children were equally like to extend the novel label to either the animate items or artifacts in the baseline condition, one-sample t (14) = -.315, p > .75, at chance. However, adults showed the artifacts bias in that condition, t (14) = -6.17, p < .01.

13

Proportion of Animal Response

1

*

*

0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Children

Adults

Figure 3. Proportion of label extension responses by age group and condition in Experiment 1.

Induction Task: The results in this section are presented in Figure 4 and they reflect the proportion of participants’ inferences that the object labeled by the novel word had animate properties. Similar to the label extension task, the main effect of the condition is also significant for the induction task, F (1, 56) = 19.253, p < .01. Both adults and young children were more likely to interpret the novel word as animate items and more likely to attribute animate properties in the animacy condition (for young children, M = 67%, SD = .11, for adults, M = 75%, SD = .31), for instance, “A dax needs water,” than in the baseline condition (for young children, M = 54%, SD = .24, for adults, M = 7.5%, SD = .26), for instance, “A dax is made by people.” 14

The interaction between age group and condition was also found for the induction task, F (1, 56) = 19.253, p < .01. Adults showed the artifact bias (t (14) = -6.38, p < .01), whereas young children did not (t (14) = .658, p > .521).

1

Proportion of Animal Response

0.9 0.8

* *

0.7 0.6

Baseline

0.5

Animacy

0.4 0.3 0.2 0.1 0

Children

Adults

Figure 4. Proportions of induction responses by age group and condition in Experiment 1.

Taken together, the results of Experiment 1 indicate that both young children and adults could map the novel word to animate objects by being told that “something was fed.” However, the underlying mechanisms remain unknown. It is not clear whether the mechanism underlying young children’s performance is the same as that underlying adults’ performance. It is also not clear what kind of knowledge help both young children 15

and adults mapping the novel word. One possibility is that young children were reasoning or testing hypotheses, and used conceptual taxonomy to help them complete the task. In this situation, when young children were told that “something was fed,” they would infer that this novel object should have some animate properties, and it should be an animal, so that the novel object should have other animal properties. Another potential explanation is that young children complete the task not via conceptual knowledge but via automatic association. Because the properties like “be fed” “need water” “animate objects,” or so forth, usually happen in the same context, the words like “feeding” trigger the other properties associated with animals automatically. However, these explanations could not be teased apart in this experiment and thus, Experiment 2 was conducted. The goal of Experiment 2 was to address the question whether there are changes of the underlying mechanism across development.

EXPERIMENT 2 The goal of Experiment 2 was to examine whether word learning in Experiment 1 was due to automatic association activated by the semantic context or participants were using conceptual knowledge to test hypothesis, and also to investigate whether the mechanism underlying the performance was the same across development. Label extension would still be used in test phase. However, the word learning part in Experiment 1 was changed to word list learning, which was introduced as a memory task. There were three betweensubject conditions: taxonomic, associative, and baseline. In the taxonomic condition, participants were asked to memorize words, all of which were familiar animals. In the 16

associative condition, the words were semantic associates of the word “animal” (the list included a combination of animals and artifacts), and they were expected to activate the word “animal.” In the control condition, the words on the list were neither animals nor semantic associates of the word “animal.” The hypotheses were that if participants use taxonomic knowledge to guide their word learning, they should successfully extend novel word to animals in the taxonomic condition. This is because all the words on the taxonomic list were animals, and participants were expected to understand that and to subsequently infer that the novel word refers to an animal. However, if they rely on semantic associations rather than on reasoning or hypotheses testing, they should infer that the novel word refers to animal in the associative condition. Methods Participants Participants were 58 four-year-old children (28 boys and 30 girls, M= 52.8 months, SD=3.6 months) recruited from day-care centers located in middle class suburbs of Columbus, Ohio. In addition, 45 undergraduates (22 women and 23 men) from the Ohio State University participated in the current experiment for a course credit. In each age group, there were 3 between-subject conditions. There were 15 adult participants in each condition. There were also 21 child participants in the associative condition, 22 child participants in the taxonomic condition, and 15 child participants in the control condition. Material and Procedure Study Phase: Word learning was introduced as a memory task and it included three between-subjects conditions: Taxonomic, Associative and Baseline. The experimenter 17

read a word list to the participants twice and participants were told explicitly to remember the words. There were nine words in the list, eight of which were familiar words (for young children familiarity of each word established by consulting the MRC Psycholinguistic Database -- http://www.psy.uwa.edu.au/mrcdatabase/uwa_mrc.htm). There was also a novel word (either fep or dax) added to each list. In the Taxonomic condition, the familiar words (e.g., “cat,” “dog,” “fish,” “bird,” etc.) belonged to the category of animal. In the Associative condition the familiar words (e.g., “zoo,” “farm,” “furry,” etc.) had forward semantic association with the word “animal”, according to the Edinburgh Word Association Thesaurus (http://www.eat.rl.ac.uk/). In Baseline condition, the familiar words (e.g., “red,” “blue,” “black,” etc.) were color words. They did not belong to the category of animal, nor had forward semantic association with the word “animal.” Word lists, semantic association strength of each word with the word “animal,” and the summed associative strengths by age group and condition were presented in Appendix B, C, and D. The associative strength in the current study was the defined as proportion of one response, for example, “animal” out of all the other responses when participants were given the target word, for example, “cat,” in the free association task. The associative strengths for the adults were taken from Edinburgh Word Association Thesaurus (http://www.eat.rl.ac.uk/), whereas those for young children were collected by 50 4-year-olds by a free association task, where the experimenter read each word to the participants and the participants were told to say the first word they thought of when they heard the word said by the experiment.

18

Test Phase: The same label extension task as in Experiment 1 was used in Experiment 2. However, prior to performing this task, they were asked to recall the words they had heard in the study phase as accurately as possible. Results and Discussion Results of performance on the label extension task are presented in Figure 5. In both the Associative and Taxonomic conditions, adults were likely to extend the novel word to an animal, (83% and 88%, respectively, both exceeding chance performance, one-sample ts> 3.6, ps< .005). A one-way ANOVA confirmed that adults learned the novel word in both Associative and Taxonomic conditions, with both conditions being above the Control condition, F (2, 43) = 3.16, p < .05, with the Control condition eliciting chance performance, (58%, one-sample t = .702, p > .494). Similar to adults, young children reliably extended the novel word to an animal in the Associative condition (70%, above chance, one-sample t (20) = 2.193, p < .040). However, in contrast to adults, their extension of novel words to animals in the Taxonomic condition (40%) did not exceed chance (p > .29) and it was reliably below the Associative condition, independent-sample t = 2.292, p < .027. These results pointed to several findings. First, the results that young children successfully extended the novel label to animals in the Associative condition indicated that early word learning could be achieved by an automatic associative process. Furthermore, when young children learned a word, it was unlikely that they relied on reasoning or hypotheses testing, as evidenced by their failure in the Taxonomic condition. Second, it also suggested that the Bayesian word learning could not account for this 19

failure. Specifically, in the Taxonomic condition, all the words referred to animals, while in the Associative condition, some words did not (e.g. zoo, furry, feeding, and farm). If young children relied on Bayesian inference to test their hypotheses, they should be able to extend the novel word to animals in the Taxonomic condition. However, the results contradicted the prediction. Finally, the results also pointed to developmental changes in word learning. Whereas young children succeeded only in the Associative condition, adults succeeded in both, the Taxonomic and Associative conditions. These findings suggest that while early word learning is likely to be driven by an associative process, later in development word learning could be achieved by hypotheses testing means. To replicate the results of the Associative condition, while further testing whether label extension was sensitive to the Bayesian likelihood, Experiment 3 was conducted, in which the associative list was shortened by leaving only the strongest semantic associates of the word animal, none of which referred to an animal (zoo, farm, furry, and feeding). This shortening should result in a decrease of the Bayesian likelihood that the novel word referred to an animal.

20

Proportion of Animal Responses

1

**

0.9

*

*

0.8 0.7 0.6

Control Taxonomic Associative

0.5 0.4 0.3 0.2 0.1 0 Children

Adults

Figure 5.Proportions of label extension responses by age group and condition in Experiment 2.

EXPERIMENT 3 Experiment 3 was to further test the hypothesis of the Bayesian approach (Xu & Tenenbaum, 2007). According to the Bayesian approach, the posterior probability of a hypothesis being true given the observation is proportional to the likelihood given the hypothesis (Xu & Tenenbaum, 2007). For example, if fep refers to dogs (Hypothesis A), the likelihood of observing an object being called as fep will be 1/(the number of all the dogs). If fep refers to mammals (Hypothesis B), the likelihood of observing an object being called as fep will be 1/(the number of all the mammals). As there are more mammals than dogs in the world, the first likelihood will be bigger than the latter one, so that the posterior of Hypothesis A, which assumes that fep refers to dogs, will be larger 21

than that of Hypothesis B, which assumes that fep refer to mammals. In another word, that fep refers to dogs is more like to be true. As the number of observation increases, the difference between the two hypotheses will increase exponentially. Similarly, in the current task, if there are more words associated with the word “animal” in the list, the more likely the novel words is also associated with the word “animal” by using the Bayesian inference. Thus, young children will more likely to extend the novel word to animals in the condition with longer list than shorter list. However, if young word learners are not testing hypotheses by the Bayesian approach or if they are not testing hypotheses at all, the number of associated words in a list should be irrelevant but the sum of associative strength would be critical. Accordingly, if we decrease the number of associative words in the memory list without decreasing the sum of associative strength significantly, and if young children’s word learning performance is not affected by the number of words, it will be the evidence that children can map novel label to the objects in context by automatic association instead of the Bayesian hypothesis testing. Moreover, Experiment 3 would be a better control than Experiment 2 as only the words that did not belong to the category of animals would remain in the associative list. Thus, word learners could not use any taxonomy knowledge in this condition. However, if any word learning occurred, it would not be explained by structured taxonomy knowledge but could be accounted by automatic semantic association.

22

Methods Participants Participants were 13 4-year-old children (M = 54.69 months, SD = 3.95 months, 10 boys and 5 girls) recruited from day-care centers located in middle class suburbs of Columbus, Ohio. Materials and Procedure Study Phase: Study phase was similar to that in Experiment 2. However, only 4 of the familiar words in the associative list were used in the memory task. The 4 words were those with the highest associative strength with the word “animal” but none of the words belongs to the category of animal. These words were furry, feeding, zoo, and farm. The total associative strength of the list equaled to 0.75 for children and 0.93 for adults. The novel word “dax” (or “fep”) was also added to the word list Test Phase: Tasks used in the test phase were the same as those used in Experiment 2. Results and Discussion The results were similar to those in Associative condition of Experiment 2: young children reliably extended the new word to an animal, 71%, above chance, one-sample t (12) = 2.17, one tail p< .05, d = 0.60. These findings replicate and further extend findings of Experiment 2. The associative list was shortened markedly (with all animal names being removed from the associative list) which should have further reduced the Bayesian likelihood. At the same time, the summed associative strength in Experiment 2 (0.75) was comparable to that in Experiment 1 (0.84). The reduction of the Bayesian 23

likelihood did not affect participants’ responses, with participants reliably using associative information in their label extension. Therefore, across Experiments 2-3, the results suggested that early word learning was sensitive to associative strength rather than to the Bayesian likelihood.

EXPERIMENT 4 The goal of Experiment 4 was to further investigate whether Bayesian approach could be the potential explanation of young children’s performance in Experiment 3. It is possible that young children are still testing hypothesis by Bayesian approach. But their hypothesis is not whether all the words coming from the same category, e.g., animal in current study, but is whether all the words are semantically associated with the word “animal.” In Experiment 3, though the word list had been shortened, all the four words in that list were positive examples of the theme “animal.” Thus, young children could still rely on reasoning that was based on thematic rather than on taxonomic knowledge. If this is the case, young children’s performance should drop if words do not represent a particular theme. For example, if a word list represents of a mixture of themes, it should be difficult to infer on the basis of hypothesis testing that the novel word refers to an animal. However, if they learn the novel word by automatic association, their performance should not drop as long as the sum of associative strength remains high. In other words, if young children are not testing hypothesis and they learn new words via automatic association, it does not matter to young word learners whether the list of words has a consistent theme. The only factor that could contribute to young children’s word 24

learning performance is the sum of associative strength. However, if they are using Bayesian approach to test hypothesis, they should be unable to learn the novel word if the list is mixed. The goal of Experiment 4 was to test these two possibilities. Methods Participants Participants were15 4-year-old children (M = 55.3 months, SD = 3.099 months, 9 boys and 6 girls) recruited from day-care centers located in middle class suburbs of Columbus, Ohio. Materials and Procedure Study Phase: Study phase was similar to that in Experiment 3. However, in addition to the 4 words used in Experiment 3, 4 more neutral words (color words) were added to the word list. Therefore, the sum of associative strength of the list, 0.75 for children and 0.93 for adults, remained the same as that in Experiment 3, but the percentage of associative words drops from 100%, indicating all the words were associated with the word “animal,” to 50%, indicating only half of the words had the association. The novel word “dax” (or “fep”) was also added to the word list. Test Phase: Tasks used in the test phase were the same as those used in Experiment 3. Results and Discussion As predicted by the automatic association account, despite the drop in percentage of the words associated with animal in the list, participants were still reliably extending the novel words to animal pictures, 77%, t (14) = 2.543, above chance, p < .012 (one-tail test). The performance here was comparable to that in the condition of Experiment 3 25

where only 4 strong associative words were presented in the study phase. Therefore, little evidence showed that children were using Bayesian approach to test the hypothesis. However, considering the same sum of associative strength cross the two experiments and the comparable performance, it might be inferred that it is the sum of associative strength with the word “animal” that triggers the label extension to animal pictures. In sum, Experiment 1 through 4 suggested that young children could extend the novel label to the animal objects in the context activating the word “animal.” However, it is also possible that the context not only activates the word “animal” but could also activate some salient feature of animacy. For example, Jones, Smith, & Landau (1991) demonstrated that the presence of eyes is a highly salient feature of animacy and therefore it is possible that labels are more likely to be extended to objects with eyes than to objects without eyes. Experiment 5 was intended to test this possibility.

EXPERIMENT 5 Experiment 1-4 indicated that reasoning is not necessary for learning new words, and that the context in which words are presented can automatically activate certain features that could provide guidance to infer the meaning of the new word. If the context could activate the semantic meaning of “animal,” it is possible that the context might have also activated other features related to animacy, for example, the presence of eyes. If this is the case, then presence of this animacy feature alone could be sufficient enough for label extension. To answer this question, novel stimuli were created for label extension task. Instead of using real animals or artifacts, Experiment 5 used geometric shapes with eyes 26

or wheels to point to the animal or the artifact categories, respectively. Eyes were used because there is evidence suggesting that they were a perceptual feature that could be used to predict animacy and were learned in early development and objects, even artifacts, with eyes are usually perceived as animate objects (see Smith and Landau, 1991 for a review). If a context could activate salient features associated with animacy, then the presence of a particularly salient feature of animacy (i.e., eyes, in this experiment) at test should be sufficient for successful label extension. Methods Participants Participants were 15 4-year-old children (M = 54.06 months, SD = 3.68 months, 6 boys and 9 girls) recruited from day-care centers located in middle class suburbs of Columbus, Ohio. Materials and Procedure Study Phase: The study phase was similar to Experiment 2 but only with the associative condition. The word list included 8 familiar associative words and a novel word “fep” or “dax” was added. Test Phase: Label extension in Experiment 5 was similar to previous tasks with one exception. In the label extension task, instead of using real animals and objects, artificial stimuli were created for Experiment 5. These stimuli were created via adding eyes or wheels to geometric shapes. For each trial, there were still four choices for participants, two for the “animal category” and two for the “artifact category” The “animal” stimuli had the same shape as the corresponding “artifact” stimuli. However, eyes were added to 27

the stimuli for animal choices but wheels for the artifact choices. Example is given in Figure 6.

Figure 6.Stimuli used in lexical extension task in Experiment 5

Results and Discussion The results were similar to that of associative condition in Experiment 2. Young children were more likely to extend the novel words to geometric shapes with eyes, 70%, t (14) = 2.449, above chance, one tail, p< .02. Therefore, a single critical feature associated with animacy (i.e., the presence of eyes) was sufficient for successful label extension.

28

Chapter 3: General Discussion Several important regularities emerge from reported experiments. Experiment 1 indicated that similar to adults, young children extract the meaning of the word from the context in which this word is presented. Experiment 2 indicated that whereas adults can rely on associative and taxonomic information when learning new words, young children can rely only on associative information. Experiment 3 indicated that reducing the number of associative words in the learning phase but not the sum of associative strength would not affect the probability of extending the novel label to animals for young children. The comparable sums of associative strength would elicit comparable performance. Experiment 4 indicated that mixing the animal words with neutral color words did not decrease the performance either. Experiment 5 indicated that the semantic associative context could not only guide word learning but could also trigger some associated features. These results suggest that young word learners could learn new words from context, in which the semantic association could activate certain features related to the novel word. Furthermore, they could do so without relying on reasoning or hypotheses testing. First, young children did not exhibit evidence of taxonomic assumption when they learn new words. The taxonomic assumption predicts that young children will assume a novel label refers to certain taxonomic category, and thus if all the exemplars they perceived are from the same category, they will use this information guide their word learning and 29

infer that the novel label refers to a certain objects from the same category. According to that prediction, young children should successfully extend the novel label to animals in the taxonomic condition of Experiment 2. However, they did not. It seems that the taxonomic assumption could not account for the findings well. Second, it is not clear how the Bayesian approach of word learning could account for the findings, though the Bayesian approach could be considered as another form of the hypothesis testing theory. According to Bayesian approach, the posteriors of hypotheses being true given the data depend on both the priors and likelihoods. In the reported experiments, the priors of hypotheses were not manipulated, but the likelihoods were, for example, whether the word lists were long or short, or whether all the words in one list belonged to the same category. However, children’s word learning performance, the probability of extending the novel word to animals, did not change in either of the situations. For example, two of the possible hypotheses in this study could be: dax is an animal (H1) or dax is any object in the world (H2). Suppose there are total M animals and N objects in the world, where M is smaller than N, since animals are in the category of all the objects). If young children have the assumption that all words in the same word list should from the same category, they should be more likely to interpret dax as an animal in taxonomic condition in Experiment 2 because there are 8 other words coming from the same category and likelihood of H1 (dax is an animal) would be (1/M)8 which is larger than the likelihood of H2 (dax is any object). In this case, it would be (1/N)8. However, in associative condition, likelihood of H1 is not available as some words were in the category of the animals while the others were not. Only H2 could be selected. Thus, they 30

should randomly interpret dax as an animal or an artifact. However, the opposite was found. It is also possible that young children did not assume that all the words in the same list came from the same category, assuming instead that they came from the same topic. This assumption could explain why they were more likely to interpret dax to be an animate objects in the associative condition in Experiment 2. However, while this assumption could explain performance in Experiment 2, it still could not explain performance in Experiments 3-4. In Experiment 3, the word list was shortened to 4 words. According to the Bayesian approach, that should negatively affect children’s inference that dax is an animate object. Because if we compare the ratio of the likelihood of H1 and H2 in Experiment 2 ((1/M) 8:(1/N)8) and Experiment 3 (((1/M)4: (1/N)4)), it could be found that the ratio in Experiment 2 was higher. And thus, it would be more reasonable to predict that in Experiment 3, H1 would less likely to be selected over H2 and participants would less likely to interpret das as animal than in Experiment 2. However, it was still not what was found. In fact, the length of word list did not have an impact on the probability that young children extend the novel label to animate objects. Similar results were found when the associative words were mixed with neutral color words. In this situation, H2 would be more likely to be selected in Experiment 4 as H1 would not be true. However, it was still not what was found. Therefore, the Bayesian approach of word learning did not seem to provide a good account for the findings in the current study. In contrast, the reported finding could be readily explained by the sum of associative strength of the word

31

“animal.” In Experiment 2, Experiment 3, and Experiment 4, the sum of associative strength did not change significantly, and neither did the word learning performance. In sum, a number of important findings stem from the reported experiments. First, young children can extract the meaning of the word from the context in which this word is presented and can rely on associative information. Second, hypothesis testing is not necessary for young word learners to learn new labels when they can rely on semantic associations. Third, young children’s word learning performance will not change significantly if the semantic associative strength is not reduced significantly. This could not be explained by the Bayesian approach, which focuses on the number of positive examples rather than the sum of associative strength. And finally, the semantic associative context could also trigger some associated features. Thus, reasoning or testing hypotheses is not necessary for word learning early in development.

32

References Balaban, M. T., & Waxman, R. (1997). Do words facilitate object categorization in 9month-old infants? Journal of Experimental Child Psychology, 64(1), 3-26. Booth, A. E., Waxman, S. R., & Huang, Y. T. (2005). Conceptual information permeates word learning in infancy. Developmental Psychology, 41(3), 491-505. Brown, R., &Berko, J. (1960). Word association and the acquisition of grammar.Child Development, 31(1), 1. Retrieved from Academic Search Complete database. Colunga, E., & Smith, B. (2005). From the lexicon to expectations about kinds: a role for associative learning. Psychological Review, 112(2), 347-382. French, R. M., Mareschal, D., Mermillod, M., & Quinn, P. C. (2004). The role of bottom-up processing in perceptual categorization by 3- to 4-month-old infants: simulations and data. Journal of Experimental Psychology: General, 133(3), 382397. Gelman, S. A., & Coley, D. (1990). The importance of knowing a dodo is a bird: Categories and inferences in 2-year-old children. Developmental Psychology, 26(5), 796-804. Gelman, S. A. (2004).Psychological essentialism in children.Trends in Cognitive Sciences, 8(9), 404-409. Markman, E. M., & Wachtel, F. (1988). Children's use of mutual exclusivity to constrain the meaning of words.Cognitive Psychology, 20(2), 121-157. Murphy, G. L. (2004).The big book of concepts. Cambridge, MA: MIT Press. Quine, W. V. (1960) Word and object.Cambridge, MA: MIT Press Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants.Science, 274(5294), 1926-1928. Sloutsky, V. M., & Fisher, V. (2004). When Development and Learning Decrease Memory: Evidence Against Category-Based Induction in Children. Psychological Science, 15(8), 553-558. 33

Samuelson, L. K., & Smith, B. (1999). Early noun vocabularies: Do ontology, category structure and syntax correspond? Cognition, 73(1), 1-33. Smith, L. B., Jones, S. S., Landau, B., Gershkoff-Stowe, L., & Samuelson, L. (2002). Object name learning provides on-the-job training for attention. Psychological Science, 13(1), 13-19. Jones, S. S., Smith, L. B., & Landau, B. (1991). Object properties and knowledge in early lexical learning. Child Development, 62(3), 499-516. Welder, A. N., & Graham, A. (2001). The influences of shape similarity and shared labels on infants' inductive inferences about nonobvious object properties.Child Development, 72(6), 1653-1673. Xu, F., & Tenenbaum, B. (2007). Sensitivity to sampling in Bayesian word learning.Developmental Science, 10(3), 288-297. Yu, C., & Smith, B. (2007). Rapid word learning under uncertainty via cross-situational statistics.Psychological Science, 18(5), 414-420.

34

Appendix A: Questions used in induction task in Experiment 1 1. Is a dax made by people? 2. Can a dax be repaired? 3. Can a dax be fixed? 4. Can a dax be packed? 5. Does a dax need water? 6. Can a dax breathe? 7. Can a dax move itself? 8. Is a dax alive?

35

Appendix B: Associative Strength of the word “animal” for both adults and children (Taxonomic List) Children

Adults

Cat

0.00

0.02

Dog

0.00

0.01

Fish

0.00

0.01

Bird

0.00

0.00

Horse

0.00

0.04

Squirrel

0.00

0.04

Cow

0.00

0.02

Rabbit

0.00

0.01

Summed Associative

0.00

0.15

Strength

36

Appendix C: Associative Strength of the word “animal” for both adults and children (Associative List) Children

Adults

Creature

0.09

0.26

Farm

0.20

0.19

Giraffe

0.00

0.09

Hamster

0.00

0.13

Bear

0.00

0.09

Furry

0.07

0.21

Feeding

0.09

0.02

Zoo

0.39

0.51

Summed Associative

0.84

1.5

Strength

37

Appendix D: Associative Strength of the word “animal” for both adults and children (Control List) Children

Adults

White

0.00

0.00

Red

0.00

0.00

Blue

0.00

0.00

Purple

0.00

0.00

Pink

0.00

0.00

Black

0.00

0.00

Green

0.00

0.00

Yellow

0.00

0.00

Summed Associative

0.00

0.00

Strength

38