The Infant Lexicon

Islands to Networks

Preferential Looking

Quick Prime Task

Priming Results

What's in a Name?

What's in a Link


Phonological Priming

Cohort Effects

More Cohorts


Graded Sensitivity


A common assumption in cognitive psychology is that a proper understanding of the adult system is little informed by an appreciation of its developmental roots. Even in areas where clear-cut changes in capacity have been documented, such as the fine-tuning of infant sensitivity to native speech contrasts, a widespread view has a somewhat static flavour – that of selection from a universal list of pre-determined contrasts to achieve alignment with those appropriate to the mother tongue.

In this talk, I want to argue for a constructivist, dynamic view of development, one that situates the developmental process as central to our understanding of the adult mature state. I’d like to suggest that a proper understanding of the mature state can only be achieved through an appreciation of its developmental history, both phylogenetically and ontogenetically: that a fractionation of the adult system into its interacting component parts can be valuably informed by considering how that system was built.

The Infant Lexicon

The focus of this talk is the infant lexicon – a cognitive system which all commentators would agree undergoes rapid development during the first years of life. Even here, however, the approach has been been somewhat static in nature with lexical development viewed as a gradual accretion of labels that can refer to objects and events, as if the simple accumulation of these isolated label-object associations provides an adequate account of how the child builds an adult-like semantic system.

Of course, this is where the developmentalist must keep an eye on the mature state – the significance of a developmental trajectory is well-informed by an appreciation of its target destination. In this case, the target destination consists of an inter-connected system of words and their meanings.

The question arises as to how the child builds such an interconnected system: Is it adult-like from the outset with words added to the system in a constrained, systematic fashion? Or does the lexicon grow in a piecemeal fashion with structure only emerging later in development, perhaps in response to the “search and find” demands of a rapidly expanding vocabulary?

Islands to Networks

Recent mathematical modelling of vocabulary development (Hills et al., 2009, Cognition) exploiting techniques from graph theory, suggests a systematicity in the manner in which words are added to early vocabularies – words with shared semantic features or associative co-occurrence relations tend to have a head start over unrelated words.

Of course, mathematical modelling of this kind does not establish the “psychological reality” of the implied mental structures. What is required is experimental evidence that lexico-semantic relations btw words can influence children’s behaviour. In the remainder of this talk, I will provide an overview of some of the techniques we have developed in Oxford to investigate semantic development in early childhood. And to anticipate our findings, I will argue that we now have good evidence that such a system exists, at least as early as 24 months of age, and that inhibitory, competitive processes are engaged when these young children attempt to access the mental representations associated with words.
We use an adaptation of the Inter-Modal Preferential Looking (IPL) task. In the “standard version” of the task, which can be thought of as a stripped down version of the Visual World Paradigm often used in adult psycholinguistic studies, infants are presented with a pair of pictures on a computer monitor and an auditory stimulus over a centrally located loudspeaker. Preferential orientation to one of the pictures in response to hearing the auditory stimulus is taken to indicate a cross-modal association btw the visual and auditory stimuli. When the auditory stimulus is a referring expression and the visual stimuli depict familiar object categories then preferential fixations can index lexical comprehension.

In our studies, eye fixations are coded every 40ms when using static camera technology or every 8ms when using automatic eye-tracking technology. The “received wisdom” in this field of research is that infants and young children will show preferential fixation of a target image over a distracter when the target is named, but of course, ANY systematic asymmetry in looking behaviour has the potential to index cross-modal associations. Nevertheless, preferential looking towards a target object when it is named is the most commonly reported measure, though other measures such as accuracy of first fixation and latency to disengage fixations from distracters and targets can by highly informative.
In our adaptation of the Inter-model preferential looking task, infants are also presented with a pair of pictures, a target and a distracter, but these pictures are preceded by a PAIR of referring expressions separated by a short period of silence.

The referring expressions are either RELATED semantically and associatively according to the Birkbeck Word Association norms or completely UNRELATED. The second referring expression, a single word in citation form, always named one of the depicted images.

No stimuli are ever repeated for an individual infant and each infant was exposed to an equal number of related and unrelated trials. Across infants, images served equally often as targets and distracters. We tested 2 groups of infants, one at 18 months and the other at 24 months. Only trials including words that were known to the infant, according to parental report, were included in the analysis.

In all conditions and at both ages, infants spent proportionally longer time looking at the target picture than the distracter picture. At 24 months, infants preference for the target over the distracter was greater in the RELATED condition compared to the UNRELATED condition. No such systematic difference was observed for the 18 month olds.

Note that the “priming effect” observed for the 24 month olds seems to be attributable to a reduction in target looking in the unrelated condition rather than an increase in target looking in related condition, pointing perhaps to inhibitory effects btw unrelated referring expressions.

It’s worth noting though that both ages were more likely to fixate the target picture first in the related condition compared to the unrelated condition, suggesting that these measures may be tapping into different aspects of the priming process or the time course of word recognition by young children. There was no correlation btw overall vocabulary size in the 24 month olds and the size of their priming effect.

This study appears to provide support for the conclusion that hearing a related prime word permits efficient processing of a subsequent target word, thereby enabling rapid and prolonged fixation of a target picture compared to priming with an unrelated word.

The existence of a priming effect in 24 month olds provides evidence for the operation of a semantic system. However, it is unclear from these results whether the priming effect is mediated through the Target Word or through a direct impact of the related prime word on target picture looking. If the latter mechanism is responsible for the priming effect observed in the study, then we have little evidence for the existence of a structured lexico-semantic system in 24 month olds.

What's in a Name?

One way to address this is issue is to design an experiment in which infants don’t always hear the target named even though they may hear a related or unrelated prime word. So we carried out a similar experiment to the one I’ve just described but infants were presented with trials belonging to 4 different conditions in a 2x2 design in which the target was named or unnamed and the prime was related or unrelated to the target picture. In the two unnamed conditions the target label was replaced with the simple directive “Look”.

We expected the related prime – named target condition to yield similar patterns of looking as the previous experiment. Of course, in the unrelated prime, unnamed condition no systematic looking preferences could be expected because no information was being provided to the infant about the target picture. The focus of interest was infant performance in the related prime, target unnamed condition and the unrelated prime target, named condition. In this experiment, we also arranged for target and distracter images to have the same label onsets so that initial target consonants alone couldn’t be used to disambiguate the target from the distracter. We also used a 21-month old age group to determine whether priming effects could be observed at a younger age.

The pattern of results observed in this experiment were rather conclusive. 18 month olds showed no evidence of a priming effect – they showed similar target preferences in both the related and the unrelated prime conditions, just so long as the target was named. If the target went unnamed, no systematic target preference was observed. At 21 months though, infants only showed evidence of target recognition when the target was named after hearing a related prime. When the prime was unrelated or only a related prime word was presented, 21 month olds showed no evidence of target recognition. This pattern replicates the priming effect observed in the previous experiment and suggests that the prime alone is inadequate to drive a target preference, at least under the timing conditions used in this experiment.

These findings point to the conclusion that by the age of 21 months, infants have begun to develop a structured lexico-semantic system. This structure is evidenced by priming effects in the related compared to the unrelated priming condition, and the failure of the prime alone to induce target looking suggests that the priming effect is lexically mediated.

21 month olds failed to show target recognition, even though the target was named but preceded by an unrelated prime. This finding points to the unrelated prime interferring with target word recognition (we know from other experiments involving repetition priming or single target word presentation that 21 month olds infants show systematic target preferences).

It’s interesting that 18 month olds don’t seem to be susceptible to the interferring effects of an unrelated prime. This points to the emergence of a lexico-semantic system acting as double-edged sword. In its absence, the 18 month olds readily identify targets irrespective of the prime’s status. Their lexical entries seem to have an island-like status impervious to their neighbours. However, once a semantic system is set in place, it can work for you or against you.

What's in a Link?

In the previous experiments we have taken advantage of what in the adult literature is called the priming boost – our related word pairings involved items that were both semantically and associatively related. We decided to do this simply to enhance the chances we would find priming effects. Of course, it is informative to ask whether infants are sensitive to both thematic/associative AND taxonomic/semantic relations btw the words in their lexicons. Privileged status of one over the other might give us some important clues as to how “edges” to use a concept from graph theory might be constructed in the infant lexicon.
Notice that I’m also making a simplifying assumption that taxonomic relations equate with semantic relations and that thematic relations equate with associative relations. This is clearly an oversimplification, but it will get us off the ground.

So we’ve used a similar design to the experiment before but eliminated the uninformative unrelated prime – unnamed target condition and instead of using related primes that are both semantically and associatively related we’ve only used related primes that are taxonomically OR thematically related. In the absence of this priming boost, we’re using slightly older age groups. Otherwise, we tested two groups of infants at each age with one group exposed to taxonomic primes and the other group to thematic primes.

The pattern of results were identical in both the taxonomically and associatively related priming conditions: 24 month olds showed a clear priming effect – they looked longer at the target in related prime compared to the unrelated prime condition, and the prime only condition was ineffective in eliciting target looking. This result indicates that the representation of some taxonomic AND thematic lexical links are in place by 24 months.

However, the 21 month olds behaved like the 18 month olds in the previous experiment: there was no priming effect but they showed similar target preferences in the related and unrelated primed conditions. The words functioned as if they were lexical islands.
Of course, we know from the previous experiment that 21 month olds do NOT behave as though their words are lexical islands under a priming boost scenario. This contrast suggests that even with infants priming effects are dynamic and subject to task constraints.


  • Overall, the results from this series of experiments indicates that the rudiments of a lexico-semantic system are in place by 24 months of age and perhaps even earlier if adequate testing conditions are used. These young toddlers seem to be able to exploit taxonomic and thematic relations btw the mental representation of words to build a structured lexicon.
  • The manifestation of this lexical structure is demonstrated by interference effects in target recognition for unrelated referring expressions.
  • These interference effects do not seem to be present in the lexicons of 18 month olds (or 21 month olds in the absence of a priming boost) where the mental representation of words function as lexical islands.
So far, I’ve focused on the development of what psychologists broadly call the semantic structure of the lexicon. What about infants’ appreciation of the relationship between the sounds of words in their lexicon. Of course, we know that they’re sensitive to the sound contrast of their mother tongue and that they can detect mispronunciations of familiar lexical items almost as young as we can test them. However, this does not tell us about their sensitivity to the phonological similarity between words.

To investigate this question, we have used another adaptation of the inter-modal preferential looking task which might call Implicit Phonological Priming. In this task infants are again shown a pair of familiar objects on a computer screen and one of them is named. However, immediately prior to this, they are briefly shown a single picture of a familiar object whose name is known to them but which they observe in silence.

The critical manipulation in the experiment concerns the relationship of the label of the unnamed prime picture to the word that is heard by the infant during presentation of the target/distracter picture pair. The stimuli are chosen so that the unheard label has the same onset consonant as the heard label or there is no phonological overlap.

In this experiment, we tested 18 month olds with equal numbers of phonologically related and unrelated picture prime trials. Our question was to determine whether a phonologically related picture prime produced a systematically different pattern of target recognition than a phonologically unrelated picture prime.

The results were quite straightforward. 18 month olds demonstrated clearcut target looking when they saw a phonologically related picture prime but their looking behaviour was entirely unsystematic when they saw an phonologically unrelated picture prime. This finding is supported both by an overall target preference in the related prime condition compared to the unrelated condition and also by moment to moment preference during the course of the trial.

It’s worth keeping in mind that infants at this age would demonstrate robust target looking in the absence of any picture prime, suggesting that phonologically unrelated picture primes interfered with target recognition.

It’s clear from this experiment that the picture prime influences target looking, but it’s influence is through label associated with that picture, not the picture itself. This provides conclusive evidence that these 18 month old infants are implicitly naming the familiar object. This finding is interesting in itself because it tells us that linguistic representations are being automatically generated by these young participants when they see a familiar category token.
For students of infant categorisation, the tendency to tag an object with a linguistic symbol may serve to highlight its category membership and facilitate communication.

For the purposes of this talk, it demonstrates that infants not only implicity name the object but that representation of the name impacts the processing of subsequent auditory stimuli – presumably through phonological overlap.

Cohort Effects

Now suppose we repeat the exact same experiment with 24 month olds – same stimuli, same timing parameters and same contrasting conditions. As you can see in the bar chart, their performance contrasts dramatically with the 18 months olds. Not only do they show systematic target recognition in both the related AND unrelated conditions but they show more target looking in the UNRELATED than the RELATED condition.

This result implies that our 24 month old toddlers also implicitly name the picture prime but it complicates our assumptions about the impact of the implicit name on the auditory label and subsequent target recognition.

It turns out that the range of words used in these experiments began with one of 5 consonants /b/, /k/, /d/, /p/ and /t/. If we count up the number of words that our 24 month olds knew in each of these cohorts a simple split emerged. /b/ words with large cohorts and the rest with smaller cohorts.

If we now reanalyse the data according to the size of the TARGET cohort, a rational pattern of responding emerges: Trials with LARGE cohorts produce less target recognition than trials with SMALL cohorts. This is a result entirely familiar to those working with adult lexical recognition. Furthermore, the priming condition attenuates target recognition in such a way that the related prime acts to support the cohort effect by further suppressing recognition. Note that none of these cohort effects are found with the 18 month olds even though there are still substantial imbalances in the cohort sizes of the set of words known to these infants.

Now if the target’s cohort is having an impact on target recognition by way of enhanced lexical competition, then should be apparent with or without a picture prime. We therefore repeated the experiment but with a third unprimed condition.

Again the results were clearcut. The picture priming conditions replicated the pattern of results in the previous experiment with large target cohorts suppressing target looking compared to small cohorts and related primes exagerating this cohort effect. Most importantly, of course, target cohort effects were still apparent in the absence of picture priming. Large target cohorts resulted in less target recognition than small cohorts.

These results provide support for the view that 24 month olds have a lexical system that embodies lexical competition, presumably via inhibitory processes. There is little evidence for these competitive processes at 18 months of age. Of course, it won’t have escaped your attention that I made a similar argument wrt the lexico-semantic system just 10 minutes ago!
We used TRACE initially to model a range of mispronunciation studies in the infant word recognition literature.
Of course, TRACE has no semantics so our simulations were based on the simplifying assumption that manipulation of word form information is adequate to capture the experimental results. Nor is TRACE a developmental model, so we simulated development by systematically manipulating the size of TRACE’s lexicon to mimick vocabulary development in the second year of life. This yielded a set of simulations that successfully mimicked vowel and consonant mispronunications of familiar words during the second year of life, and also generated some interesting predictions regarding the changing sensitivity to these segment during development.
I don’t have time to describe these results here – but I do want to show you how we simulated, with the same model, a different set of results that demonstrated graded sensitivity to the degree of mispronunciation of familiar words by young infants.
Using an adaptation of the intermodal preferential looking task,
White & Morgan found that 19 month old infants show a graded sensitivity to the degree of mispronunciation of familiar words, as measured in terms of number of features: With a larger number of featural changes infants showed less target recognition than correct or a small number of featural changes. We can interpret these results as evidence for the psychological reality of these featural segments at 19 months of age – though the authors acknowledge that other acoustic metrics of mispronunciation might do just as good or even a better job.
Our concern here was to model these findings in TRACE. At first blush this might seem a trivial matter. However, it turned out that we could only this if achieve this if we switched off one of TRACE’s central mechanisms – lexical competition. This is because the cohort sizes in White & Morgan’s different mispronunciation conditions forced TRACE to show a non-linear pattern of responding. As you can see, 3 feature mispronunciations have the lowest cohort sizes whereas 2 feature mispronunciations have the highest.
The corollary of this theoretical finding is that lexical competition must be absent or substantially diminished at 19