Stanford’s Word Bank

http://wordbank.stanford.edu/

Wordbank is an open database for storing information about children’s vocabulary growth.

Wordbank is approved by the MB-CDI advisory board, and is a project of the Stanford Language and Cognition Lab, PI Michael C. Frank. Contributors include Dan Yurovsky, Virginia Marchman, Ranjay Krishna, Mika Braginsky, and Benji Nguyen.

Wordbank archives data from the MacArthur-Bates Communicative Development Inventory (MB-CDI), a family of parent-report questionnaires. The Wordbank database enables researchers to recover data filtered by source, age, gender, word, and a host of other variables, enabling simple export of plain-text data for further analysis.

Wordbank also includes a number of reports based on recent research on children’s vocabulary: see how children’s vocabulary grows and changes across early childhood.

Wordbank is open access! The graphical search interface is currently under construction, but see this tutorial to start using R to analyze the data today.

And, there’s R-code for accessing  it: http://rpubs.com/dyurovsky/using-wordbank

Instruments from SpatialLearning.org

http://spatiallearning.org/index.php/resources/testsainstruments

All sorts of spatial tests here, including the following areas. Some tests are specifically geared for children.

  • Force and motion
  • Mental bending
  • Mental (de)fragmentation
  • Mental folding
  • Mental slicing & penetrative thinking
  • Mental rotation
  • Navigation
  • Perspective taking
  • Scaling
  • Spatial assembly
  • 2D to 3D translation
  • Understanding topographic maps

Wither consciousness?

Bargh, J. A., & Chartrand, T. L. (1999). The unbearable automaticity of being. American Psychologist, 54, 462-479.

What was noted by E. J. hanger (1978) remains true today: that much of contemporary psychological research is based on the assumption that people are consciously and systematically processing incoming information in order to construe and interpret their world and to plan and engage in courses of action. As did E. J. hanger, the authors question this assumption. First, they review evidence that the ability to exercise such conscious, intentional control is actually quite limited, so that most of moment-to-moment psychological life must occur through nonconscious means if it is to occur at all. The authors then describe the different possible mechanisms that produce automatic, environmental control over these various phenomena and review evidence establishing both the existence of these mechanisms as well as their consequences for judgments, emotions, and behavior. Three major forms of automatic self-regulation are identified: an automatic effect of perception on action, automatic goal pursuit, and a continual automatic evaluation of one’s experience. From the accumulating evidence, the authors conclude that these various nonconscious mental systems perform the lion’s share of the self-regulatory burden, beneficently keeping the individual grounded in his or her current environment.

Continue reading

Schacter, Chiu, and Ochsner: Implicit memory

Schacter, D. L., Chiu, P. C. -Y., Ochsner, K. N. (1993). Implicit memory: A selective review. Annual Reviews of Neuroscience, 16, 159-182.

Memory is not a unitary phenomenon. Some memory processes are driven by conscious effort to bring to mind previously encountered information (declarative memory). Similarly, we might consciously try to recreate previous experiences, recalling where we were, what we were feeling, what we were sensing, etc. (episodic memory). These are explicit memory processes; whcih I think of as memory of. (I have a memory of an episode in my life, or an explict memory of a list that I learned.)

Other times, though, memory is apparent without conscious effort, and without any explicit effort either during encoding or retrieval. This is memory for; our knowledge of how to do things, contrasted with the propositional knowledge above (“knowledge that”).

Continue reading

Baddeley’s working memory may not play a role in speech production

Gathercole, S., & Baddeley, A. (1993). Working memory and language (pp. 161-175). Hillsdale, NJ: Lawrence Erlsbaum.

Gathercole and Baddeley argue that working memory might play two roles in speech production.

Working memory in a between-processes role.
Working memory might serve as a buffer for output. If speech production is a stage-based process, with different levels of representation of language content (message, function, position, sound), working memory may hold the output from processing at one level so that it can be picked up as input for processing at another.

Is the phonological loop identical with a motor planning buffer? Probably not. In studies of speech-onset timing, there is little difference between simple reaction time experiments–where a participant is told a target word to say, and then presented with a cue to say it– and experiments where they are asked to fill the time intervening between target word and cue with nonsense syllables. If the phonological loop were a working space for motor planning, rehearsing the nonsense sylalbles should impede the production of the target word. It is possible that we construct rather lengthy motor programs (above the word or stress group level), by combining sub programs in some working space, but that space may not be the phonological loop.

Some correlational evidence in neurophysiology might speak to a connection between the phonological loop and a motor planning buffer: e.g. Broca’s aphasics show deficits in fluent speech and in short term memory. Importantly, though, there seems a quite demonstrable dissociation: there are patients with relatively profound loss of short-term memory skills, but who, nonetheless, retain fluent speech planning.

It may be that, while the phonological loop is not necessary as a buffer for output from speech planning systems, it may well be that the same processes that underly articulation underlie subvocal rehearsal processes (repeating things to oneself), and might show correlated effects from damage or interference.

Working memory in a within processes role.
Alternatively, or perhaps additionally, working memory might serve a _within-processes_ role. For instance, “to achieve a positional level representation, the phonological specifications of the main lexical items must then be retrieved, a planning frame for the sentence constructed, the words inserted into the planning frame, and the appropriate affixes and function words assessed and inserted.” All of this activity requires interactivity between the cognitive “thing” that will eventually be the utterance, and all this stuff in long-term memory (meanings of words, syntactic structures appropriate to the language, etc.). Working memory may serve as the workshop for all of this cognitive processing. The central executive may be most at work here.

Loading the central executive with tasks (like remembering 6 digits while also trying to form interesting, grammatical sentences) results in speech that is more stereotyped and predictable than when the central executive is less loaded with activity. The effect is small, and present only in relatively difficult conditions (e.g. 6 digit task) rather than in easier condition that might be handled solely by the phonological loop (a 3 digit task). More work needs to be done here.

Developmental trajectories.
Though the phonological loop may not be synonymous with the motor buffer for speech production in adults, it is possible that it plays a greater role in the speech production of children. Similarly, the automaticity of much of adult speech, and it’s relative independence from other cogitive and physically demanding activty, may suggest a relatively minor role of the central executive in production. However, in children who haven’t yet the levels of practice to automate the task of speaking, the central executive may be more involved, and speech less independent of other tasks.

My questions

  1. Would we expect deaf children (particularly of deaf parents) to have a “cheremological loop?” Would we expect differences (in function, capacity, etc.) in such a system? Would it be different still from the visuo-spatial sketchpad?
  2. Does the evidence for dissociation between perception/memory tasks and speech production, which Gathercole believes fails to support a meaningful connection between working memory processes and a buffer for motor planning activities, pose similar challenges for a connectionist model of production? A symbolic approach to mental functioning would seem quite amenable to this idea of dissociation–my idea of a “chair” need not be tied to any particular acquisition event or any particular type of behavioral output (including speech) related to chairs. In a connectionist model, wouldn’t we anticipate more of a connection — if my idea of “chair” is really a certain mental state of shared activation across input, output, and hidden nodes, wouldn’t we predict more of a relationship between perception, memory, and production?
  3. In the first paragraph of the “Speeded Speech Production” section, Gathercole discusses the simple reaction time and choice reaction time tests. This may be a problem with her paraphrase, rather than a design problem in the experiments, but in the simple reaction time experiment participants were “told” (one would believe via speech) the target word. In the choice reaction time experiment, participants read words visually. Could some of the syllable-length effects be due to the fact that participants who are told a word are given a pretty good model of how to articulate that word, while reading requires a bit more translation into an appropriate motor plan?

Gathercole & Baddeley: Introduction to working memory

Gathercole, S., & Baddeley, A. (1993). Working memory and language (pp. 1-12). Hillsdale, NJ: Lawrence Erlsbaum.

Working memory, here, is “the short-term memory system, which is involved in the temporary processing and storage of information.” Baddeley’s model is a “resources” model (as opposed to a discrete slots model, or a decay model, or an interference model) of short-term memory.

The working memory model proposed is tripartite. The first component, a central executive, monitors two slave systems: a phonological loop and a visuo-spatial sketchpad. (This latter is little involved in speech perception, so I will ignore it here.)

The central executive is involved in selective control of action, planning, coordination of tasks, possibly consciousness. This executive might be a unitary process, or it might be several cooperative subprocesses. Tasks that require the inhibition of prepotent responses in favor of more novel responses would seem to involve the central executive.

The authors propose that the phonological loop has two processes. The first of these is a passive buffer of sorts, that takes in phonological information from the environment. This buffer is subject to word length effects: the more syllables in a set of target words, the more difficult those targets are to remember, possibly because the ribbon of the loop is too short to capture them all. The buffer is also subject to articulatory suppression effects–when we are prohibited from beginning to subvocally rehearsing, retention suffers.

The second process is an articulatory rehearsal process. This process is subject to phonological similarity effects: when target words share phonological characters, they are more difficult to remember. It is subjct to irrelevant speech effects, too. When non-target words share similar phonology–independent of whether they might share semantic or lexical similarity–they interfer more readily with retention of target words.

Some of my questions

  1. Do fast talkers routinely test better on working memory than slow talkers do?
  2. The authors offer up as evidence of capacity limits for the phonological loop that longer words (more syllables and/or more time necessary to articulate) results in lower recall scores. Is word length confounded with corpus frequency? Do the effects remain when controlling for distributional differences? [2b] In visual attention/working memory research, we see that sometimes what’s touted as resource-driven, qualitative effects (e.g. lower memory span for complex objects than for simple objects; as in Alvarez & Cavanaugh, 2004), can be explained in favor of a simple, relatively high fidelity “slot” model, where the object is stored well (complex or not), but where participants are just really bad at making comparisons (Awh, Barton, & Vogel, 2007). I wonder if something similar might occur in the phonological loop?
  3. “The probability of losing a phonological feature which discriminates the item form other members of the memory set will be greatest when the number of discriminating features is smallest.” Two questions about this:
    1. If we’re considering counts of features, this would seem to make sense. However, what about a different dimension, like temporal extent of features? Would we expect a length-limited “tape” in the phonological loop to record with high-fidelity features that have relatively short duration, and perhaps to suffer when important information spans time longer than the tape?
    2. In aggregate, it might be true that items in a set discriminated by n features are less likely to be remembered than items discriminated by n + 1. However, shouldn’t we expect that features enjoy different weighting, and that this cannot be a simple, linear, additive model? For some reason, I’m thinking of the Family Guy use of the utterance “Cool Whip” with the initial consonant in whip oddly aspirated. (Here’s where my ignorance of linguistics starts to show.) In English, aspirated consonants are not generally contrastive (right?), but might we expect sometimes that violations of expectations laid down by the distribution of our experience to be quite marked, and notable even if it’s just a single feature? Similarly, if one lives in a society where post-vocalic /r/ is some marker of group membership, or status, or whatever, might we expect this feature to carry more weight than some other, more culturally neutral, single feature?

Deviation change test in R

The deviance change test is one method of comparing nested models. A significant change test indicates that the model containing more parameter estimates accounts for significantly more variability in the outcome variable than the model containing fewer parameter estimates.

If the number of parameters that differ between two models is limited to one, the deviance change test can be a method of determining if the contribution of a predictor variable is significantly greater than nothing.

Deviance tests only work with nested models. Models are nested if all of the models effects (fixed or random) in the simpler model are also included in the more complex model. In other words, the “bigger” model is achieved simply by attaching one or more additional predictors to the “smaller” model.

Here’s a little function to get the job done:

deviance <- function(a,b) {
 diffneg2LL <- (-2*as.numeric(logLik(a))) - (-2*as.numeric(logLik(b)))
 dfneg2LL <- (attr(logLik(b), "df") - attr(logLik(a), "df"))
 return(1 -pchisq(diffneg2LL, dfneg2LL))
}

Once this is defined in your script. You simply have to pass it two model objects. The smaller, more reduced model as “a” and the bigger, fuller model as “b.” You’ll need to have created the two models using lm or lmer or some such.

Here is an example of two, nested models. These are both multilevel models, one with random slopes and intercepts, the other with just random intercepts.

full <- lmer(Y ~ X*W + (X|GROUP), data = dataFrame)
reduced <- lmer(Y ~ X*W + (1|GROUP), data = dataFrame)

Where Y is the outcome variable, X is the level 1 predictor (often a continuous measure of an individual within a group), W is the level 2 predictor (often a category describing groups), and GROUP is often some ID variable unique attributed to each group. dataFrame holds all the values: Y, X, W, and GROUP will be columns in this data set.

Once the models and deviance function are defined in your script, you just need to call the function and pass it the model objects.

deviance(reduced,full)

What does the script do? Using the logLik function, it finds the log-likelihood of each model. These are each then multiplied by negative two. The -2 log likelihood is known in the R package lme4 as the “REML criterion at convergence.”

Because the fuller model includes additional predictors, it counts for more variability than the reduced model (or, at worst, no less variability). This smaller -2LL value is subtracted from the larger and stored as a variable, diffneg2LL. (It’s probably not the most streamlined approach to create additional variables like this, but I opted for explicitness over efficiency here.)

The logLik function also ends up determining the degrees of freedom for the model. The fuller model includes more df when calculating the log likelihood, as it contains more parameters. In the case of the two models above, the full model has two more degrees of freedom than the reduced model, because the reduced model lacks both an estimate for the variance for the random intercepts and an estimate of the covariance between intercepts and slopes.

The difference in degrees of freedom between the two models is equal to the number of parameters that differ between the models.

-2 log likelihoods are chi-square distributed, and the degrees of freedom for the chi-square distribution is the number of parameters that differ between the two models. The last line of the function, then, calls the pchisq function, passing it the -2LL value and the number of parameters. Subtract this number from 1, and you get a p-value indicating the probability of obtaining a -2LL this large (or larger) if the parameters added to the fuller model really contributed nothing to improving predictions.

Ontogenetic niche

West, M. J., & King, A. P. (1987). Setting nature and nurture into an ontogenetic niche. Developmental Psychobiology, 20, 549-562.

All organisms inherit parents’ genes, but many also inherit parents, peers, and the places they
inhabit as well. We suggest the term ontogenetic niche to signify the ecological and social legacies that accompany genes. A formal name is needed to give the idea of the inherited environment equal status with i�s conceptual cognates; nature and nurture. We argue here that increased recognition of the inherited environment facilitates unification efforts within the developmental sciences by emphasizing the affinity, rather than opposability, of ontogenetic processes.

The authors are rather Whorfian in their motivation:

The addition of a [ontogenetic] niche might seem only a semantic art. We argue later that it is not, but such an interpretation does not bother us because words matter greatly. If we are told nature and nurture compete, we assume divisibility, and we look for the strong and the weak. If we are told they are rivals, we go so fare as to create numerical scores to determine which prevailed. But if we are told nature and nurture are allies, what different ontogenetic process might be proposed and what metrics might evolve to measure their collaborative effects?

Ontological niche as legacy. Parents bequeath an endowment of resources, peers, social standing, and customs at least as important for inheritance as their endowment of genetic material. In this conception, the authors adopt a sense of environment as habitat, with an ecological approach to “nonarbitrary connections between species-typical surroundings and species-typical behavior.” This is such an important point. We often point to things like language or bipedal locomotion in humans as utterly genetic–they are universals, after all, so they must be evolved, “nature” traits. This ignores, of course, that there are a lot of universals of early experience for human kids, and this commonality of experience may be as contributory as any genetic predisposition. Also, this makes me think of B.I.Z.A.R.R.E. chimpanzees.

Ontological niche as link. Because the niche is not just the physical environment, but the social, “niches rest on transgenerational social dynamics.” Our niche reflects our dependencies on one another. In our early niche, we are nourished and cared for. We learn appropriate and inappropriate social behaviors, food preferences, sexual behaviors, parenting skills, Parents nurture offspring; responsive offspring nurture good parenting.

Ontological niche as way of life. A niche “specifies the behavioral adaptations of its occupants.” For mammals, certainly, and likely other critters, play is a behavior often thought of as an ontological adaptation, a chance for the young to, relatively safely, practice and perfect the behaviors (social aggression and cooperation, mating, food gathering, hunting, tool use, etc.) that would be necessary as an adult. The authors warn us, though, of seeing the world of the child through “adultomorphically biased lenses” (yes!).

Watching animals play is seeing them at their best. It is to view professionals in the game of growth.

The authors quote Piaget: play and youthful behaviors “supply evolution with its principle motor.” That is, they argue, playful behaviors, with their inherent variety and novelty are a force for breaking tradition and changing culture. Perhaps. Though I’d note that, in the motor learning literature, for example, the best way to establish a strong schema for an action is through varied practice. That is, learning of the “core” of an action is strongest when we around the that core a bit–rather than massing practiced instance of the core behavior itself. Perhaps the variety we see in play, dancing around the adult forms, is a way to cement culture as much as break it down.