Chapter 13 General discussion

This experiment tested how children responded to familiar words, one-feature mispronunciations of those familiar words, and unambiguous novel words. These three types of stimuli allow us to examine features of children’s representations of familiar words and their processing of unfamiliar words.

13.1 A lexical processing account of the results

Children were sensitive to mispronunciations of familiar words. In terms of lexical processing, the mispronounced syllable onset leads a child down a phonological garden path. The “d” in dirl activates a neighborhood a “d”-initial words. The rest of the syllable, however, provides information needed to activate girl. Children therefore are slower to build up activation of the word because of the onset-mismatch, and they show less activation overall because the spoken word only matches the rime of the word. Put differently, they get a late start and have to work with a poor-fitting form of the word. The finding that children are better at processing mispronunciations at age 5 suggests that older children are better able to build up more activation to these candidate rime words.

Children also processed mispronunciations differently based on the image they were fixated on. One complication with a simple lexical processing explanation is that this is a two-image task. Children have ample time to view each image before the noun onset and build up expectations about the words they might hear named. This possibility leads me to speculate that children might be more uncertain in the trials where they hear “shoup” while fixated on soup because they build up the activation of soup. This prepotent activation would make the mismatch from the mispronunciation more severe leading to greater uncertainty. This explanation, however, implies some kind of inhibition where soup suppresses the activation of “sh”-initial words. The results from Study 1, which do not provide any evidence for changes in inhibition, make me skeptical of this explanation. Thus, the effect of children’s fixation location on their immediate response to speech provides an avenue for further research in this area.

13.2 A nonword is just a word you haven’t learned yet

I hypothesized that the nonword condition might be more difficult than the real word condition, particularly for younger children as seen in Bion et al. (2013). This prediction did not bear out at all, and indeed, there was an advantage in the nonword condition in later ages. Part of this advantage may be a novelty bias. Mayor and Plunkett (2014) used the TRACE model of word recognition (McClelland & Elman, 1986) to simulate these kinds of situations. In one set of simulations, the novel object receives a novelty/salience boost to resting activation. During presentation of the nonword, none of the child’s known words build up enough activation to overtake the novel word. In an alternative set of simulations, the novel word is added as to the lexicon as a low-frequency word, and the absence of competition of any familiar words causes the novel word to win out. In both of these accounts, children can quickly associate the novel object with a nonword because there are no familiar words to interfere with the processing.

The results here probably support both processing accounts. During data reduction, I separated trials based on initial-fixation location because children become increasingly likely to start nonword trials on the novel object. This bias affected 50% of nonword trials at age 3, 55% at age 4, and 59% at age 5. Thus, there is a novelty preference that for these trials gets stronger with development. Plus, children are also learning during this task: At age 5, children were on average were able to recall the unfamiliar image paired with 5/6 nonwords. This learning is consistent with the strategy of simulating a nonword as a low-frequency lexical item.

The findings of Swingley and Aslin (2007) can help us understand the mispronunciation retention results. In that study, toddlers were better able to retain unneighbored nonwords (like shang or meb) compared to neighbored nonwords (mispronunciations like tog [dog] or gall [ball]). They concluded that part of fast-referent selection involves a probability calculation in which children “evaluat[e] the likelihood that an utterance conveys a new word”. In this experiment (that is, Study 2), we presented an image of the familiar (mispronounced) object during the mispronunciation trials, which further reduces the likelihood that the mispronunciation indeed reflects a new, as-yet unlearned word. Thus, the children in the retention task averaged around 3–4 mispronunciations correct because those words had a low likelihood of conveying a new word. In fact, the more children discounted that probability—that is, the more they looked to the familiar word on the mispronunciation trials—the less likely they were to retain the mispronunciation items.

The interference from the familiar word on encoding and retaining the mispronunciation and image pairing is beneficial for word learning. Suppose that wordform-object associations are encoded in proportion to the activation of the wordforms and objects. For the unambiguous nonwords, the strong activations allow the associations to be built up quickly, whereas for the mispronunciations, most of activation is spent on the familiar word, meaning that the association is built up more slowly. The use of the word likelihood in Swingley and Aslin (2007) also suggests an analogous framing using Bayesian terminology: Known words are like priors which influence (regularize) how unfamiliar words are interpreted, so it takes more exposure to overcome these priors and encode a mispronunciation as a new word. In both framings (lexical activation or Bayesian inference), children showed less retention of the mispronunciations because the words the children know made the mispronunciations ambiguous and that ambiguity made each instance of the mispronunciations less informative. The slowed encoding is a good thing, because spurious associations from mispronunciations should not be learned. McMurray et al. (2012) argues the same point: “Slow [word] learning may be more optimal in that it prevents children from committing too strongly to a single (perhaps erroneous) mapping before they have enough data” [p. 870].

13.3 Limitations and implications

The primary limitation for this study is that it applied a procedure designed for toddlers (White & Morgan, 2008) on preschoolers. That is to say, the two-image task was too easy for there to be large year-over-year developmental changes in children’s performance. Children were successful at recognizing real words and fast-selecting referents for nonwords at age 3, and by age 4, a quarter of the children performed at ceiling on the nonword condition. The most difficult condition, based on the absence of ceiling effects, was the mispronunciation condition in which children showed much more uncertainty on how to process these words. For this condition, a developmental trend was also observed where preschoolers at age 5 had a larger preference for the familiar object on the trials.

Another limitation here is that the mispronunciations are a particular kind of mispronunciation: One-feature onset mismatches. That initial segment sets the stage for the processing of a word as it activates a word’s onset neighbors. Just as it takes longer for a rime to influence processing of word—as evidence needs to pile up from multiple compatible sound in order to overcome an initial mismatch—it should also take longer for a child to recover from an onset-mispronunciation. Conversely, we might also expect vowel or rime mispronunciation to be less disruptive for word recognition because of the useful information on the starting segment. Indeed, Swingley (2009) demonstrates that this is probably the case. That study tested onset and coda mispronunciations in an eyetracking task that used two familiar objects. Both adults and toddlers responded to coda mispronunciations (e.g., “dut” for duck) by showing delayed looks away from the (mispronounced) target image.

As I note in Appendix E, the small repertoire of mispronunciations is another limitation for this study. It is conceivable that specific mispronunciations change in severity with development, even though the canonical form of the word is a familiar and well known word. In this study, for example, the distance between the girl and dirl increased each year, even though girl was well known to three-year-olds, but this distance was driven by children becoming faster and more efficient at recognizing girl. Put another way, a mispronunciation penalty also reflects children’s knowledge of the canonical word. In this data, the age-5 rice receives as many looks as the age-4 dirl. Do children know rice less or accept dirl more? With so few items, it is unclear whether these differences are accidental or systematic.

One implication for this research is that children’s recognition of familiar words and referent selection for nonwords improved a modest amount over the preschool years. Although these tasks can be done by toddlers, full mastery does not begin to emerge until age 4. This finding agrees with the main finding from my analysis of familiar word recognition: Although children can ostensibly know a word very well, their recognition and processing of that word can still improve during preschool.

The finding that mispronunciations were harder to retain than nonwords also has implications for teaching or intervention. Namely, teaching words that are confusable with known or relevant words is more difficult because a child’s known words can influence whether the child accepts a taught word as a novel word.

References

Bion, R. A. H., Borovsky, A., & Fernald, A. (2013). Fast mapping, slow learning: Disambiguation of novel word–object mappings in relation to vocabulary learning at 18, 24, and 30 months. Cognition, 126(1), 39–53. doi:10.1016/j.cognition.2012.08.008

Mayor, J., & Plunkett, K. (2014). Infant word recognition: Insights from TRACE simulations. Journal of Memory and Language, 71(1), 89–123. doi:10.1016/j.jml.2013.09.009

McClelland, J. L., & Elman, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18(1), 1–86. doi:10.1016/0010-0285(86)90015-0

Swingley, D., & Aslin, R. N. (2007). Lexical competition in young children’s word learning. Cognitive Psychology, 54(2), 99–132. doi:10.1016/j.cogpsych.2006.05.001

McMurray, B., Horst, J. S., & Samuelson, L. K. (2012). Word learning emerges from the interaction of online referent selection and slow associative learning. Psychological Review, 119(4), 831–877. doi:10.1037/a0029872

White, K. S., & Morgan, J. L. (2008). Sub-segmental detail in early lexical representations. Journal of Memory and Language, 59(1), 114–132. doi:10.1016/j.jml.2008.03.001

Swingley, D. (2009). Onsets and codas in 1.5-year-olds’ word recognition. Journal of Memory and Language, 60(2), 252–269. doi:10.1016/j.jml.2008.11.003