Chapter 15 General discussion of both studies

In this chapter, I integrate results from the two studies. First, I describe the mechanisms underlying children’s word recognition. I then briefly discuss some clinical implications of this research, and I outline the main contributions of the research.

15.1 Mechanisms of word recognition

What cognitive or word-recognition mechanisms can explain the data observed from these two studies? These are the essential findings that the model of word recognition needs to account for:

• Developmental improvements in familiar word recognition
• Early advantage of phonologically similar words (over unrelated words)
• Late advantage of semantically similar words (over unrelated words)
• Developmental changes in the advantage of these similar words
• Disrupted processing of onset-mispronunciations
• Effortless processing of unambiguous nonwords
• Individual differences in familiar word recognition

As a baseline for word recognition mechanisms, I will start with 1) a continuous activation model 2) that uses different levels of representation—in other words, TRACE (McClelland & Elman, 1986). In my preceding interpretations of the data from Study 1 and Study 2, I assumed a TRACE-like architecture, so it is helpful to briefly review what this model does.

TRACE interprets an input pattern by spreading activation (energy) through a network of processing units. The pattern of activation over the network is its interpretation of the input signal, so that more active units represent more likely interpretations. Over many processing cycles, the network propagates energy among its connections until it settles into a stable pattern of activation. (Activation also decays over cycles so that the model can start from and return to a resting state.) This activation process is continuous; the model’s interpretation evolves continuously. We can ask at any point during (or after) presentation of a word what the model’s interpretation of that word is. Thus, the listener does not need to hear a whole word to generate a plausible guess for that word (e.g., Fernald et al., 2001).

The model involves three levels of representation: perceptual/phonetic features, phoneme units and lexical units. The input for TRACE is a mock-speech signal that activates the perceptual feature-detectors. These units respond to phonetic features like voicing or vocalic resonance. The perceptual units activate phoneme units, and the phoneme units activate lexical word units. For example, the bundle of features representing /b/ would activate /b/ but to a lesser extent also activate the phonetically similar /d/ (different place), /p/ (voice), /v/ (manner), or /m/ (nasality). The initial /b/ sound activates a neighborhood of words containing /b/, and the phonetically similar phonemes like /d/ or /p/ also activate compatible similar words, albeit to a weaker extent.

The combination of continuous processing and these levels of representation means that ambiguities can arise during word recognition. Suppose that after /b/, the sound /i/ arrives, activating a set of phoneme units and in turn activating words containing /i/. The sequence of /bi/ favors a particular neighborhood of cohorts: be, bee, beam, beak, beat, beetle, etc. At this point, however, the signal is ambiguous. Any of the words in the cohort are plausible interpretations, and more information is needed to refine the interpretation. In Swingley et al. (1999), 24-month-olds were slower to respond to trials of doggie versus doll, compared to doggie–tree or doll–truck trials, where the delay reflected the momentary ambiguity from the words sharing an onset consonant and vowel.

The mechanisms described thus far can account for the advantage of the phonological competitors over unrelated words from the first study. The initial phoneme in a word activates a cohort of words that share that sound, so the cohorts briefly represent more plausible interpretations of the target than words that are not phonologically related. A child acts on that early information and shifts their gaze to the phonological competitor.

Words in TRACE compete with each other through lateral inhibition, so that an active word will dampen the activation of other competitors. Inhibition allows the model to reinforce or revise an interpretation. In the earlier example, the arrival of /m/ after /bi/ would strongly favor beam as the most plausible interpretation of the word, and beam will inhibit the other candidates like beak or beat so that it can be the decisive interpretation of the word. The transient effect of the phonological competitor suggests lateral inhibition: The advantage of the phonological competitor over the unrelated word is short lived because the target word builds up activation and inhibits the phonological competitor.

To account for the effect of the semantic competitor, we need to make a few more assumptions. Semantic information is not explicitly included as a part of TRACE, but we can stipulate that semantic information is part of a word’s lexical representation. We also need a way for semantically related words to coactivate, so that hearing bee will generate some spurious looks to fly. In this case, we can assume that there are excitatory connections between semantically related words so that hearing a word also activates its semantic relatives. In my earlier discussions, I used the term cascading activation to describe this arrangement. For children to generate looks to the semantic competitor, they first need to build up activation of the word and that activation would cascade over to semantic relatives. The time course of cascading activation here is consistent with the late effects of the semantic competitor. The semantic competitor exerts an advantage over the unrelated word after semantic information comes online.

The relative advantage of the phonological and semantic competitors increased each year, as did children’s overall recognition of the familiar word. In other words, children became better at activating the target and the words related to the target. In Chapter 7, I argue that these developmental changes in the first study reflected stronger bottom-up phoneme–word connections (for greater activation of the target and phonological competitors) and stronger semantic connections between words. Alternatively, one might assume that phonologically similar words coactivate in a similar way as the semantically related words activate each other. The problem with this interpretation is that it would not resolve lexical ambiguity to have similar sounding words supporting each other. The phonological similarity between words lives not in the connections between them but in the phonemes that the words share and that mutually activate them. The phonologically related words compete with each other, and they may inhibit each other so that the most plausible interpretation can quickly suppress competing interpretations. For these data, I did not observe any developmental changes in inhibition, so I favored an interpretation that focused on stronger bottom-up connections. (I discuss inhibition more below when I discuss open questions.)

One prediction of TRACE is that rhymes and rimes (one-syllable rhymes) can affect word recognition. But these rhymes are at a disadvantage. Early in the processing of a word, all the action is in the bottom-up connections from the phonetic features to the phonological units onto the words. Cohorts show an early advantage in word recognition because they receive activation before lexical units start to inhibit each other. A rhyme mismatches the input from the start of the word, so it undergoes inhibition early on. But as the word unfolds, subsequent phonemes can build up activation of the rhyme word, and the word can overcome the initial disadvantage. Allopenna et al. (1998) found a strong similarity between TRACE’s activation patterns and adult listeners’ looking patterns. Namely, adults can hear beaker and look to the word, but they also might generate spurious early looks to a cohort (beetle) and late looks to a rhyme (speaker). (Anecdotally, my name is Tristan, but in grade school, I always snapped to attention whenever Kristen’s name was called.)

The mechanisms that predict how rhymes can engage in lexical competition also explain the disruptive mispronunciation effects observed in Study 2. The initial /s/ in suze sends the listener down a lexical garden path, activating /s/-initial words. The arrival of the rest of the word—plus the presentation of an image shoes onscreen—supports shoes as an interpretation of the word. But there is much less certainty in this situation. At age 3, I observed 80% looking to the image of the shoes for the real word shoes compared to 50% looking (to the shoes) for suze.⁹ There was a small developmental improvement for the mispronunciation and real word conditions. For example, at age 5, suze reached 60% looking to the familiar image and shoes reached 87% looking. Developmentally, children became more likely to activate the familiar word when given a mispronunciation, and this change likely reflects general improvements in activation efficiency. Gains in activation efficiency are consistent with the results for familiar word recognition in Study 1 where children showed increases in overall looking probability and in how quickly looking probabilities changed during a trial.

What about the effortless processing of the unambiguous nonwords? Surely, children do not have a lexical item geeve to activate the first time they hear the word. On these trials, however, the children did know sock and know that geeve was not the name for the sock, so they looked to the trolley instead. For McMurray et al. (2012), the problem facing a child is reference selection: Children have to select a visual referent for a spoken word. In their model, all words can refer to all visual referents initially, so the model has to prune away unnecessary connections to build up selective word recognition. Development of the sock-sock pairing pruned away other visual referents or words from activating sock. Thus, geeve is not likely to activate sock but the viability of a geeve-trolley pairing allows the child to select the correct referent for the nonword. In TRACE simulations, Mayor and Plunkett (2014) handled this situation by treating the nonword as a low-frequency word. In both situations, a novel nonword is recognized despite not being well known to the child. This recognition is possible because the new word is not affected by lexical competition from any other plausible alternatives.

This framework also allows us to account for the differences in retention for the nonwords and mispronunciations at age 5. In McMurray et al. (2012), learning was associative. The model developed connections between spoken words, lexical items, and visual referents when spoken words and visual referents occurred together, and each co-occurrence built up the connections. On the mispronunciation trials from Study 2, a child heard a mispronunciation of a familiar word and also saw an image of the familiar (mispronounced) word. On average, they tended to interpret the mispronunciation as the familiar word. Thus, the familiar word competed with the mispronunciation, leading the child to develop a weaker association between the novel object and the mispronunciation. The effect of looking behavior, where children who looked more to the familiar image on mispronunciation trials showed poorer retention, helps explain how the familiar image could impede the association of the mispronunciation and the novel object. In contrast, for the unambiguous nonword trials, children could associate the novel object and novel word more strongly. This difference in lexical competition manifested in the retention performance where children were better able to retain nonwords than mispronunciations.

So far, I have described a general framework of word recognition, and I claimed children’s developmental changes in word recognition reflect more efficient representations and activation pathways. I now describe task differences and individual differences under this framework.

Word learning is a matter of degree. I like to draw a distinction between “shallow” receptive knowledge and “deeper” expressive knowledge, based on the idea that recognition is easier than generation. But we can imagine a finer continuum with degrees of recognition ability. For example, a word can be recognized in one situation but may not be recognized in a more challenging situation. For example, McMurray et al. (2012) tested a word-learning model’s comprehension by simulating alternative-forced choice (AFC) tasks where a named target was displayed and pitted against visual competitors. The model showed graded performance, with better comprehension on 3-AFC (2 competitor) tests than 5-AFC tests, and better performance on 5-AFC tests than 10-AFC tests. Thus, the 4-AFC task in my first study provided a more challenging word-recognition environment than the 2-AFC task in the second study. For example, children demonstrated ceiling performance on the nonword condition at age 4, whereas children had room to develop each year in the 4-AFC task.

Individual differences in word recognition reflect differences in children’s lexicons and their lexical representations. Although all the words on the 4-AFC were familiar to preschoolers, children differed in their peak looking probabilities and rate of fixating on the target. In lexical processing terms, children differed in peak activation and the rate at which activation reached the target word. Differences in word recognition were stable from year to year. Even though all the children became faster, more reliable and more certain during word recognition with age, the children who were faster and more reliable at age 3 were also faster and more reliable at age 5. The children who performed better at age 3 had more familiarity with the words and more reliable representations of them—thus, these children had a head start and they built on top of that advantage as they grew older. This interpretation can also account for how word recognition performance at age 3 correlated with vocabulary scores at later ages.

15.1.1 Open questions about word recognition mechanisms

There are three immediate open questions from this research. First, how does lexical inhibition change over this developmental window? The results from Study 1 show that phonologically and semantically similar words become more relevant during word recognition as children grow older. (The words became more active, compared to the unrelated word.) I did not observe any clear changes in how quickly those words were rejected as possible interpretations of the input, and thus, I could not make any claims about the development of inhibition.

In principle, developmental changes could have been observed in this experiment. Lexical inhibition would affect how quickly the phonological competitor’s advantage decays as the target word becomes the favored interpretation. A developmental change in lexical inhibition would cause the competitor’s activation to decay more quickly (or more slowly) at older ages. But the growth curves observed here were parallel; they decayed at the same rate. Changes in lexical inhibition are detectable by an experiment paradigm like this one (with a target, competitor and an unrelated word), but in the present case, I did not observe these changes. Thus, developmental change in lexical inhibition during the preschool years remains an open question.

Based on other work, I expect older children to show greater inhibition. Rigler et al. (2015) showed that 9-year-old children were more sensitive to phonological cohorts and rimes than 16-year-old listeners, suggesting children need to develop inhibitory connections that suppress the interference from these words. Blomquist and McMurray (2017) used a cross-splicing paradigm to test lexical inhibition in 7–8-year-old versus 12–13-year-old children. In this paradigm, a target like cap is created by splicing an initial ca onset with a different token (ca(p)p), with a cohort competitor (ca(t)p) and a nonword (ca(k)p), the idea being that the sublexical information in the cohort splice will favor cat and therefore inhibit cap whereas a nonword splice cannot inhibit cap. This manipulation held in both groups, but the older children were more disrupted by the cohort splice. It would be revealing to see both paradigms applied to this age range. For the preschool years, however, the development trajectory seems to be the strengthening of connections so that the phonological competitors can participate in word recognition with later childhood being a time to develop inhibitory connections. In other words, a child has to develop sensitivity to cohorts first in order to demonstrate the ability to quickly inhibit them.

A second open question is when does a nonword engage in lexical competition and interfere with word recognition in children at this age. For adults, nonwords can affect processing very quickly. Kapnoula et al. (2015) used a cross-splicing paradigm with adults and observed that newly learned words compete with familiar ones immediately. Magnuson, Tanenhaus, Aslin, and Dahan (2003) trained adult participants with artificial lexicons and observed that after one day of training, cohort and rimes effects within the artificial lexicon were comparable, but after a second day, the cohort showed an early advantage. For preschoolers, I would expect them to show cohort and rime effects with enough training. The prediction is based on the competitor effects observed in the first study where age-5 children showed the early advantage of the phonological competitor over the unrelated word. Sensitivity to lexical inhibition via cross-splicing is an open question for preschoolers in general, even for familiar words. If lexical inhibition develops over later childhood, it is conceivable that preschoolers could show equal sensitivity to cross-splicing from cohorts and nonwords.

A third open question, given the previous discussion of models and mechanisms, is whether a word recognition model like TRACE can replicate the developmental changes observed here. There is no reason to assume that it would not be able to simulate the results from each year, given that it has been used to simulate word-recognition data from adults (Allopenna et al., 1998), adults with aphasia (Mirman, Yee, Blumstein, & Magnuson, 2011), toddlers (Mayor & Plunkett, 2014), and adolescents with specific language impairment (McMurray et al., 2010). These simulations have shed light on their respective listener populations. For the toddler data, Mayor and Plunkett (2014) had to use reduced lexical inhibition parameters in order to replicate graded mispronunciation effects of White and Morgan (2008), suggesting that lexical inhibition is not a crucial feature of toddler word recognition. McMurray et al. (2010) used TRACE simulations with different modeling parameters to test different theories of specific language impairment. Ultimately, they found that lexical decay—“the ability to maintain words in memory” (p. 23)—was the most important model parameter, implying that individual differences in word recognition for listeners with specific language impairment are rooted in lexical processes (and not perceptual or phonological ones).

For the current data, the goal of the simulations would be the developmental story: Which parameters would need to change each year to have the model match the empirical data? I would posit that the changes would involve some manipulation of the rate of lexical activation so that bottom-up information can activate relevant words the more quickly. I would also expect changes in the degree of lexical inhibition to play a role, based on the simulations in McMurray et al. (2010) in which cohort effects increased as lexical inhibition decreased. Even though I did not observe any changes in lexical inhibition in terms of how quickly the competitor advantages decayed in the first study, it is still plausible that lexical inhibition changes are needed to accommodate increased bottom-up activation.

15.2 Clinical implications

The results of Study 1 remind us that words are not simply acquired—they are recognized, learned, and integrated. In the first study, children’s recognition of highly familiar words improved each year. Children’s representations of familiar words will continue to develop, even when they ostensibly know the word. One might attribute this development change to improvements in visual processing, sensory processing, or some other nonlinguistic factor. This study cannot rule out those explanations. The increasing effect of the phonological and semantic competitors, however, suggests that changes in lexical representations are needed to explain these results.

Part of the promise of eyetracking-based research is that word recognition can predict later outcomes. One common conclusion in this research is that word recognition may provide an early screening tool: “[t]ime-course measures of comprehension in very young language learners could ultimately prove useful in improving early identification of children at risk for persistent language disorders” (Fernald & Marchman, 2012, p. 219). The developmental results here stress that such a tool has to be developmentally appropriate. Children’s processing of real words on the two-image task did not predict language outcomes, but the slightly more challenging nonword condition did yield a small predictive effect. For the more difficult four-image task, individual differences were greatest and most predictive at age 3 and the range of variability decreased with age. Thus, recognition of familiar words is perhaps best understood as a lexical measure that needs to be scaled with children’s vocabulary norms.

Finally, the observed difficulty of retaining mispronunciations emphasizes that children will err on the side of known words during nonword referent selection when the known words are plausible interpretations. In particular, it does not seem like a contrastive method of teaching, say, pear by having it compete against a known word bear would be effective because the known word would interfere with the encoding of the nonword.

15.3 Contributions

The most important contribution of this research is that children became more sensitive to phonological and semantic competitors as they grew older. When they erred, they were more likely to look to a relevant word. This result indicates that children improve in word recognition by being able to activate phonologically plausible words quickly, from partial information, and activate semantically related words on the basis of cascading activation. The developmental trend is that of more engagement (Leach & Samuel, 2007), with children developing the connections among related words and harnessing those similarities to their advantage.

Another contribution is the description of individual differences in word recognition: Namely, differences are stable over time but diminish in magnitude, so that early differences are more predictive than later ones. Although there has been ample evidence of how word recognition in toddlers predicted later outcomes, it was not clear whether those differences held over the preschool years. The results here indicate that the differences are task-specific: A more age-appropriate four-image task can better differentiate preschoolers than a simpler two-image one.

A final contribution comes from Study 2. This study was limited by how apparently easy it was for preschoolers. After all, it was a 2-AFC task where one of the images was always a familiar object and the other was an unfamiliar object. That limitation was revealing, showing that this design does not scale up for preschoolers developmentally. Children had mastered mutual-exclusivity-type referent selection on this task by age 4. Children could effortlessly associate nonwords to novel objects, provided that the nonword is not under competition from any known words, as was observed for the mispronunciations.