Word-based categorization in 14 to 16 month-old infants : evidence from a Slavic language

We investigate early word-based categorization by testing one sample of infants at the ages of 14 and 16 months using an object manipulation task. Specifically, we investigate the facilitating effect of labeling on: a) categorization of unknown objects with a shared label and a similar shape; b) categorization of unknown objects with a shared label, but no perceptual similarity. The results indicate that an average infant can categorize novel objects with a similar shape and a shared label starting at 16 months of age. Categorization of novel objects with a shared label but no perceptual similarity is at chance level at both ages. These results confirm and extend previous research findings to a new language.

There is a plethora of evidence that very early, already during the first months of their lives, babies start developing the ability to parse the world around them into rudimentary categories (Quinn, 2002;Rakison & Oakes, 2003;Sloutsky & Fisher, 2011).For example, results obtained by the familiarization/ novelty-preference procedure (see Quinn, 2002, for a detailed description) indicate that, by the age of three to four months, infants can form separate and mutually exclusive categorical representations for cats and dogs (Arterberry & Bornstein, 2001;Quinn, 2002;Rakison & Oakes, 2003).Although the role of language in categorization has been much debated (see, for example, Sloutsky & Fisher, 2011;Waxman & Gelman, 2009) a facilitating effect of language on categorization was demonstrated already at the age of three months (Ferry, Hespos, & Waxman, 2010).However, on more demanding tasks that include real object manipulation, categorization of novel objects, and socio-pragmatic cues to be taken into account, word-based categorization becomes visible later, different objects in several different situations.Thus, a child can hear the adults use the word "cow" when looking at a) a very big, brown, animate object, b) a two-dimensional purple picture on the chocolate wrapping, and c) a little blue plush toy.Is it the word "cow" that stimulates the child to look for the common properties of these quite different objects (Waxman & Markow, 1995;Waxman & Braun, 2005)?
Already at the age of three months, that is, long before infants actually begin to understand or use language manifestly, words come into scene as a tool for categorization (Ferry, Hespos, & Waxman, 2010).A number of studies have reported that, from the age of six months on, presenting different stimuli with the same label promotes categorization in infants by turning their attention to the common properties of the objects in question (Balaban & Waxman, 1997;Booth & Waxman, 2002;Fulkerson & Waxman, 2007;Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007;Rakison & Oakes, 2003;Waxman, 2002).The facilitating effect of labeling on categorization gradually becomes apparent towards the end of the first year, and by the end of the second year language becomes one of the most powerful means for categorization (see, for example Balaban & Waxman, 1997;Booth & Waxman, 2002;Fulkerson & Waxman, 2007;Plunkett, Hu, & Cohen, 2008;Waxman, 2002 etc.; but see also Sloutsky & Fisher, 2011 for the opposite view).
Most of the above-cited studies on early categorization are based on the preferential looking paradigm and use familiar objects as stimuli, thus raising some important methodological issues such as: ecological validity vs. experimental control, category knowledge prior to the experiment and the importance of social referencing (Cohen & Brunt, 2009).In a series of studies that is of particular interest for our research, an object manipulation task was used (improving the ecological validity of the experiment) to examine the effect of labeling on categorization of novel objects (assuring no category knowledge prior to the experiment) in a game-like interaction with a real person (allowing social referencing) (Booth & Waxman, 2002;Nazzi, 2005;Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007).These experiments show that, on an object manipulation task, infants can categorize two new objects presented with the same label from the age of 16 months, but only if these objects share at least some perceptual or functional features.It is only at 20 months that infants can categorize two new objects in the absence of any common perceptual or functional feature, based on the shared label only (Booth & Waxman, 2002;Nazzi, 2005;Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007).These studies have also suggested that there might be a specific link between the vocabulary size and word-based categorization abilities during the second year of life.

Overview of the experiment
The main goal of the current study was to further explore early categorization abilities in a new linguistic and cultural setting.In order to assure that any observed differences between the age of 14 and the age of 16 months would be due to the age effect alone, and not the cohort effect we decided to use the longitudinal design (Teti, 2005, p.8).Our focus was on categorization "in use", which occurs later than categorization evidenced by looking-time measures alone (e.g.Ferry, Hespos, & Waxman, 2010;Plunkett, Hu, & Cohen, 2008).For this purpose, we used an object manipulation task in a game-like interaction with a real person and a procedure quite similar to that used by Nazzi & Gopnik (2001) and Nazzi & Pilardeau (2007).Mothers' estimates of children's productive vocabulary were collected in order to examine the potential link between the vocabulary size and the ability to use names to categorize things (Nazzi & Gopnik, 2001).
There was, however, an important difference in our procedure.We performed the first testing session at 14 months.After that session, mothers were instructed to practice the experimental procedure with the infants using the same task, but in a familiar context, with familiar objects.First, in this way we made sure that infants at 16 months would understand the instructions.Second, such procedure ensured that infants at 16 months would be familiarized with the task, while the objects to be categorized and their labels would be completely unknown to them.Specifically, we investigated the facilitating effect of labeling on: a) categorization of unknown objects with a similar shape and a shared label; and b) categorization of unknown objects with a shared label, but no perceptual similarity.Testing was performed at home, in a familiar socio-pragmatic context, on a familiar task, thus maximizing the ecological validity of the experiment.
At 14 months of age performance on both categorization tasks (unknown objects with a similar shape and a shared label and unknown objects with a shared label, but no perceptual similarity) will be at chance level.2.
At 16 months of age performance on the task of categorization of unknown objects with a similar shape and a shared label will be above chance level.Performance on the task of categorization of unknown objects with a shared label, but no perceptual similarity will be at chance level.
Our expectations with regard to the longitudinal design did not differ from the expectations in a cross-sectional design.Although we used alternate forms of testing material to avoid the practice effect, it was not likely to be found in infant population given the limitations in long-term memory at this age (Teti, 2005, p.8).For the same reason we expected the familiarity with the task only to minimize the effect of infants' failure to understand the instructions, rather than to improve their categorization abilities.

Participants
Thirty-one fourteen-month-old infants (range 13;25 to 14;15) from monolingual Serbian families in Bosnia & Herzegovina (regions of Banja Luka and Derventa) were recruited for this study.13 boys and 18 girls were included in the final sample.The same infants were re-tested at 16 (15;25 to 16;15) months.Three infants refused to participate in the task at 16 months for different reasons (fusiness, refusal to follow the instructions, illness), so the sample at 16 months consisted of 13 boys and 15 girls.
Infants were recruited through the listing of the Banja Luka Center for preschool education and through participating mothers' networks of acquaintances.All infants came from Bosnian middle-class backgrounds, from complete families with one or both parents employed full-time.
Two triads of familiar objects (plush toys, balls, blocks, etc) were used in the familiarization phase.In the testing session, six triads of small objects were chosen so that the infants would be unfamiliar with them and would not have a name for them.By designing our own stimuli we made sure that infants were not familiar with categories, and that any categorization occured during the experiment.Three of the sets were made up of a pair of similar-shaped objects and a third object that differed from the other two in shape, color, and texture (see Figure 1).The remaining three sets were made up of three disctinct objects, that differed in shape, color, and texture (see Figure 2).
The objects were given the following made-up labels: chook/dal, moz/djig, vab/ newk,nol/cag, jed/leef, zoop/teech 1 .Most of these pseudowords had previously been used in Bijeljac-Babic et al. ( 2009), but were adapted to conform with Serbian phonological and phonotactical rules.We used six pairs of very distinct monosyllabic pseudowords in order to maximally reduce the cognitive load associated with the task.

Procedure
Each infant is tested individually, at home, in a quiet atmosphere.Infants are either seated on the floor, or in a feeding chair.Mothers are instructed to stay close to the child and silently watch the session.They are told to put the child in the lap if needed, but not to help or interfere in any way with the child's response.Provided that the experimenter obtained consent from the parents the experiment is recorded with a camera.
During the familiarization phase, the experimenter shows the infant two toys of the same kind, e.g. two different rubber ducks, and a third, different toy, e.g. a plastic dog.The experimenter demonstrates the task to the infant in the following way: "Look what I've got, it's a duck!Here, take the duck."The infant is allowed to play with the toy until the experimenter says: "Look what else I've got, it's a dog!Here, take the dog."The infant is allowed to play with the second toy.Then, the toys are put on the table, and the experimenter says: "Look what I've got now, it's another duck!Look carefully now!I'm going to put my duck in the basket.(the experimenter puts the duck in the basket) Which one will go with the duck?Look, another duck!(the experimenter puts the second duck in the basket) One duck and the other duck, they go together."The second familiarization trial is done following exactly the same procedure, but using three familiar, inanimate objects (e.g. a clothes peg, and two differentcolored blocks).After the demonstration, the task is repeated with the infant.The first two steps of the trial are the same, whereas in the third step, instead of putting the second toy in the basket, the experimenter asks the infant: Which one will go with the duck?Put it in the basket!".If the infant puts the right toy in the basket, the experimenter says: "Good, one duck and the other duck, they go together."and passes on to the next familiarization trial.If the infant puts the wrong toy, both toys, or no toy in the basket the experimenter corrects him saying: "No, it's the duck that goes with the duck -duck and duck go together!".The familiarization trial with the infant is repeated up to three times before passing on to testing.
The testing session contains six trials with only a pass/fail option, that is, the experimenter passes on to the next trial regardless of the infant's success on the task.The procedure in the testing phase is the same, except that all objects and their labels are unknown to the infant (see the Stimuli section).Each object name is repeated six times and the names are embedded in carrier sentences.The testing session is completed after six trials.If an infant continously fails to complete the task (e.g.playing with the toys without putting them in the basket), or if he gets fussy, the session ends immediately.
After the session, mothers are instructed to practice this "game" with the infants during the following two months, two times a week.They are left with written instructions on how to play "the game" using familiar objects such as toys, and they are asked to always use the same final sentence: "Which one will go with the _____".
The entire procedure is repeated with the same infants at 16 months of age.Following each testing session, mothers fill out the productive vocabulary list of the Croatian version of the Mac Arthur CDI Communication Development Inventory (Fenson et al., 1993) 2 , adapted to Serbian language for the purpose of this study.

Design
The experiment consisted of two familiarization trials and two testing conditions with three trials each.The first condition was the similar condition, where infants had to choose the object with a similar shape and a shared label with the target object.The second condition was the dissimilar condition, where infants had to chose the object with a shared label, but no perceptual similarity with the target object.
The similar and the dissimilar sets were presented in an alternating order, always starting with the similar trial in order to encourage the child by first presenting the easier task.After that, the order was alternated so so as to avoid fatigue as a potential confound if the dissimilar trials had always been the three last ones.To control for the possible effect of hand preference and order of stimulus exposition, the order of presentation of the objects (from left to right and vice versa) and the side of the object to be picked up by the infant were counterbalanced across trials.In order to avoid potential perceptual bias in selecting the objects, the pairs defined by the made-up names were counterbalanced across subjects.

Statistical analysis
Children's responses were scored as 1 for the correct answer and 0 for the wrong answer.Thus we calculated the sum scores on familiarization tasks (0 to 2), the similar condition (0 to 3) and the dissimilar condition (0 to 3).These scores are restricted count measures; consequently, they were treated as both categorical and numerical variables, as was the case in previous similar research (Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007).In order to obtain comparable results we used chi-square tests and appropriate confidence intervals to examine whether at least 50% of the specific age population could solve the particular task.To test for within-subject difference in performance between 14 and 16 months we used McNemar's Chi-square (exact p-values).In addition, we tested whether an average infant's performance at the specified age was above chance level by using one-sample t-test.Two-factor mixed ANOVA was used to explore the effect of condition (within-subject factor: similar or dissimilar) and gender (between-subject factor) on the success in the testing phase.
The vocabulary score was calculated by summing up the number of words an infant can produce at a given age based on the inventory filled out by the mother.Important individual differences in vocabulary rates, which are expected at this age (Fenson et al., 1993), produced a positively skewed score distribution on the CDI.This is why we used Spearman's rank correlation coefficients to test for the association of the vocabulary rate and the task effectiveness.

Results
Familiarization phase 14 months.One sample t-test revealed that, on average, infants at 14 months choose the right object (the object that belongs to the same basic-level category and has the same name as the target object) significantly above chance level (M = 1.54,SD = 0.74, t(27) = 3.8, p <.001) when familiar objects such as toys are used as stimuli.As shown in Table 1, 19 out of 28 infants chose the right object in both familiarization tasks at 14 months.We performed additional analyses by grouping the infants into two categories: those who succeeded on less than 50% of the trials (0 or 1 correct answers out of 2) versus those who succeeded on more than 50% of the trials (2 out of 2 correct answers).However, the chi-square analysis revealed that the distribution of the 14-month-olds was not different from chance (χ2(1) = 3.57, p = .09).Specifically, in our sample 67.9% of infants chose the right object in both familiarization trials, but the 95% confidence interval range [49.2%, 82.2%] suggests that this is really a transition period in which we could reasonably assume that at least 50% of infants have already mastered this capacity while others might still be in the process of learning it.16 months.As can be seen in Table 1, infants were clearly more successful on the familiarization task at 16 months, at which point 8 infants performed better than they had at 14 months (p = .039).Both by comparing mean values of correct answers (M = 1.93,SD = 0.26, t(27) = 18.73, p <.01) and by comparing the number of infants who succeeded on more vs.less than 50% of the trials (χ2(1) = 20.57,p <.001, 95% CI [76.2%, 99.1%]) we come to conlusion that categorizing two familiar objects with a shared label is an easy task for an average 16-month-old.14 months.The categorization task consisted of 6 trials, 3 for the similar and 3 for the dissimilar condition.As depicted in Table 2, there are no infants at 14 months who succeeded in more than 1 trial per condition.As much as 75% of the infants scored 0 in the similar condition and 92.9% scored 0 in the dissimilar condition.

Testing phase
One sample t-tests revealed that infants at 14 months chose the right object, that is the object that has the same name as the target object, significantly below chance level in both the similar (M =0.25, SD = 0.44, t(27)= -15, p <.001) and the dissimilar condition (M = 0.07, SD = 0.26, t(27) = -28.82,p <.001).Since none of the infants in both conditions succeeded in more than 50% of the trials we did not perform the chi-square analysis.The 95% confidence interval for this result ranges from 0 to 14.3%, suggesting that this task is definitely too difficult for an average 14-month-old.
Submitting these results to a two-factor mixed ANOVA condition (withinsubject factor: similar vs. dissimilar) and gender (between-subjects factor: girls vs. boys) revealed no statistically significant effects of condition, gender, or their interaction.However, a medium effect size of condition suggests that these results failed to achieve statistical significance only due to a relatively small sample size (F(1,26) = 3.72, p = .065,classical η 2 = .06).In other words, even though all infants scored lower than 50% on this task, those who succeded on some trials, did so more often in the similar condition.
The analysis of the CDI results showed that 14-month-olds had a mean productive vocabulary of 23.6 words (SD = 28.2;Mdn = 16.5;range 0-142).The correlation between the results on the categorization task and infants' productive vocabulary was not significant (r s (26) = .05,p = .809).
16 months.One sample t-tests showed that infants at 16 months chose the object similar to the target object significantly above chance level (M = 2.0, SD = 0.94, t(27) = 2.81, p = .009).In the dissimilar condition, their performance remained at chance level (M = 1.11,SD = 1.2, t(27) = -1.74,p <.094).A twofactor mixed ANOVA was conducted to assess the effects of condition and gender.A large significant main effect of condition (F(1,26) = 12.26, p = .002,classical η 2 = .15)confirmed the predicted outcome: 16-month-olds reliably chose the object that has the same name as the target object only if these two objects also shared at least some perceptual features.Infants in the similar condition chose the right object on average in 2 out of 3 trials (M = 2.00, SD = 0.94), significantly more often than in the dissimilar condition (M = 1.10,SD = 1.97).Neither the main effect of gender nor the effect of interaction condition x gender were significant.
Although the chi-square test was not significant (χ2(1) = 2.29, p = .18),18 out of 28 (64.3%, 95% CI [45.8%, 79.4%]) infants succeeded on more than 50% of the trials in the similar condition.Thus, our results seem to support previous findings that more than 50% of the 16-month-old infants can categorize unknown, perceptually similar objects with a shared label, but a larger sample is needed to make a more precise estimation.
When it comes to unknown dissimilar objects, only 9 out of 28, or 25.7% of the infants succeeded on more than 50% of the trials (χ2(1) = 3.57, p = .09,95% CI [17.8%, 50.8%]).This result confirmed our predictions based on previous studies that a shared label is not strong enough a cue for an average 16-month-old to categorize two objects in absence of perceptual similarity on an object manipulation task, although the upper end of the confidence interval lets us assume that this task might be solved by 50% of the 16-month-olds.
The analysis of the CDI results showed that 16-month-olds had a mean productive vocabulary of 52.6 words (SD = 52.9;Mdn = 36; range 2-228).The correlation between the results on categorization tasks and infants's productive vocabulary at 16 months was not significant (r s (24) = .09,p = .650).

Discussion
This study makes two primary contributions.First, the results obtained on English and French infants are now extended to Serbian language.Since crosslinguistic research has become crucial for determining which links between word-learning and object-categorization are universal and which ones are language-specific (McCardle, Colombo, & Freund, 2009;Slobin et al., 2008;Waxman, 2009), our study is an important step towards achieving this goal.
Although there is no reason to believe that infants differ in cognitive abilities across linguistic and cultural settings, we cannot generalize the results obtained on non-random samples in Western laboratories to infants in general for at least two reasons: first, English is not a very good basis for understanding how children learn a wide range of languages due to its fixed word order and rather simple morphology (Lieven & Stoll, 2010); second, there might be significant socio-economic and cultural differences in different samples across different cultures.
Moreover, it is, to our knowledge, the first study in which one sample of infants was tested at both 14 and 16 months.In that way we were able to avoid any potential confound due to the cohort-effect (Teti, 2005, p.8) and draw stronger conclusions about age-related changes in early categorization.
Our results demonstrate that, when an object-manipulation task is used, infants learn to categorize familiar objects from a same basic-level category on average around 14 months.Even though the mean score for all infants was significantly above chance level, there were 12 out of 31 infants for which two clear demonstrations and instructions repeated at each trial were not enough to succeed on the familiarization phase.
Our predictions regarding 14-month-olds' ability to categorize unknown objects with a similar shape and a shared label as well as unknown objects with a shared label, but no perceptual similarity were confirmed.As previously shown by Booth and Waxman (2002) and Nazzi and Pilardeau (2007), this ability develops between 14 and 16 months of age for similar objects with a shared label.Our results are also in line with previous findings concerning categorization of dissimilar objects with a shared label reported by Nazzi and Gopnik (2001).Apparently, in highly interactive procedures that require the use of socio-pragmatic cues, categorization of novel objects based on naming information alone is a task too difficult for infants aged 14 months, and it is only at 20 months that infants start developing this ability.
One possible explanation for the 14-month-olds' failure on these tasks is that they did not form word-object associations necessary for word-based categorization to take place.Werker et al. (1998) reported that at 14 months infants can rapidly form arbitrary word-object associations with minimal exposure, without any social or contextual support.In our task, the exposure to novel objects and corresponding labels was thoroughly supported by repetition, experimenter's gaze and infant-directed speech.Nevertheless, Werker and colleagues concluded that this ability develops around 14 months, thus it is possible that the failure on this task is not only due to the lack of categorization abilities, but also to the inability of infants to memorize several short words and associate them with several novel objects in a short time.Perhaps instead of using monosyllabic words alone, it would be better to use a combination of long and short words so that infants can differentiate new words and objects more easily.
It is worth noting that, in our experiment, infants aged 14 months performed below chance level on the task of categorization of both similar and dissimilar objects with a shared label.There are several potential explanations for this result.First, in most cases, infants did not get through the whole testing session. 2 familiarization trials and 6 test trials made a session quite demanding for an average 14-month-old.Even when they sat through the whole session, after the first 3 to 4 trials infants started to play with the objects in their own way, regardless of the instructions.An important reason for the 14 montholds' failure to form categories could also be the alternate order of similar and dissimilar conditions.For infants who succeded on the familiarization trials and the first test trial (which was always the similar one), the dissimilar trial seemed to disrupt the understanding of the task.It is possible that, if the similar trials came in a consecutive order the result on the similar condition would have been better, an assumption also supported by the fact that14-month-olds' results were better in the similar vs. dissimilar condition.This time, we opted for the alternate order so as to avoid fatigue as a potential confound if the dissimilar trials had always been the three last ones, but this is certainly an important factor for further studies to be taken into account.
The 16-month-olds showed above chance level performance on the similar categorization task, but chance level performance on the dissimilar categorization task.We can now extend the findings established by relevant studies in other languages (e.g.Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007;Rakison & Oaks, 2003) to Serbian: 16-month-olds can categorize novel objects in an interactive task requiring the use of socio-pragmatic cues only when provided with some perceptual cue, whereas categorization does not occur based on naming alone.This result goes in favor of the classic piagetian sensorimotor view of infancy, "child-as-data-analyist" view (Waxman & Gelman, 2009) and the assumption that infants rely on apparent features and statistical regularities rather than on their conceptual abilities (Sloutsky & Fisher, 2011).Yet, numerous studies demonstrated that, even when perceptual cues are available, object naming enhances infants' success on object categorization (Balaban & Waxman, 1997;Booth & Waxman, 2002;Fulkerson & Waxman, 2007;Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007;Rakison & Oakes, 2003;Waxman, 2002).In other words, infants' success on the similar condition is most likely due to both perceptual similarity and labeling.
Our infants' performance was slightly lower than the performance of English and French infants on a similar task, in spite of the potentially facilitating effect of familiarity with the setting and instructions in our experiment.Inability to form word-object associations in a short time could be one reason for somewhat lower success on this task by Serbian infants.Nazzi and Pilardeau (2007) showed that at 16 months infants can learn three names for three objects in a similar object-manipulation task, although they still cannot use those names to categorize dissimilar objects.However, Serbian might be more difficult for this task compared to English and French, given its complex morphology: adding inflections to object labels potentially enhances the cognitive load associated with the task (Slobin, 1982).It is also possible that the performance of Serbian infants was sligthly lower due to the alternate order of the similar and dissimilar condition (the problem explained earlier in discussion), whereas in Nazzi & Gopnik (2001) the similar trials were exposed first.
The failure of 16-month-olds in the dissimilar condition confirms findings eshablished by Nazzi and colleagues (Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007).A possible explanation for this result is the connection between accelerated lexical development and word-based categorization.Given that the so-called lexical explosion, or vocabulary spurt occurs, on average, between 13 and 18 months of age (Cohen & Brunt, 2009;Plunkett et al., 2008;Schafer & Plunkett, 1998;Werker et al., 1998), it is reasonable to expect that word-based categorization comes fully into scene only after the infants have mastered the process of word-learning.
No significant correlation between the success on categorization tasks and productive vocabulary at 14 and 16 months goes in favor of the fact that infants categorize objects based on perceptual cues but not based on naming information.This finding is also consistent with previous studies which suggest that correlation of the success on categorization tasks and vocabulary measures occurs only around 20 months, as the importance of words for categorization becomes clearer to infants (Nazzi & Gopnik, 2001;Nazzi & Pilardeau, 2007).
The main limitation of our study seems to be a relatively small sample size.Another potential limitation might be in the task itself.During the study, we realized that the very action of putting an object into the basket was so engrossing for some infants that, regardless of the instructions, they enjoyed putting all three objects into the basket.An alternative procedure where infants should simply point to the target object, or give the target object to the experimenter should be considered in future research.Additionally, the familiarization task that would include exemplars with the same label but no perceptual similarity (both during the actual experiment, and during the practising phase with the mothers) would be desirable.
To sum up, we can now extend to a new language the finding that infants up to at least 16 months of age rely primarily on perceptual information to categorize novel objects in interactive tasks requiring the use of socio-pragmatic cues.Word-based categorization in this situation probably becomes more important during the second half of the second year, a hypothesis that is being investigated in our ongoing research.

Table 1
Frequency of infants who gave 0, 1, or 2 correct answers on the familiarization task

Table 2
Frequency of infants who gave 0, 1, 2, or 3 correct answers on the categorization tasks