Please post your comments on the papers and presentations 9/30 and 10/2 here.
29 thoughts on “Paper presentations”
Monica Bennett
The Goldstein and Schwade (2008) experiment used mothers as the feedback-givers. It seems logical that care-givers would have the strongest influence on an infant’s acquired phonology, since they tend to be the ones whose language the infant is most often exposed to, but it made me wonder how far this effect goes. Is there any evidence that infants will restructure their babbling from exposure to people outside the role of care-giver? I’m interested in how this interacts with a bilingual environment as well.
I am also curious about under which conditions or context infants will restructure their babbling as found by Goldstein & Schwade (2008). Specifically, if there would be differences in infants’ babbling response if primary caregivers vocalized using infant-directed speech versus using non-infant directed speech? Do infants recognize or differentiate when their primary caregiver is directing speech toward them based on the sound or pitch of the voice (i.e., infant directed speech vs. more adult-directed speech)?
How are syllables represented in memory? I’ve often thought of syllabification as a chunking operation over segments; however, Newport & Aslin’s (2004) results seem to argue against this. A chunking account would seem to predict an inversion of their results such that unpacking a chunk (syllable) to compare inter-syllabic segments would be more difficult than a comparison between chunks (syllables) that would not require unpacking.
In the same vein, if transitional probabilities are computed between adjacent segments within a segmental tier, do these computations occur before or after syllabification?
There appears to be a confound in Newport & Aslin’s (2004) study. In the non-adjacent syllable studies, there are two segments (one syllable) intervening between the non-adjacent syllables, while in the non-adjacent segment studies, there is only a single segment intervening between the non-adjacent segments. It seems a little far-fetched to me, but is it possible that there is simply more memory decay across two segments than across one?
Josh: I think that Newport and Aslin consider syllables being represented on a different tier, which is not the combination of the consonants and the vowels tiers. Therefore, it is just a matter of zooming into the appropriate tier, where the units of operation consist of more than one segment.
I believe that according to autosegmental theory, a segment can be associated with more than one tier simultaneously. Therefore, a segment can be part of a syllable on one tier, and just an element in a stream of its likes in the consonantal or vowel tier.
But on the other hand, they hint that it also might be a general cognitive grouping mechanism. In this case I would imagine that the elements are clustered into different groups and then the probabilities are calculated within each group but not in their original setting.
I guess it also raises a question how automatic this process is…
Monica: That’s a good question, and you are right to make the bilingual connection, since it’s hard to know how you could measure this outside of the Goldstein and Schwade experimental paradigm without a language difference between the caregiver and others. I don’t know the babbling literature that well, but this looks like a good place to start – Google didn’t give me anything newer than this, and this is very good lab:
Kristine Yu of Linguistics has been recently investigating the prosodic characteristics of IDS, and would be a good person to talk to if you get interested in this.
My question is on a minor aspect of the Lewandowsky (2011) paper. He found practice effects for the “difficult categories,” where average performance improved as the session went on. He followed Wilhelm and Oberauer (2006) and removed the practice effects and counterbalancing noise. In a task where you’re really only looking at learning performance, is this kind of correction/transformation okay? He says that the scores “preserved all individual differences and mean performance for each problem,” which sounds okay, but I am a little skeptical about removing any type of learning from your dependent variable that’s measuring learning.
In this paper, the authors were looking at the effect of socially contingent speech and actions on infants babbling. To do this, the experimenters split up mothers and their infants in to two groups, one in which the parents were to respond to their infant in a contingent way with certain (experimenter controlled) phonological patterns (vowel only or CV), and the other in which the parents did not respond in a contingent way, instead using the experimental groups responses. Both groups did interact socially with their infants. Results showed that infants in the experimental group altered their babbling to include phonological patterns that their mothers used.
I thought this paper was really interesting, as one of the few that combined both learning and social aspects. I would really like to see different variations of this, such as still acting in a contingent way, but instead manipulating who’s voice the infants hear.
Lewandowsky (2011) demonstrated the relationships between working memory capacity and visual categorization using SEM. He found that working memory capacity, or rather, the latent variable that drove the working memory task performance, was highly correlated with a single latent variable that predicted all 6 of the Shepard et al. (1961) categorization tasks, though only some of those tasks are supposedly rule-based, which working memory is supposed to be particularly important for. Also, though some but not all of the tasks are thought to particularly rely on dimensional attention, dimensional attention did not differ by task. The ALCOVE model was able to fit the data and the parameter that most explained individual differences was a learning parameter, rather than the parameter associated with dimensional attention as might be thought for a categorization task. This learning parameter was also related to working memory capacity.
It’s interesting that though working memory could be thought of as being particularly relevant for rule-based categorization tasks, it was important for all of the categorization tasks. I would like to ensure that the latent variable can really be thought of as working memory capacity, rather than just some other unspecified individual difference. Of course, we know that working memory is important for a large number of cognitive tasks, so for falsifiability, what task could we include that would show that the latent variable of working memory is NOT relevant to just any cognition? Also, I thought that a particularly class-relevant follow-up to this paper would be to replicate the findings with phonological categorization tasks, which would presumably be more like information-integration tasks, since the task strategy is not verbalizable.
Hirotanie and colleagues (2009) present evidence that infants better learn novel word-object associations when taught in a social versus non-social context. Hirotanie and colleagues (2009) found an increased negativity to congruent compared to incongruent word-object pairs between 200 and 600ms post stimulus onset. The authors suggested this early negativity was reflective of word familiarity or word recognition. Additionally, authors found a significant increase in negativity to incongruent compared to congruent word-object pairs occurring between 800 and 1200ms post stimulus onset. This adult-like late N400 is suggested to reflect semantic priming or in-depth information integration.
Overall this article seemed to support and extend previous behavioral and ERP research investigating the development of word acquisition. One critique concerns the methods section. Information concerning the total number trials infants could see versus the average number of trials completed was not clear. Additionally, standard information found in ERP literature such as the average number of trials that were included in the final analyses for each condition did not seem to be included. Event-related potentials provide electrophysiological evidence for the type of mechanisms elicited during word learning and recognition; however methods such as eye-tracking may provide additional insight to potential differences in visual strategies used across social and non-social learning contexts. Hirotanie and colleagues’ (2009) findings could be applied to research related to early identification/detection of social processing disorders such as Autism Spectrum Disorder.
Wang and Kuhl (2003) provide strong reasoning for why there was a need to re-examine the Critical Period Hypothesis and how it applies to learning speech sounds from a second language. Although previous research has examined adult second language learning little research had focused on neither childhood nor the transition between pre-adolescence into adolescence and adulthood. Overall Wang and Kuhl examined whether second language learning is a function of biological time (i.e., maturation) or experience. This study found evidence which did not support the Critical Period Hypothesis. Instead, they provide evidence that learning speech sounds of a second language is a function of experience. Specifically, across 4 ages children and adults equally improved their ability to identify speech sounds post a 2-week training period.
A possible direction these findings could be extended to is situations that are almost the opposite of the second language training experience described in the study. What I mean is that since second language learning may be a function of experience, how do experiences where a child’s first language does not continue to be their first language, but may later re-learn this language as an older child or adult? To better understand how experience (and possibly in combination with biological maturation) effects learning speech sounds it would make sense to look at whether early experience with a language has lasting effects that may support or hinder re-learning that language as an adult. For example, children who are trans-nationally adopted from one country to another country with a different primary language. Will that child’s early experience with their initial L1 support re-learning this language as an adolescent or adult?
Goyet et al. (2010) run two ERP experiments with French-learning 12-month-olds in an effort to examine learning of word-segmentation. Prior studies using a Headturn Preference Procedure (HPP) had shown that while English-learning infants use trochaic units as cues to word learning (e.g. Juscyzk et al., 1999), French-learning infants use syllabic units as cues to word learning (Nazzi et al, 2006), as is expected given the prevalence of trochees in English and their absence in French. However, Nazzi et al.’s (2006) results were surprising in that they only showed evidence of segmentation of bisyllabic words at 16-months-old, which is later than the age by which French infants already know many bisyllabic words. Goyet et al. (2010) suppose that this inconsistency is a limitation of HPP, citing prior research in Dutch-learning infants that word segmentation effects are detectable at an earlier age with ERP.
The first experiment familiarizes infants with 10 different tokens of the same low-frequency word. In the test phase, 12 sentences were presented, six containing the familiarized word, and six not containing the familiarized word. ERPs were time-locked to the onset of the first and second syllables and the two conditions were compared 350-500 ms after the onset of either syllable. There was a main effect of familiarity in the critical region after the onset of the first syllable, but not in the critical region after the onset of the second syllable.
The second experiment attempts to hunt for a familiarization effect on the second syllable of a target word by only familiarizing infants with the second syllable of the target word. The researchers fail to find a main effect of familiarization, but claim that there is a significant segmentation effect on the basis of an interaction with quadrant.
I don’t think that experiment 2 answers the questions raised by experiment 1. No syllable from the familiarized words in experiment 1 appears in any of the test sentences (aside from the target word). As the authors discuss, this allows for the possibility of not needing to attend to the second syllable of the familiarized word; it always co-occurs with the first syllable. They try to address this problem with experiment 2, but familiarizing infants on a single syllable seems to change the task from word segmentation to syllable identification. Perhaps a better follow-up would be to familiarize infants on bisyllabic words as in experiment 1 (e.g. dorade) but then to have the control word in the test phase have the same second syllable as the familiarized word (e.g. tirade).
McMurray et al. (2008) try to compare effects of expectancy across the domains of speech and music. The Ganong (1980) effect shows that when subjects are presented with stimuli from an artificial continuum, such as from /froot/ to /froop/, where the final consonant of words from the medial portion of the continuum is ambiguous as to whether it is a [t] or a [p], perception of these ambiguous phones is modulated by contextual (in this case lexical) expectation. Subjects are biased to perceive a larger range of the ambiguous stimuli as /froot/, because a final [t] forms a valid English word, whereas a final [p] does not.
Using this analogy, McMurray et al. (2008) construct artificial continua of musical chords that range from major to minor. They then present tokens from a particular continuum in a harmonic context that is designed to bias expectation of either a major or a minor chord. Contrary to the Ganong (1980) effect, McMurray et al. (2008) report a perceptual bias in the opposite direction of the contextual expectation. They conclude that perceptual processing of music is different from perceptual processing of speech.
However, it may be the case that the Ganong (1980) effect is the wrong analogy for musical expectations. Where /froop/ is unambiguously not a word in English, it is hard to say that any harmonic progression is categorically unacceptable.
Kenneth J. Kurtz (2004) proposes a novel implementation of neural network architecture as an explanation of human category learning. In order to shore up the claim that this type of network actually mimics human category learning, Kurtz tests it on the Shepard, Hovland, and Jenkins (1961) shape category learning tasks.
Kurtz’s Divergent Autoencoder or DIVA model operates using multiple autoencoder neural networks. These networks tend to be effective only when tested on unsupervised learning problems. Kurtz’s model effectively turns supervised learning problems, involving the teaching of multiple categories via example, into a form on which autoencoders can operate. The model then runs two autoencoders in parallel and makes a classification response based on the amount of reconstructive error in each. The model performs effectively, and in a human-like manner, in a variety of category classifications that other models have had trouble interpreting.
I appreciated Kurtz’s invoking a direct and reasonable test of category learning in order to determine the effectiveness of the model. I do still find most of the neural network modeling literature a bit overly dependent upon hand-waving, though. It seems like the argument that the model performs in a similar manner to human beings therefore the mechanisms underlying human learning must be similar feel weak sometimes. I can make a neural network model that mimics human-like coin-flipping and it won’t mean much.
Since we didn’t get a chance to get to any developmental disabilities that are associated with phonology, word comprehension, etc, I thought it would be interested to review a paper that looked at dyslexia in infants and children.
In this study, Zuijen and colleagues (2013) did a longitudinal study that began with 26 2-month old infants who were followed until 2nd grade (about 7 years old). These were infants who had a predisposition to dyslexia, and there was also a control group. At 2-months old, infants were brought in to the lab and EEG and EOG were used. The stimuli used were naturally spoken CVC words. When the children were in 2nd grade, reading fluency was measured with a standardized word reading fluency test. At ages 3 and 4, their non-verbal intelligence and language comprehension were measured.
Results showed that infants who were able to process the different speech sounds at 2 months became fluent readers later in life, but the opposite result of distinguishing non-fluent readers was not found. The author does cite the fact that this was only found for a certain group could be due to many outside environmental factors that shape the child throughout life.
Berent, Steriade, Lennertz, and Vaknin (2007) found that monosyllabic nonsense words with unattested onset clusters that more strongly violated the Sonority Sequencing Principle (SSP) were more likely to be misidentified as disyllabic by English- speaking listeners. Furthermore, when English-speaking listeners were asked to discriminate between a monosyllabic word and its disyllabic counterpart created by inserting a schwa within the onset cluster (e.g., lbif and lebif), English-speaking listeners were more likely to report hearing two instances of the same word rather than two different words. These results not only highlight perceptual difficulties among listeners in representing SSP-violating words, but perceptual epenthetic repair of these violations. In addition, attestedness was shown to be an important factor. Performance for Russian speakers, whose language includes the onset clusters used in this study, was far superior, although similarly modulated by sonority.
Davidson (2006) examined the effects of sonority on articulation among English speakers. English speakers were less accurate in producing words with unattested onset clusters that plateaued or fell in sonority (e.g., zvabu and zdaba) compared to those that rose in sonority (e.g. zmafe and znafe). Furthermore, there was both a main effect of the first consonant on accuracy (f>z>v) and an interaction between the first and second consonants, suggesting that sonority profile alone does not explain the entirety of the accuracy effects. Importantly, the researchers found that vowel insertion between the initial consonants was the most common error among speakers. Results from a follow-up experiment, in which speakers were asked to produce CC onset clusters and CәC counterparts (e.g., zvabu and zevabu), showed that vowel insertion was shorter in duration for CC compared to CəC clusters. Given these results, the researchers argue that speakers accurately represent and attempt to produce unattested onset clusters, but failures in articulatory coordination, rather than epenthetic repair, result in vowel insertion. However, prior to articulation, speakers in this study received an orthographic presentation of each word in addition to an auditory presentation, which may have strengthened the veracity of their representations.
Nosofsky & Palmeri (1996) Attempt to replicate the classic Shepard, Hovland, and Jenkins ’61 visual category learning study. However, instead of using the classic visually separable stimuli, they chose to use integral stimuli. They define integral stimuli as stimuli whose feature dimensions are all continuous. That is, in this case they used colors differing in saturation, hue, and brightness. Each of these features is a continuous measure and the end result is a continuous whole rather than an easily separable set of features. Essentially, they find that the standard SHJ results reverse II>IV to IV>II if you use integral stimuli. The explanation they offer is that selective attention is a likely driver of this effect. Separable visual elements can be selectively attended, whereas integral objects cannot easily be seen as being composed of separable parts that can be attended to individually. Without selective attention, subjects cannot show a II>IV effect, and rely more on exemplars that allow them to put together a “feel” of IV quickly, even if the rule remains unverbalized.
I had a couple of questions regarding some of the presentations on projects that were done in class, and since there isn’t a place to post those I feel like this could be an ok area.
Regarding Lisa’s project, I was first wondering what the difference between audiovisual and audio/visual was? I was a little confused when we were discussing that. Now, even though this is different than what she was proposing, I recently completed a study in infants where we were presenting them with stimuli that were visually different and made different sounds (it was a social evaluation experiment though). In a follow up condition, we took away one of those cues (appearance) and did not get as strong as an effect as when they were together. We’re currently doing the opposite and just using appearance, so it will be interesting to see what those effects are. Personally, I would feel that stimuli that were both auditory and visual would have the biggest impact on learning, since there is so much stimulation happening it may be easier to learn those.
Marcus et al. (1999) raise the question of what learning mechanisms are available to infants, proposing at least two possible mechanisms. One of these is statistical, as, for example, transitional probabilities; some researchers have suggested that statistical mechanisms are sufficient for language learning. Another potential mechanism is something that allows infants to learn algebraic rules, i.e. abstract relationships between different objects. These would be rules of the form a=b+3, where any value could be substituted in for b to determine the value of a, or S = NP + VP, where anything of the categories NP and VP could be substituted in to form something of the category S.
To pull these explanations apart, Marcus et al. designed an experiment combining familiar grammatical patterns with novel lexical items. Infants were habituated to one of two grammars, then exposed to test items using novel lexical items. Across three experiments, a vast majority (15/16, 15/16, 16/16) of the infants showed a longer attention time to the test sentences that were inconsistent with the training grammar.
As it seems clear (to Marcus et al., and to me as well!) that infants could not appeal to transitional probabilities for these test items, these experiments demonstrate that infants are capable of learning and applying more general rules.
As a generative linguist, I’m inclined to celebrate Marcus et al.’s findings (without much surprise at their conclusions), but others are maybe not so enthusiastic. Gerry Altman, a creator of one of the neural network models critiqued in the Marcus et al. paper, published this response shortly afterward.
In his response, Altman claims that his model was able to learn the same patterns as in the Marcus et al. experiment with novel test items, and draws from this the conclusion that infants’ learning may not significantly differ from statistical learning mechanisms after all.
There’s a short response from Marcus in the same document; in this response, he claims that the decision about what the model is “learning” depends on one’s interpretation of the model’s outputs. In another interpretation, “in terms of what word is most active at any given moment (a standard way of analyzing models of this general class),” (Marcus 875a) in fact the model isn’t learning the target grammar, but oscillating between several possible grammars.
Marcus also claims that in his work on this model, he has found that the model does not genuinely learn an abstract pattern, but actually to map a given lexical item’s representation onto a new lexical item, which may not be what infants are doing; he sketches an experiment that might pull these possibilities apart.
I don’t know enough specifics of the Dienes, Altman, and Gao model to have an informed opinion on this particular back and forth, but it is an interesting question — can we actually appeal to statistical learning to explain what looks like abstract rules?
Newport and Aslin (2004) present three experiments which explore the adult learners’ ability to compute transitional probabilities between non-adjacent elements and to use this knowledge for word segmentation. In Experiment 1 native English speakers were exposed to concatenated three syllable nonsense words, in which transitional probabilities were the highest between the first and the last syllable within each word. This type of language was proven to be impossible to learn even when the stimuli inventory was kept to only three words and the exposure occurred over more than one session. In Experiments 2 and 3 the authors test the hypothesis that the learning of non-adjacent dependencies is constrained by the types occurring in natural languages. Experiment 2 employed nonsense words in which transitional probabilities were the highest between the three consonants of each word; in Experiment 3 the transitional probabilities were the highest between the vowels of each stimulus word. Both types of languages were successfully acquired.
As the authors themselves note, their study supports the general cognitive principle that learning regularities between non-adjacent elements interleaved with elements of other type is relatively easy; learning such dependencies between non-adjacent elements of the same type is harder.
With respect to the word segmentation tasks, previous infant studies show that transitional probabilities can be overridden by other cues for word boundaries (eg pitch, coarticulation Johnson and Jusczyk 2001). A possible follow up would be to augment the language used in Experiment 1 with acoustic cues such as stress, to explore whether their presence will allow the adult speakers to learn the language with non-adjacent syllable regularities.
The results of this study seem to disagree with the evidence from Johnson and Jusczyk (2001) which found that on given conflicting statistical and acoustic cues, English learning infants rely on acoustic and not statistical cues. It is unclear at this point whether the differences in the results of the two studies is because English and Spanish have different priorities in their usage of word segmentation cues, or because infants and adults differ in their attention to those cues.
Barlow and Gierut (1999) discuss child phonological acquisition and its theoretical analysis in the context of Optimality Theory (OT).
OT is similar to MaxEnt in that both posit the existence of underlying forms which may or may not differ from the output forms produced by speakers. In OT, unlike MaxEnt, these constraints derive output forms by rankings rather than weights.
Barlow and Gierut follow Ferguson and Garnica (1975)’s criteria for evaluating the adequacy of a general linguistic theory as a theory of acquisition — namely, that such a theory must capture (1) the data about discrepancies between children’s and adults’ productions, (2) individual grammar differences, and (3) the changing nature of a developing grammar. All of these can be captured, they show, by OT — positing a reranking of constraints over the course of child development.
Barlow and Gierut show that child phonology can be derived using the same constraints as adult phonology, and that OT meets the three criteria above. The first is captured in OT by generally ranking markedness constraints higher than faithfulness constraints. The second is captured by positing differing constraint rankings for different individuals. The third is captured by changing constraint rankings for learners over the course of their acquisition.
While it’s certainly true that in most cases, any given output form can be derived by showing a varied constraint ranking, I’m a little skeptical about OT’s true adequacy in these second two criteria, particularly as compared with a MaxEnt model. Unlike the gradient descent models we discussed in class, it’s not clear exactly how this reranking over time might occur; moreover, MaxEnt gives a clearer picture of how to model individual variation than OT.
Norris, McQueen, and Cutler (2003) examined whether listeners retune phonetic boundaries based on lexical information, using stimuli that ended in a sound between [f] and [s]. The most ambiguous step in the continuum between [f] and [s] was selected based on pretests. In Experiment 1, 49 native speakers of Dutch performed a lexical decision task as training. The stimulus set in this part of the experiment either contained the ambiguous sound in contexts that formed a word with /f/ in its place (but not /s/; e.g., “meatloa?” in English) as well as clear, natural instances of word-final /s/ (e.g., “tennis”), contexts that formed a word with /s/ in its place (but not /f/; e.g., “tenni?” in English) as well as clear instances of word-final /f/ (e.g., “meatloaf”), or only instances of clear word-final [f] and [s]. There were no other instances of /f/ or /s/ in the words.
Afterward, participants performed a short phonetic categorization task, in which they would indicate whether a range of ambiguous fricatives–steps in a continuum from [É›f] to [É›s]–were [f] or [s]. Results demonstrated that participants who heard the ambiguous fricative during training in contexts lexically biased for [f] tended to categorize more steps on this task as [f] than those who hard heard it in contests biased for [s]. The authors conclude that this shift suggests that listeners rapidly used lexical knowledge to retune phonetic categories. Experiment 2 further supported these results, using several controls to rule out alternative explanations, such as selective adaptation and contrast effects.
Minda, Desroches, and Church conducted a series of experiments on both children and adults to see how well each age group performed on SHJ Type I-VI categories. They were interested in using the two age groups to test dual-process models of categorization, hypothesizing that children would perform poorly on the Shepard types that required a complex verbal rule (taxing working memory, which is less developed in children) but well on the types which relied upon perceptual learning (where categorization decisions cannot be summarized by clearly-defined rules). An additional prediction is that young children should perform worse than a group of older children (i.e., teenagers). The authors also hypothesized that adults would perform better overall than children, showing no significant performance difference between rule-defined and non-rule-defined category learning. These hypotheses were all presented in the framework of COVIS, a multiple-systems theory of category learning. The specific predictions with regard to Shephard types were that there would be no children/adult differences in Types I and IV category learning (learned through procedural system) but that differences would emerge for Types II and III (which require explicit, complex verbal rules).
In Experiment 1, children and adults were tested on 4/6 Shepard types. Results supported some of the authors’ predictions, in that adults out-performed children (although the youngest children, 3-year-olds, did poorly even on Type I), with only adults performing well on Type II. Both age groups did poorly on Type IV.
In Experiment 2, adults were tested to see whether the explicit system and verbal working memory play a role in learning rule-defined categories but not non-rule-defined ones. Adults were randomly assigned to one of three groups: 1) no concurrent task, 2) verbal concurrent task, and 3) nonverbal concurrent task. The first group did not differ in any way from the adults tested in Experiment 1. In the verbal concurrent task group, participants had to perform verbal coarticulation (i.e., saying a word aloud when it appeared on the screen) while they did their categorization, and in the nonverbal concurrent task condition, participants had to tap their index finger each time an asterisk (*) appeared on screen. It was predicted that performance would suffer in Types II and III for people in the verbal concurrent task condition because those systems rely upon verbal working memory. Indeed, adults in the verbal concurrent task group showed degraded performance in Type II categorization, but no difference was found between the groups for the nonverbal concurrent task participants. Together, Experiments 1 and 2 provide evidence for a multiple-systems account of category learning.
Experiment 3 tested whether childrens’ struggle with Type II could be alleviated if their working memory capacity was burdened less. To accomplish this, children were asked to repeat all of the names of the stimulus items prior to categorizing them, and they were also told that items were to be classified into “families.” Children were tested only on Types II and IV, and their performance was much better in Type II categorization than it was in Experiment 1 (stimuli in Exps. 1 and 3 were the same). When the results of all three experiments are considered, the authors conclude that there is sufficient evidence to support a multiple-systems view of categorization of Shepard types.
Thiessen and Saffran were interested in how young infants (7-9 months of age) perform word segmentation, specifically whether segmentation is guided by stress cues or statistical cues (i.e., transitional probabilities). Previous research has suggested that infants first pick up stress cues to help them with word segmentation, but this poses a chicken and egg problem of sorts: how can they learn any words at all if stress patterns are their first cues to word boundaries? If stress is the first cue that they use to segment words, then they can’t be learning about the stress pattern of English from words that they have segmented from fluent speech streams. Which comes first?
The goal of Experiment 1 was to determine whether 9-month-olds preferentially attend to stress cues over statistical cues in a word segmentation task using synthesized speech sounds. Two artificial languages were synthesized for use during a familiarization period, one iambic (stress on second syllable in bisyllabic word) and one trochaic (stress on first syllable of bisyllabic word). During the familiarization phase, two of the bisyllabic words occurred twice as often as the other two, to adjust for item frequency during the test phase (i.e., so that the babies heard the whole word and words created at the boundaries an equal number of times). Results showed that the difference in looking time between words and part-words (i.e., bisyllabic segments spanning a word boundary, such as BA#tu, in the sement diBA#tuNI) was significant, with babies looking longer at part words. Further, the overall picture of looking time shows a differential pattern of responses: 9-month-olds look more at words than non-words in the trochaic pattern condition, but this pattern reverses in the iambic pattern condition, where they look more at non-words. This interaction was significant. Since they were looking longer at part-words in the iambic condition, this means that they were treating the part-words as words, guided by the stress pattern they might expect (i.e., stress on the first syllable, which is more common in English). Therefore, this experiment confirms the authors’ prediction that 9-month-olds will use stress cues over statistical cues when they are in conflict.
Experiment 2 was essentially a replication of Experiment 1 with a younger subject pool (6.5 to 7-month-old infants). The stimuli were identical to those used in Experiment 1. A 2×2 ANOVA (item x condition) revealed a main effect of item (infants looked longer at part-words than words), a main effect of condition (looked longer at iambic than trochaic), and no interaction. These results suggest that younger infants are achieving the same segmentation results after hearing both iambic and trochaic languages, such that this younger group of infants was weighing statistical cues more heavily than stress cues compared to the 9-month-old group.
A final experiment, Experiment 3, tested 9-month-old infants on the same languages used in Exps. 1 and 2, but they were synthesized in a monotone (i.e., stress removed). In the absence of stress cues, the babies showed a novelty preference, which was to be expected if they had been relying upon stress cues when they were present. Overall, these results demonstrate that younger babies use statistical cue knowledge to segment words while older babies rely more upon stress cues.
@Josh, I’m a little confused about the second paper summary you posted — is /froot/ and /froop/ one of the actual examples they give in the paper? I only have a survey size of one native speaker, but I’m pretty sure my phonotactics allow /froop/ just as easily as /froot/.
I’m also interested in how they “bias expectation of a major or minor chord.” I don’t know the details of this, but it sounds like there’s some assumption that the preceding sounds/chords license later sounds/chords, right? I’m wondering how it would go to do a similar music study, but parallel to the English onset restriction. So under this assumption of licensing by prior sounds, one could say that in the /pnit/ vs. /prit/ minimal pair, the single phoneme /p/ rules out /n/ in some way that it does not rule out /r/. Setting aside the question of whether this is in fact what’s happening, in the music experiment, could you use only one preceding chord to bias for a following chord?
Yes, the phonotactics of /frut/ and /frup/ are both legal. froot-froop is not one of the examples that they give in the paper, but it is emblematic of stimuli used to demonstrate the Ganong effect. Along a continuum between [t] and [p], some ambiguous phones that are physically more similar to [p] are nevertheless heard as [t] given the preceding context [fru] because [frut] forms a word in English (“fruit”) while [frup] does not.
As for the music, the tonality (“musical context”) is what the authors used to manipulate the expectation biases. You’re right that there’s definitely an assumption that preceding chords will license later chords, particularly in the paradigm that they used where there was a forced choice between only two chords. (Outside the laboratory, there’s probably not a single chord that’s categorically licensed or ruled out by any given context, but rather some distribution of expectations over all possible chords. I imagine something is similar for phonotactics — I’ve definitely heard the restaurant “Panera” pronounced as [pnɛɹʌ])
Actually, Bharucha and Krumhansl (1983) used two chords, as you suggest, but they didn’t explicitly examine expectation bias. Rather, they examined the perceptual relatedness of the two chords — subjects were asked to rate how well the second chord followed the first chord on a 7-point scale. Using multidimensional scaling, they found clear patterns of relatedness that conform with theoretical notions of tonality. Two theoretically related chords were perceived as being more related than two theoretically distant chords. Whether this means that the first chord set up an expectation for the second chord, I can’t say.
The Goldstein and Schwade (2008) experiment used mothers as the feedback-givers. It seems logical that care-givers would have the strongest influence on an infant’s acquired phonology, since they tend to be the ones whose language the infant is most often exposed to, but it made me wonder how far this effect goes. Is there any evidence that infants will restructure their babbling from exposure to people outside the role of care-giver? I’m interested in how this interacts with a bilingual environment as well.
I am also curious about under which conditions or context infants will restructure their babbling as found by Goldstein & Schwade (2008). Specifically, if there would be differences in infants’ babbling response if primary caregivers vocalized using infant-directed speech versus using non-infant directed speech? Do infants recognize or differentiate when their primary caregiver is directing speech toward them based on the sound or pitch of the voice (i.e., infant directed speech vs. more adult-directed speech)?
How are syllables represented in memory? I’ve often thought of syllabification as a chunking operation over segments; however, Newport & Aslin’s (2004) results seem to argue against this. A chunking account would seem to predict an inversion of their results such that unpacking a chunk (syllable) to compare inter-syllabic segments would be more difficult than a comparison between chunks (syllables) that would not require unpacking.
In the same vein, if transitional probabilities are computed between adjacent segments within a segmental tier, do these computations occur before or after syllabification?
There appears to be a confound in Newport & Aslin’s (2004) study. In the non-adjacent syllable studies, there are two segments (one syllable) intervening between the non-adjacent syllables, while in the non-adjacent segment studies, there is only a single segment intervening between the non-adjacent segments. It seems a little far-fetched to me, but is it possible that there is simply more memory decay across two segments than across one?
Josh: I think that Newport and Aslin consider syllables being represented on a different tier, which is not the combination of the consonants and the vowels tiers. Therefore, it is just a matter of zooming into the appropriate tier, where the units of operation consist of more than one segment.
I believe that according to autosegmental theory, a segment can be associated with more than one tier simultaneously. Therefore, a segment can be part of a syllable on one tier, and just an element in a stream of its likes in the consonantal or vowel tier.
But on the other hand, they hint that it also might be a general cognitive grouping mechanism. In this case I would imagine that the elements are clustered into different groups and then the probabilities are calculated within each group but not in their original setting.
I guess it also raises a question how automatic this process is…
Monica: That’s a good question, and you are right to make the bilingual connection, since it’s hard to know how you could measure this outside of the Goldstein and Schwade experimental paradigm without a language difference between the caregiver and others. I don’t know the babbling literature that well, but this looks like a good place to start – Google didn’t give me anything newer than this, and this is very good lab:
http://www.psych.mcgill.ca/perpg/fac/genesee/13.pdf
Charisse: Here’s a nice overview on IDS:
http://www.cfs.purdue.edu/ITSI/docs/briefs/InfantDirectedSpeech.pdf
Kristine Yu of Linguistics has been recently investigating the prosodic characteristics of IDS, and would be a good person to talk to if you get interested in this.
My question is on a minor aspect of the Lewandowsky (2011) paper. He found practice effects for the “difficult categories,” where average performance improved as the session went on. He followed Wilhelm and Oberauer (2006) and removed the practice effects and counterbalancing noise. In a task where you’re really only looking at learning performance, is this kind of correction/transformation okay? He says that the scores “preserved all individual differences and mean performance for each problem,” which sounds okay, but I am a little skeptical about removing any type of learning from your dependent variable that’s measuring learning.
Goldstein & Schwade (2008) Summary:
In this paper, the authors were looking at the effect of socially contingent speech and actions on infants babbling. To do this, the experimenters split up mothers and their infants in to two groups, one in which the parents were to respond to their infant in a contingent way with certain (experimenter controlled) phonological patterns (vowel only or CV), and the other in which the parents did not respond in a contingent way, instead using the experimental groups responses. Both groups did interact socially with their infants. Results showed that infants in the experimental group altered their babbling to include phonological patterns that their mothers used.
I thought this paper was really interesting, as one of the few that combined both learning and social aspects. I would really like to see different variations of this, such as still acting in a contingent way, but instead manipulating who’s voice the infants hear.
Lewandowsky (2011) Summary:
Lewandowsky (2011) demonstrated the relationships between working memory capacity and visual categorization using SEM. He found that working memory capacity, or rather, the latent variable that drove the working memory task performance, was highly correlated with a single latent variable that predicted all 6 of the Shepard et al. (1961) categorization tasks, though only some of those tasks are supposedly rule-based, which working memory is supposed to be particularly important for. Also, though some but not all of the tasks are thought to particularly rely on dimensional attention, dimensional attention did not differ by task. The ALCOVE model was able to fit the data and the parameter that most explained individual differences was a learning parameter, rather than the parameter associated with dimensional attention as might be thought for a categorization task. This learning parameter was also related to working memory capacity.
It’s interesting that though working memory could be thought of as being particularly relevant for rule-based categorization tasks, it was important for all of the categorization tasks. I would like to ensure that the latent variable can really be thought of as working memory capacity, rather than just some other unspecified individual difference. Of course, we know that working memory is important for a large number of cognitive tasks, so for falsifiability, what task could we include that would show that the latent variable of working memory is NOT relevant to just any cognition? Also, I thought that a particularly class-relevant follow-up to this paper would be to replicate the findings with phonological categorization tasks, which would presumably be more like information-integration tasks, since the task strategy is not verbalizable.
Hirotanie and colleagues (2009) present evidence that infants better learn novel word-object associations when taught in a social versus non-social context. Hirotanie and colleagues (2009) found an increased negativity to congruent compared to incongruent word-object pairs between 200 and 600ms post stimulus onset. The authors suggested this early negativity was reflective of word familiarity or word recognition. Additionally, authors found a significant increase in negativity to incongruent compared to congruent word-object pairs occurring between 800 and 1200ms post stimulus onset. This adult-like late N400 is suggested to reflect semantic priming or in-depth information integration.
Overall this article seemed to support and extend previous behavioral and ERP research investigating the development of word acquisition. One critique concerns the methods section. Information concerning the total number trials infants could see versus the average number of trials completed was not clear. Additionally, standard information found in ERP literature such as the average number of trials that were included in the final analyses for each condition did not seem to be included. Event-related potentials provide electrophysiological evidence for the type of mechanisms elicited during word learning and recognition; however methods such as eye-tracking may provide additional insight to potential differences in visual strategies used across social and non-social learning contexts. Hirotanie and colleagues’ (2009) findings could be applied to research related to early identification/detection of social processing disorders such as Autism Spectrum Disorder.
Wang and Kuhl (2003) provide strong reasoning for why there was a need to re-examine the Critical Period Hypothesis and how it applies to learning speech sounds from a second language. Although previous research has examined adult second language learning little research had focused on neither childhood nor the transition between pre-adolescence into adolescence and adulthood. Overall Wang and Kuhl examined whether second language learning is a function of biological time (i.e., maturation) or experience. This study found evidence which did not support the Critical Period Hypothesis. Instead, they provide evidence that learning speech sounds of a second language is a function of experience. Specifically, across 4 ages children and adults equally improved their ability to identify speech sounds post a 2-week training period.
A possible direction these findings could be extended to is situations that are almost the opposite of the second language training experience described in the study. What I mean is that since second language learning may be a function of experience, how do experiences where a child’s first language does not continue to be their first language, but may later re-learn this language as an older child or adult? To better understand how experience (and possibly in combination with biological maturation) effects learning speech sounds it would make sense to look at whether early experience with a language has lasting effects that may support or hinder re-learning that language as an adult. For example, children who are trans-nationally adopted from one country to another country with a different primary language. Will that child’s early experience with their initial L1 support re-learning this language as an adolescent or adult?
Goyet et al. (2010) run two ERP experiments with French-learning 12-month-olds in an effort to examine learning of word-segmentation. Prior studies using a Headturn Preference Procedure (HPP) had shown that while English-learning infants use trochaic units as cues to word learning (e.g. Juscyzk et al., 1999), French-learning infants use syllabic units as cues to word learning (Nazzi et al, 2006), as is expected given the prevalence of trochees in English and their absence in French. However, Nazzi et al.’s (2006) results were surprising in that they only showed evidence of segmentation of bisyllabic words at 16-months-old, which is later than the age by which French infants already know many bisyllabic words. Goyet et al. (2010) suppose that this inconsistency is a limitation of HPP, citing prior research in Dutch-learning infants that word segmentation effects are detectable at an earlier age with ERP.
The first experiment familiarizes infants with 10 different tokens of the same low-frequency word. In the test phase, 12 sentences were presented, six containing the familiarized word, and six not containing the familiarized word. ERPs were time-locked to the onset of the first and second syllables and the two conditions were compared 350-500 ms after the onset of either syllable. There was a main effect of familiarity in the critical region after the onset of the first syllable, but not in the critical region after the onset of the second syllable.
The second experiment attempts to hunt for a familiarization effect on the second syllable of a target word by only familiarizing infants with the second syllable of the target word. The researchers fail to find a main effect of familiarization, but claim that there is a significant segmentation effect on the basis of an interaction with quadrant.
I don’t think that experiment 2 answers the questions raised by experiment 1. No syllable from the familiarized words in experiment 1 appears in any of the test sentences (aside from the target word). As the authors discuss, this allows for the possibility of not needing to attend to the second syllable of the familiarized word; it always co-occurs with the first syllable. They try to address this problem with experiment 2, but familiarizing infants on a single syllable seems to change the task from word segmentation to syllable identification. Perhaps a better follow-up would be to familiarize infants on bisyllabic words as in experiment 1 (e.g. dorade) but then to have the control word in the test phase have the same second syllable as the familiarized word (e.g. tirade).
McMurray et al. (2008) try to compare effects of expectancy across the domains of speech and music. The Ganong (1980) effect shows that when subjects are presented with stimuli from an artificial continuum, such as from /froot/ to /froop/, where the final consonant of words from the medial portion of the continuum is ambiguous as to whether it is a [t] or a [p], perception of these ambiguous phones is modulated by contextual (in this case lexical) expectation. Subjects are biased to perceive a larger range of the ambiguous stimuli as /froot/, because a final [t] forms a valid English word, whereas a final [p] does not.
Using this analogy, McMurray et al. (2008) construct artificial continua of musical chords that range from major to minor. They then present tokens from a particular continuum in a harmonic context that is designed to bias expectation of either a major or a minor chord. Contrary to the Ganong (1980) effect, McMurray et al. (2008) report a perceptual bias in the opposite direction of the contextual expectation. They conclude that perceptual processing of music is different from perceptual processing of speech.
However, it may be the case that the Ganong (1980) effect is the wrong analogy for musical expectations. Where /froop/ is unambiguously not a word in English, it is hard to say that any harmonic progression is categorically unacceptable.
Kenneth J. Kurtz (2004) Summary:
Kenneth J. Kurtz (2004) proposes a novel implementation of neural network architecture as an explanation of human category learning. In order to shore up the claim that this type of network actually mimics human category learning, Kurtz tests it on the Shepard, Hovland, and Jenkins (1961) shape category learning tasks.
Kurtz’s Divergent Autoencoder or DIVA model operates using multiple autoencoder neural networks. These networks tend to be effective only when tested on unsupervised learning problems. Kurtz’s model effectively turns supervised learning problems, involving the teaching of multiple categories via example, into a form on which autoencoders can operate. The model then runs two autoencoders in parallel and makes a classification response based on the amount of reconstructive error in each. The model performs effectively, and in a human-like manner, in a variety of category classifications that other models have had trouble interpreting.
I appreciated Kurtz’s invoking a direct and reasonable test of category learning in order to determine the effectiveness of the model. I do still find most of the neural network modeling literature a bit overly dependent upon hand-waving, though. It seems like the argument that the model performs in a similar manner to human beings therefore the mechanisms underlying human learning must be similar feel weak sometimes. I can make a neural network model that mimics human-like coin-flipping and it won’t mean much.
Since we didn’t get a chance to get to any developmental disabilities that are associated with phonology, word comprehension, etc, I thought it would be interested to review a paper that looked at dyslexia in infants and children.
In this study, Zuijen and colleagues (2013) did a longitudinal study that began with 26 2-month old infants who were followed until 2nd grade (about 7 years old). These were infants who had a predisposition to dyslexia, and there was also a control group. At 2-months old, infants were brought in to the lab and EEG and EOG were used. The stimuli used were naturally spoken CVC words. When the children were in 2nd grade, reading fluency was measured with a standardized word reading fluency test. At ages 3 and 4, their non-verbal intelligence and language comprehension were measured.
Results showed that infants who were able to process the different speech sounds at 2 months became fluent readers later in life, but the opposite result of distinguishing non-fluent readers was not found. The author does cite the fact that this was only found for a certain group could be due to many outside environmental factors that shape the child throughout life.
Berent, Steriade, Lennertz, and Vaknin (2007) found that monosyllabic nonsense words with unattested onset clusters that more strongly violated the Sonority Sequencing Principle (SSP) were more likely to be misidentified as disyllabic by English- speaking listeners. Furthermore, when English-speaking listeners were asked to discriminate between a monosyllabic word and its disyllabic counterpart created by inserting a schwa within the onset cluster (e.g., lbif and lebif), English-speaking listeners were more likely to report hearing two instances of the same word rather than two different words. These results not only highlight perceptual difficulties among listeners in representing SSP-violating words, but perceptual epenthetic repair of these violations. In addition, attestedness was shown to be an important factor. Performance for Russian speakers, whose language includes the onset clusters used in this study, was far superior, although similarly modulated by sonority.
Davidson (2006) examined the effects of sonority on articulation among English speakers. English speakers were less accurate in producing words with unattested onset clusters that plateaued or fell in sonority (e.g., zvabu and zdaba) compared to those that rose in sonority (e.g. zmafe and znafe). Furthermore, there was both a main effect of the first consonant on accuracy (f>z>v) and an interaction between the first and second consonants, suggesting that sonority profile alone does not explain the entirety of the accuracy effects. Importantly, the researchers found that vowel insertion between the initial consonants was the most common error among speakers. Results from a follow-up experiment, in which speakers were asked to produce CC onset clusters and CәC counterparts (e.g., zvabu and zevabu), showed that vowel insertion was shorter in duration for CC compared to CəC clusters. Given these results, the researchers argue that speakers accurately represent and attempt to produce unattested onset clusters, but failures in articulatory coordination, rather than epenthetic repair, result in vowel insertion. However, prior to articulation, speakers in this study received an orthographic presentation of each word in addition to an auditory presentation, which may have strengthened the veracity of their representations.
Nosofsky & Palmeri (1996) Attempt to replicate the classic Shepard, Hovland, and Jenkins ’61 visual category learning study. However, instead of using the classic visually separable stimuli, they chose to use integral stimuli. They define integral stimuli as stimuli whose feature dimensions are all continuous. That is, in this case they used colors differing in saturation, hue, and brightness. Each of these features is a continuous measure and the end result is a continuous whole rather than an easily separable set of features. Essentially, they find that the standard SHJ results reverse II>IV to IV>II if you use integral stimuli. The explanation they offer is that selective attention is a likely driver of this effect. Separable visual elements can be selectively attended, whereas integral objects cannot easily be seen as being composed of separable parts that can be attended to individually. Without selective attention, subjects cannot show a II>IV effect, and rely more on exemplars that allow them to put together a “feel” of IV quickly, even if the rule remains unverbalized.
I had a couple of questions regarding some of the presentations on projects that were done in class, and since there isn’t a place to post those I feel like this could be an ok area.
Regarding Lisa’s project, I was first wondering what the difference between audiovisual and audio/visual was? I was a little confused when we were discussing that. Now, even though this is different than what she was proposing, I recently completed a study in infants where we were presenting them with stimuli that were visually different and made different sounds (it was a social evaluation experiment though). In a follow up condition, we took away one of those cues (appearance) and did not get as strong as an effect as when they were together. We’re currently doing the opposite and just using appearance, so it will be interesting to see what those effects are. Personally, I would feel that stimuli that were both auditory and visual would have the biggest impact on learning, since there is so much stimulation happening it may be easier to learn those.
Marcus et al. (1999) raise the question of what learning mechanisms are available to infants, proposing at least two possible mechanisms. One of these is statistical, as, for example, transitional probabilities; some researchers have suggested that statistical mechanisms are sufficient for language learning. Another potential mechanism is something that allows infants to learn algebraic rules, i.e. abstract relationships between different objects. These would be rules of the form a=b+3, where any value could be substituted in for b to determine the value of a, or S = NP + VP, where anything of the categories NP and VP could be substituted in to form something of the category S.
To pull these explanations apart, Marcus et al. designed an experiment combining familiar grammatical patterns with novel lexical items. Infants were habituated to one of two grammars, then exposed to test items using novel lexical items. Across three experiments, a vast majority (15/16, 15/16, 16/16) of the infants showed a longer attention time to the test sentences that were inconsistent with the training grammar.
As it seems clear (to Marcus et al., and to me as well!) that infants could not appeal to transitional probabilities for these test items, these experiments demonstrate that infants are capable of learning and applying more general rules.
As a generative linguist, I’m inclined to celebrate Marcus et al.’s findings (without much surprise at their conclusions), but others are maybe not so enthusiastic. Gerry Altman, a creator of one of the neural network models critiqued in the Marcus et al. paper, published this response shortly afterward.
In his response, Altman claims that his model was able to learn the same patterns as in the Marcus et al. experiment with novel test items, and draws from this the conclusion that infants’ learning may not significantly differ from statistical learning mechanisms after all.
There’s a short response from Marcus in the same document; in this response, he claims that the decision about what the model is “learning” depends on one’s interpretation of the model’s outputs. In another interpretation, “in terms of what word is most active at any given moment (a standard way of analyzing models of this general class),” (Marcus 875a) in fact the model isn’t learning the target grammar, but oscillating between several possible grammars.
Marcus also claims that in his work on this model, he has found that the model does not genuinely learn an abstract pattern, but actually to map a given lexical item’s representation onto a new lexical item, which may not be what infants are doing; he sketches an experiment that might pull these possibilities apart.
I don’t know enough specifics of the Dienes, Altman, and Gao model to have an informed opinion on this particular back and forth, but it is an interesting question — can we actually appeal to statistical learning to explain what looks like abstract rules?
Newport and Aslin (2004) present three experiments which explore the adult learners’ ability to compute transitional probabilities between non-adjacent elements and to use this knowledge for word segmentation. In Experiment 1 native English speakers were exposed to concatenated three syllable nonsense words, in which transitional probabilities were the highest between the first and the last syllable within each word. This type of language was proven to be impossible to learn even when the stimuli inventory was kept to only three words and the exposure occurred over more than one session. In Experiments 2 and 3 the authors test the hypothesis that the learning of non-adjacent dependencies is constrained by the types occurring in natural languages. Experiment 2 employed nonsense words in which transitional probabilities were the highest between the three consonants of each word; in Experiment 3 the transitional probabilities were the highest between the vowels of each stimulus word. Both types of languages were successfully acquired.
As the authors themselves note, their study supports the general cognitive principle that learning regularities between non-adjacent elements interleaved with elements of other type is relatively easy; learning such dependencies between non-adjacent elements of the same type is harder.
With respect to the word segmentation tasks, previous infant studies show that transitional probabilities can be overridden by other cues for word boundaries (eg pitch, coarticulation Johnson and Jusczyk 2001). A possible follow up would be to augment the language used in Experiment 1 with acoustic cues such as stress, to explore whether their presence will allow the adult speakers to learn the language with non-adjacent syllable regularities.
Toro-Soto, RodrÃguez-Fornells and Sebastián-Gallés (2007) explore the use of stress as a cue for word segmentation in Spanish. The authors cite on the previous literature which had established that speakers apply their prevalent stress patterns on novel stimuli (eg Vroomen et al 1998). Based on that, the authors hypothesize that the Spanish speakers will use stress as a cue for penultimate syllable, which is the most common stress position. To test their hypothesis, the authors composed four languages from three syllable stimuli identical to Saffran et al’s (1996). The words were concatenated in a stream with no pauses. Three languages had a stress cue – raised pitch – on either initial, medial or final syllable. The fourth language had flat stress and only the transitional probabilities served as a cue for word boundaries. Contrary to the author’s expectations, the penultimate stress language was the only one which the speakers did not manage to learn (40% success), while they managed to learn the remaining three languages with a similar rate of success (70% success). The results suggest that the stress cue can be used productively for word segmentation only if it coincides with the word boundaries and thus supports the statistical cues. Otherwise it contradicts the statistical cues and is interfering with the word learning.
The results of this study seem to disagree with the evidence from Johnson and Jusczyk (2001) which found that on given conflicting statistical and acoustic cues, English learning infants rely on acoustic and not statistical cues. It is unclear at this point whether the differences in the results of the two studies is because English and Spanish have different priorities in their usage of word segmentation cues, or because infants and adults differ in their attention to those cues.
Barlow and Gierut (1999) discuss child phonological acquisition and its theoretical analysis in the context of Optimality Theory (OT).
OT is similar to MaxEnt in that both posit the existence of underlying forms which may or may not differ from the output forms produced by speakers. In OT, unlike MaxEnt, these constraints derive output forms by rankings rather than weights.
Barlow and Gierut follow Ferguson and Garnica (1975)’s criteria for evaluating the adequacy of a general linguistic theory as a theory of acquisition — namely, that such a theory must capture (1) the data about discrepancies between children’s and adults’ productions, (2) individual grammar differences, and (3) the changing nature of a developing grammar. All of these can be captured, they show, by OT — positing a reranking of constraints over the course of child development.
Barlow and Gierut show that child phonology can be derived using the same constraints as adult phonology, and that OT meets the three criteria above. The first is captured in OT by generally ranking markedness constraints higher than faithfulness constraints. The second is captured by positing differing constraint rankings for different individuals. The third is captured by changing constraint rankings for learners over the course of their acquisition.
While it’s certainly true that in most cases, any given output form can be derived by showing a varied constraint ranking, I’m a little skeptical about OT’s true adequacy in these second two criteria, particularly as compared with a MaxEnt model. Unlike the gradient descent models we discussed in class, it’s not clear exactly how this reranking over time might occur; moreover, MaxEnt gives a clearer picture of how to model individual variation than OT.
Summary for Norris, McQueen, and Cutler (2003):
Norris, McQueen, and Cutler (2003) examined whether listeners retune phonetic boundaries based on lexical information, using stimuli that ended in a sound between [f] and [s]. The most ambiguous step in the continuum between [f] and [s] was selected based on pretests. In Experiment 1, 49 native speakers of Dutch performed a lexical decision task as training. The stimulus set in this part of the experiment either contained the ambiguous sound in contexts that formed a word with /f/ in its place (but not /s/; e.g., “meatloa?” in English) as well as clear, natural instances of word-final /s/ (e.g., “tennis”), contexts that formed a word with /s/ in its place (but not /f/; e.g., “tenni?” in English) as well as clear instances of word-final /f/ (e.g., “meatloaf”), or only instances of clear word-final [f] and [s]. There were no other instances of /f/ or /s/ in the words.
Afterward, participants performed a short phonetic categorization task, in which they would indicate whether a range of ambiguous fricatives–steps in a continuum from [É›f] to [É›s]–were [f] or [s]. Results demonstrated that participants who heard the ambiguous fricative during training in contexts lexically biased for [f] tended to categorize more steps on this task as [f] than those who hard heard it in contests biased for [s]. The authors conclude that this shift suggests that listeners rapidly used lexical knowledge to retune phonetic categories. Experiment 2 further supported these results, using several controls to rule out alternative explanations, such as selective adaptation and contrast effects.
Summary for Minda, Desroches, and Church (2008):
Minda, Desroches, and Church conducted a series of experiments on both children and adults to see how well each age group performed on SHJ Type I-VI categories. They were interested in using the two age groups to test dual-process models of categorization, hypothesizing that children would perform poorly on the Shepard types that required a complex verbal rule (taxing working memory, which is less developed in children) but well on the types which relied upon perceptual learning (where categorization decisions cannot be summarized by clearly-defined rules). An additional prediction is that young children should perform worse than a group of older children (i.e., teenagers). The authors also hypothesized that adults would perform better overall than children, showing no significant performance difference between rule-defined and non-rule-defined category learning. These hypotheses were all presented in the framework of COVIS, a multiple-systems theory of category learning. The specific predictions with regard to Shephard types were that there would be no children/adult differences in Types I and IV category learning (learned through procedural system) but that differences would emerge for Types II and III (which require explicit, complex verbal rules).
In Experiment 1, children and adults were tested on 4/6 Shepard types. Results supported some of the authors’ predictions, in that adults out-performed children (although the youngest children, 3-year-olds, did poorly even on Type I), with only adults performing well on Type II. Both age groups did poorly on Type IV.
In Experiment 2, adults were tested to see whether the explicit system and verbal working memory play a role in learning rule-defined categories but not non-rule-defined ones. Adults were randomly assigned to one of three groups: 1) no concurrent task, 2) verbal concurrent task, and 3) nonverbal concurrent task. The first group did not differ in any way from the adults tested in Experiment 1. In the verbal concurrent task group, participants had to perform verbal coarticulation (i.e., saying a word aloud when it appeared on the screen) while they did their categorization, and in the nonverbal concurrent task condition, participants had to tap their index finger each time an asterisk (*) appeared on screen. It was predicted that performance would suffer in Types II and III for people in the verbal concurrent task condition because those systems rely upon verbal working memory. Indeed, adults in the verbal concurrent task group showed degraded performance in Type II categorization, but no difference was found between the groups for the nonverbal concurrent task participants. Together, Experiments 1 and 2 provide evidence for a multiple-systems account of category learning.
Experiment 3 tested whether childrens’ struggle with Type II could be alleviated if their working memory capacity was burdened less. To accomplish this, children were asked to repeat all of the names of the stimulus items prior to categorizing them, and they were also told that items were to be classified into “families.” Children were tested only on Types II and IV, and their performance was much better in Type II categorization than it was in Experiment 1 (stimuli in Exps. 1 and 3 were the same). When the results of all three experiments are considered, the authors conclude that there is sufficient evidence to support a multiple-systems view of categorization of Shepard types.
Summary for Thiessen and Saffran (2003):
Thiessen and Saffran were interested in how young infants (7-9 months of age) perform word segmentation, specifically whether segmentation is guided by stress cues or statistical cues (i.e., transitional probabilities). Previous research has suggested that infants first pick up stress cues to help them with word segmentation, but this poses a chicken and egg problem of sorts: how can they learn any words at all if stress patterns are their first cues to word boundaries? If stress is the first cue that they use to segment words, then they can’t be learning about the stress pattern of English from words that they have segmented from fluent speech streams. Which comes first?
The goal of Experiment 1 was to determine whether 9-month-olds preferentially attend to stress cues over statistical cues in a word segmentation task using synthesized speech sounds. Two artificial languages were synthesized for use during a familiarization period, one iambic (stress on second syllable in bisyllabic word) and one trochaic (stress on first syllable of bisyllabic word). During the familiarization phase, two of the bisyllabic words occurred twice as often as the other two, to adjust for item frequency during the test phase (i.e., so that the babies heard the whole word and words created at the boundaries an equal number of times). Results showed that the difference in looking time between words and part-words (i.e., bisyllabic segments spanning a word boundary, such as BA#tu, in the sement diBA#tuNI) was significant, with babies looking longer at part words. Further, the overall picture of looking time shows a differential pattern of responses: 9-month-olds look more at words than non-words in the trochaic pattern condition, but this pattern reverses in the iambic pattern condition, where they look more at non-words. This interaction was significant. Since they were looking longer at part-words in the iambic condition, this means that they were treating the part-words as words, guided by the stress pattern they might expect (i.e., stress on the first syllable, which is more common in English). Therefore, this experiment confirms the authors’ prediction that 9-month-olds will use stress cues over statistical cues when they are in conflict.
Experiment 2 was essentially a replication of Experiment 1 with a younger subject pool (6.5 to 7-month-old infants). The stimuli were identical to those used in Experiment 1. A 2×2 ANOVA (item x condition) revealed a main effect of item (infants looked longer at part-words than words), a main effect of condition (looked longer at iambic than trochaic), and no interaction. These results suggest that younger infants are achieving the same segmentation results after hearing both iambic and trochaic languages, such that this younger group of infants was weighing statistical cues more heavily than stress cues compared to the 9-month-old group.
A final experiment, Experiment 3, tested 9-month-old infants on the same languages used in Exps. 1 and 2, but they were synthesized in a monotone (i.e., stress removed). In the absence of stress cues, the babies showed a novelty preference, which was to be expected if they had been relying upon stress cues when they were present. Overall, these results demonstrate that younger babies use statistical cue knowledge to segment words while older babies rely more upon stress cues.
@Josh, I’m a little confused about the second paper summary you posted — is /froot/ and /froop/ one of the actual examples they give in the paper? I only have a survey size of one native speaker, but I’m pretty sure my phonotactics allow /froop/ just as easily as /froot/.
I’m also interested in how they “bias expectation of a major or minor chord.” I don’t know the details of this, but it sounds like there’s some assumption that the preceding sounds/chords license later sounds/chords, right? I’m wondering how it would go to do a similar music study, but parallel to the English onset restriction. So under this assumption of licensing by prior sounds, one could say that in the /pnit/ vs. /prit/ minimal pair, the single phoneme /p/ rules out /n/ in some way that it does not rule out /r/. Setting aside the question of whether this is in fact what’s happening, in the music experiment, could you use only one preceding chord to bias for a following chord?
@Megan
Yes, the phonotactics of /frut/ and /frup/ are both legal. froot-froop is not one of the examples that they give in the paper, but it is emblematic of stimuli used to demonstrate the Ganong effect. Along a continuum between [t] and [p], some ambiguous phones that are physically more similar to [p] are nevertheless heard as [t] given the preceding context [fru] because [frut] forms a word in English (“fruit”) while [frup] does not.
As for the music, the tonality (“musical context”) is what the authors used to manipulate the expectation biases. You’re right that there’s definitely an assumption that preceding chords will license later chords, particularly in the paradigm that they used where there was a forced choice between only two chords. (Outside the laboratory, there’s probably not a single chord that’s categorically licensed or ruled out by any given context, but rather some distribution of expectations over all possible chords. I imagine something is similar for phonotactics — I’ve definitely heard the restaurant “Panera” pronounced as [pnɛɹʌ])
Actually, Bharucha and Krumhansl (1983) used two chords, as you suggest, but they didn’t explicitly examine expectation bias. Rather, they examined the perceptual relatedness of the two chords — subjects were asked to rate how well the second chord followed the first chord on a 7-point scale. Using multidimensional scaling, they found clear patterns of relatedness that conform with theoretical notions of tonality. Two theoretically related chords were perceived as being more related than two theoretically distant chords. Whether this means that the first chord set up an expectation for the second chord, I can’t say.