Writing Sample I, from “Semi-Supervised Learning for Connectionist Networks”

From Robare, R.J. (2010). Semi-supervised learning for connectionist networks. Unpublished doctoral dissertation. Graduate School and University Center, City University of New York.

At the environmental level, language learning is increasingly being shown to be a highly social process. According to some researchers, social cues to word meaning begin to be used between 10 and 19 months of age, supplanting perceptual salience, which is a more dominant cue in infants aged 10 to 12 months. The changes appear to be gradual; 12-month-olds are better at making inferences from social cues than 10-month-olds, though they are not yet improved enough to learn mappings from such cues (Hollich, Hirsh-Pasek, & Golinkoff, 2000, cited by Hirsh-Pasek et al., 2004). At about the same age, 13 to 25 months, children are increasingly willing to extend words to perceptually dissimilar exemplars, again showing the decreasing importance of perceptual salience (and, perhaps, the shape bias?) at this time.

An excellent demonstration of the importance of social-environmental cues is seen in studies of the relationship between caregiver-child joint attention and lexical acquisition. In one longitudinal study (Carpenter, Nagell, & Tomasello, 1998), joint engagement of attention was measured in time duration and number of episodes of looking over a ten-minute observational period. Based on the amount of time the infants spent looking at the same objects as their caregivers in the experimental setting, infants were divided into four groups: early, middle, and late (referring to the month at which joint engagement was observed), and never. Mothers made utterances of three kinds: follow, in which they named an object to which the child was already attending; lead, in which they used an object name to draw the child’s attention to an object; and other, which were all other kinds of utterances. Follow and lead utterances did not differ in frequency, although Nelson (1988) had found that adults infer the object of children’s attention in naming, and not the other way around. Vocabulary comprehension and production were measured with the Communicative Development Inventory (CDI), which surveys parents on the words their children understand and produce at various ages (Dale & Fenson, 1996; Fenson et al., 1993), at several intervals. There were significant positive correlations of joint attention at 11, 12, and 13 months of age with word comprehension between 9 and 15 months, echoing the finding that social cues to word meaning begin to be used between 10 to 19 months (Hollich, Hirsh-Pasek, & Golinkoff, 2000, cited by Hirsh-Pasek et al., 2004). The early attention group comprehended significantly more words, and had a significantly greater rate of vocabulary increase, than the late attention group (Carpenter et al., 1998). Lexical acquisition is thus shown to be highly related to the caregiver’s behavior, which can be considered part of the language learner’s environment.

It is unlikely, however, that words used during periods of joint attention account for all word learning. Joint attention is relatively infrequent (Hoff & Naigles, 2002, cited by Woodward, 2004), so it is likely that word learners also use non-joint-attention contexts to interpret words (Akhtar, Jipson, & Callanan, 2001, cited by Woodward, 2004). (One example of this may be mutual attention, in which the mother and child pay attention to each other.) Word learning when caregiver and child do not share attention also fits cross-linguistic accounts of cultures in which the Western practice of directly giving an infant the name of an object does not occur.

The frequency with which a caregiver uses a word – in other words, the frequency of that word in the learner’s environment — also impacts the learning of that word (Goodman, Dale, & Li, 2008). A direct test of the hypothesis that more frequent exposure to a word leads directly to an earlier age of acquisition for that word used the Communicative Development Inventory (CDI; an inventory by which parents can record which words their child understands at which ages; Fenson et al., 1993) and the CHILDES database (MacWhinney, 2000), larger databases than had previously been used to look at the relationship between frequency of caregiver use and child learning (Goodman, Dale, & Li, 2008). The authors additionally assert a conservative measure of input frequency, because of the variability in frequency across caregiver-infant pairs. Finally, the large databases eliminate the effects of particular pairs focusing on particular categories or types of language, for example, referential (using words to name objects and actions) versus social (using words such as “hi” and “bye”). The CDI was used to determine the age of acquisition for various words, and the CHILDES corpus to estimate the frequency with which caregivers used those words. The study of these databases showed that overall, noun labels were used least frequently by parents but learned earliest by children, as opposed to closed-class words (e.g., function words and pronouns), which were used most frequently and learned the latest. Note that there are prosodic differences between open and closed-class words as well as frequency differences; the authors’ point is that words used more commonly by caregivers appear to be learned more easily, and the only exception to this is in the relatively unstressed closed-class words. The only words used frequently by caregivers but not learned by the age of two (the oldest age examined in this study) fell into this closed-class category. No words learned by age two were absent from the measure of caregiver use. In other words, according to these data, children did not learn words their caregivers did not use. Within lexical categories, however, the expected effect held true, with children learning earliest the words which occurred most frequently. These correlations were stronger using frequencies derived from CHILDES than from comparative databases such as the Francis-Kucera frequency norms, which are derived from written texts (Kucera & Francis, 1967). Goodman, Dale, and Li (2008) additionally suggest that frequency becomes more important for nouns but less important for function words over the course of development, but this is not confirmed and parsing this change in the importance of frequency for various word classes requires additional study (Goodman, Dale, & Li, 2008).

Of course, the influences of the environment interact with object and category properties during word learning. Labeling can serve as a cue for the learner to attend to an object that has appeared consistently while the label was repeated. This finding does not hold true for verbs, reinforcing the special status of object names in word learning. (Echols, 2002, cited by Echols & Marti, 2004).

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: