Chapter 7 General discussion
This study examined the development of familiar word recognition over the preschool years. The word recognition data came from a visual-world eyetracking experiment which recorded children’s fixations to images in response to prompts like see the bear. The trials featured a target noun (e.g., bear) along with a phonological competitor (bell), a semantic competitor (horse), and an unrelated image (ring). To describe children’s word recognition ability, I analyzed how the probability of fixating on the target image changed over the time course of a trial. The presence of the competitor images also allowed additional analyses about children’s sensitivities to the phonological and semantic competitors. The experiment was conducted as part of a three-year longitudinal study; children were 28–39 months-old at the age 3 visit, 39–52 at age 4, and 51–65 at age 5. The longitudinal design allowed me to describe developmental changes in word recognition.
7.1 How to improve word recognition
Children showed year-over-year improvements in word recognition, as measured by average looking probabilities, peak looking probabilities, and the rate of change in looking probabilities. Children became more reliable, less uncertain, and faster at recognizing familiar words as they grew older. At the same time, children also became more sensitive to the phonological and semantic competitors, compared to the unrelated images. With each year, children looked more to the target image, but when they erred, they were more likely to err on a lexically relevant word.
We can interpret these developmental patterns in terms of lexical activation and processing dynamics. In this task, children hear a stream of speech and activate some phonetic, phonological, lexical, and semantic representations that match the speech input. As they hear more of a word, the activation builds until a particular word is favored, and children shift their gaze onto the named image. Let’s imagine that we have to engineer this system. To make word recognition more efficient, we have to find ways to increase the relative activation of the correct word. In particular, we can boost the strength of connections so that activation can propagate more quickly through the system, and we can also allow inhibition among competing words so that the correct word can win out over its competitors more quickly.
The results from these studies indicate that children become more efficient at activating the target word and related words over the preschool years. As they grew older, children were faster to look at a named image and more likely to fixate on the phonological competitor (compared to the unrelated image). These two findings reflect changes in how partial acoustic information can propagate to activate phonologically plausible words. The phonological competitors shared the same syllable onset as the target noun (e.g., dress–drum), so the early part of the word matched both words. That children became more sensitive to the phonological competitor means that they learned and somehow encoded the phonological similarities among words because part of a word could activate a neighborhood of phonologically plausible matches. This developmental change supports faster word recognition because the listener can channel activation to relevant words more quickly. A similar line of reasoning applies to the semantic competitors: Relative looks to the semantic competitors increased with age, suggesting that children had learned semantic connections among words and activated semantically related words during word recognition.
The other mechanism we might tune to improve word recognition is inhibition. Children’s looks to the phonological or semantic competitors were temporary: Looks increase to some peak level and then quickly decrease. Behaviorally, the drop in looking probability reflects the rejection of an interpretation: for example, a child hears “dr”, shifts looks to dress, but hears “um”, revises the interpretation and jumps to drum. We can read these corrections as evidence for an inhibitory process: Corrections indicate a change in relative activation where a different word overrides an initial interpretation. But the evidence for developmental changes in lexical inhibition from these data was scant. The rate of rejection of the phonological competitor—that is, how quickly looks fall from their peak value—did not change from age 4 to age 5, although the rate did increase for the semantic competitor from age 4 to age 5. Preschoolers did demonstrate inhibition by revising their interpretations of nouns, but there were no clear developmental changes in inhibition.
Previous simulation work can help identify more specific mechanisms at play. McMurray, Samelson, Lee, and Tomblin (2010) used the TRACE model of word recognition (McClelland & Elman, 1986) to simulate looks to a target and phonological competitors (cohorts and rimes) in adolescents with specific language impairment. The authors tuned a number of model parameters and analyzed how those changes affected simulated looks to the target and competitors. In the current dataset, I observed a developmental trend where the relative looks to the phonological competitors peak higher each year. In those TRACE simulations, looks to the cohort competitor peak higher if 1) the rate of lexical activation increased, 2) the rate of lexical decay decreased, or 3) strength of lexical inhibition decreased. Of these options, the growth curve for the decrease in lexical inhibition best matches the shape of the current data. The similarity does not mean that children inhibited words any less as they grew older. That would be too simplistic: Developmental changes in preschoolers are the result of simultaneous changes in many mechanisms. But those simulation results suggest that an increase in lexical inhibition is not one of the key developmental changes in preschoolers’ word recognition.
7.2 Learn words and learn connections between words
Preschoolers showed increased activation of the target noun and semantically and phonologically related words but little developmental change in lexical inhibition. Paired with the findings from older children, these results lead to a compelling developmental story. Rigler et al. (2015) compared 9- and 16-year-olds on a Visual World word recognition experiment with phonological (cohort and rime) competitors. The younger children were slower to look to the target image and showed more looks to the competitors. The implications are that children’s word recognition is still developing in late childhood and that in particular, children’s inhibition of lexical competitors became stronger with age.
The current study with 3-, 4-, and 5-year-olds followed a different pattern: Relative looks to the competitor images increased with age. Taken together, these two studies suggest an interesting progression for the development of lexical processing. During the preschool years, children learn many, many words, and they establish phonological and semantic connections between these words. These connections support the immediate activation of neighborhoods of related words. Later childhood, based on the Rigler et al. (2015) findings, then is a time for refinement of those connections so that sensitivity to the competitors decreases. This refinement could follow from more selective activation channels, increased lexical inhibition, changes in resting activation (to favor more frequent words), or likely a combination of these factors.
7.3 Individual differences are most important at younger ages
Another dimension of this study concerned individual differences in word recognition. Some children were faster or more accurate during word recognition, and these children also were more likely to be faster or more accurate at later ages. The magnitude of these differences diminished over time, as children approached a more mature level of performance.
In terms of lexical processing dynamics, we might think of early differences as reflecting early differences in the burgeoning lexicon. Children may have different numbers of words, different degrees of experience with some words, less established connections among words, and at a lower level, different phonetic and speech perception abilities, given the links between speech perception in infancy and early vocabulary development (Cristia, Seidl, Junge, Soderstrom, & Hagoort, 2014). Differences in word recognition are greatest early on in development because this is when the differences among children’s lexicons are greatest. The task of learning new words, and more importantly, of developing representations and associations to organize words normalizes the early differences among children’s lexicons. That pressure would make the overall variability among children decrease over time while still preserving a relative ordering among children.
We can also interpret the predictive power of word recognition measures in terms of lexical processing and lexical organization. Correlations between word recognition performance and future vocabulary were strongest for the age-3 growth curve features, particularly for the peak probability of looking to the target word. The peak probability measures the overall certainty in word recognition and how strongly the target word is activated. Children with more efficient representations of familiar words at age 3 have a stronger foundation for encoding and integrating future words, and as a result, they showed larger vocabularies at age 4 and age 5.
Initially, I had expected processing speed—as approximated by growth curve slopes—to be the most predictive measure of vocabulary growth. Children who can more quickly recognize words, the reasoning goes, can take in information more quickly and devote extra processing resources towards learning.8 Processing speed was indeed correlated with future vocabulary size, yet peak probability was a stronger predictor of future vocabulary size. Granted, these two processing measures are highly related; to hit a higher peak by time x, a growth curve needs to start from higher baseline or have steeper slope. The idea of uncertainty suggests an alternative explanation of the predictive power of word recognition: Children who are more accurate (or less uncertain) during word recognition can extract and activate more information from the speech signal.
7.4 Limitations and implications
The discussion of processing speed and word recognition certainty highlights one limitation of this research: The experiment’s four-image, eyetracking-based design meant that a clean measure of processing speed was not feasible. Other eyetracking studies with two images can use the latency of how long it takes the child to shift between images as a measure like reaction time. This approach does not translate to the four-image design, as children can visit multiple images on their way to the target. Visual World studies with older participants can obtain an explicit reaction time measure by means of a mouse click or tap on a touchscreen, but those additional task demands may not translate to young children like those in this study. Thus, this study could not address directly whether the predictive power of word recognition performance reflects a more developed lexicon, a general reaction-time-like speed advantage, or both.
The lack of an explicit selection behavior, such as a mouse click, also means that word recognition accuracy was never directly measured but rather inferred. As a result, the interpretation of peak looking probability as a measure of word recognition certainty comes with a caveat: It reflects certainty averaging over many familiar words and maybe a few unfamiliar words. The idea is as follows: Suppose at age 3 a child does not know four of words well. If they had to click or tap an image, they would have to guess on these trials. We could exclude those trials where they guessed incorrectly, leaving just the trials where the child correctly recognized the word. In this scenario, we would be more justified in interpreting a growth curve peak as a measure of certainty during familiar word recognition because trials involving incorrectly identified words had been excluded. (It bears mentioning that explicit selection behaviors during the experiment are just one way to test a child’s knowledge of items; another is a receptive vocabulary test after the experiment which checks whether the child can point to the words from the experiment.)
As it stands here, there is no clear way to tease apart whether the lower growth curve peaks at age 3 reflected greater uncertainty during lexical processing or a greater number of words being unfamiliar (or unknown) to the children. I favor the former interpretation because these were highly familiar words and because children’s word recognition improved from age 4 to age 5. We piloted the images/words in two preschool classrooms, using only items that were at least 80% recognizable to children. But even if some words were unfamiliar at age 3, the number of unfamiliar words at age 4 was likely to be very small and therefore unknown words would have exerted a minimal effect on the lexical processing measures. The average peak looking probabilities increased by about .13 at age 4 (from .55 to .68) and by about .09 at age 5 (to .77)—the magnitude of these changes are comparable. Because children also showed improvements at age 5, when the effect of unknown words would be very small, age-related improvements in word recognition certainty likely reflect changes in lexical processing, as opposed to changes in the average mixture of known and unknown words.
The experimental design included semantic and phonological competitors on every trial, so isolating out the semantic and phonological competition effects required some subtlety. As a result, the lexical competition effects are only indirectly observed A more direct design would compare different types of trials: for example, trials with a target vs. three unrelated images intermixed with trials with a target vs. a competitor vs. two unrelated images. The trials also used different kinds of phonological and semantic competitors. For example, two of the phonological competitors rhymed with the target, so they could not be included in the analysis of phonological competitors (which focused on just competitors with the same syllable onset as the target). The current design limited the number of trials that could be used in the analyses of the competitors and weakened the power of the analyses.
A related limitation is that the phonological competitors used here are weak competitors. Adult studies tend to use phonological competitors with substantial overlap between the target and the competitor. For example, the landmark study of adults by Allopenna, Magnuson, and Tanenhaus (1998)—which showed that participants’ eyetracking probabilities matched lexical activations from the TRACE model of word recognition (McClelland & Elman, 1986)—used target–cohort pairs that shared a whole syllable: beaker–beetle, candle–candy, carrot–carriage, castle–casket, dollar–dolphin, paddle–padlock, pickle–picture, sandal–sandwich. With this degree of overlap, there is much more phonological and temporal ambiguity for the cohort to build up activation and compete with the target. In contrast, the words used in this study were all one syllable and the amount of overlap was limited to syllable onset (e.g., flag–fly, pen–pear). This reduced overlap limits the degree over temporal ambiguity and thus limits the degree to which the competitors can participate in lexical competition. These words were weak phonological competitors, compared to others studies in the adult literature. As a result, the brief advantage of the phonological competitor over the unrelated word may underestimate children’s sensitivity to phonological competitors: Preschoolers probably will show much more interference from competitors that have a larger degree of overlap. Moreover, with more interference from the competitors, individual differences could emerge more clearly so that child-level measures like speech perception can predict processing differences.
A final limitation includes the changes in the experiment procedure over the course of the longitudinal study. From age 3 to age 4, we re-recorded the stimuli (with the same original speakers) so that the noun durations between the two different dialect versions of the experiment were similar. From age 4 to age 5, we also shortened the duration of the trials by removing attention-getting prompts (e.g., this is fun!) from the ends of the trials. These small procedural changes mean that year-to-year differences do not reflect pure development differences. It is implausible, however, that the robust year-over-changes owe more to procedural changes than a year of learning and language development.
The findings from this study have implications for our understanding of word recognition and word learning. The first is the overall developmental narrative. Preschool children become better at recognizing words by learning similarities among words and using those similarities to activate neighborhoods of lexically relevant words. Rather than just measuring vocabulary size, word recognition reveals how well words have been integrated into the lexicon. The developmental trends here show that familiar words become more integrated and more connected over the preschool years. Even if a child knows a word at age 3 well enough to recognize or express it, their knowledge of the word will strengthen over time as the word develops connections to other similar words.
From this perspective, we can think of individual differences in word recognition as differences in lexical development. Variability in word recognition diminishes over time, so that differences are more predictive and discriminating at younger ages. Thus, if we wanted to intervene on word recognition, these results indicate that early intervention is better and that intervention should build connections among words and should target words that build onto existing semantic and phonological networks. The natural closing of gaps in word recognition performance with age, however, suggests that word recognition in and of itself may not be an important intervention target. Rather, word recognition measures should serve to supplement other vocabulary measures as an indicator of lexical processing and lexical integration.
References
McMurray, B., Samelson, V. M., Lee, S. H., & Tomblin, J. B. (2010). Individual differences in online spoken word recognition: Implications for SLI. Cognitive Psychology, 60(1), 1–39. doi:10.1016/j.cogpsych.2009.06.003
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
Rigler, H., Farris-Trimble, A., Greiner, L., Walker, J., Tomblin, J. B., & McMurray, B. (2015). The slow developmental time course of real-time spoken word recognition. Developmental Psychology, 51(12), 1690–1703. doi:10.1037/dev0000044
Cristia, A., Seidl, A., Junge, C., Soderstrom, M., & Hagoort, P. (2014). Predicting individual variation in language from infant speech perception measures. Child Development, 85(4), 1330–1345. doi:10.1111/cdev.12193
Allopenna, P. D., Magnuson, J. S., & Tanenhaus, M. K. (1998). Tracking the time course of spoken word recognition using eye movements: Evidence for continuous mapping models. Journal of Memory and Language, 38(4), 419–439. doi:10.1006/jmla.1997.2558
“The infant who identifies familiar words more quickly has more resources available for attending to subsequent words, with advantages for learning new words later in the sentence, as well as for tracking distributional information about relations among words… Being slow to identify the referent of a familiar word could interfere with lexical activation and impede success in tracking distributional regularities and managing attentional resources in real time (Evans, Saffran, & Robe-Torres, 2009)” (Fernald & Marchman, 2012, p. 217).↩