Can infants retain statistically-segmented words and mappings across a delay?

Abstract

Infants are sensitive to statistics in spoken language that aid word-form segmentation and immediate mapping to referents. However, it is not clear whether this sensitivity influences the formation and retention of word-referent mappings across a delay, two real-world challenges that learners must overcome. We tested how the timing of referent training, relative to familiarization with transitional probabilities (TPs) in speech, impacts English-learning 23-month-olds’ ability to form and retain word-referent mappings. In Experiment 1, we tested infants’ ability to retain TP information across a 10-minute delay and use it in the service of word learning. Infants successfully mapped high-TP but not low-TP words to referents. In Experiment 2, infants readily mapped the same words even when they were unfamiliar. In Experiment 3, high- and low-TP word-referent mappings were trained immediately after familiarization, and infants readily remembered these associations 10 minutes later. In sum, although 23-month-old infants do not need strong statistics to map word forms to referents immediately, or to remember those mappings across a delay, infants are nevertheless sensitive to these statistics in the speech stream, and they influence mapping after a delay. These findings suggest that, by 23 months of age, sensitivity to statistics in speech may impact infants’ language development by leading word forms with low coherence to be poorly mapped following even a short period of consolidation.

1. Introduction

The discovery that human infants readily identify patterns among co-occurring elements in auditory and visual streams has profoundly reshaped our estimation of infants’ cognitive capacities by providing evidence that infants are well-equipped to learn (for reviews see Aslin, 2017; Cannistraci et al., 2019; Saffran & Kirkham, 2018). This ability, called statistical learning, was originally demonstrated in the domain of word segmentation by Saffran, Aslin, and Newport (1996; see also Aslin, Saffran & Newport, 1998). Using a simple artificial language, stripped of prosodic cues, they showed that 8-month-old infants are sensitive to syllable co-occurrence patterns (i.e., transitional probability, or TP; the probability of Y given X in the sequence XY). Following familiarization with one of these languages, infants discriminate between “words”, or syllable sequences with high TP (i.e., TP=1.0), and partwords, or sequences with low TP (i.e., TP=.33), even when high and low TP sequences were equally frequent (Aslin, Saffran, & Newport, 1998). Infants of the same age are also able to track TPs in complex natural language stimuli (Pelucchi, Hay, & Saffran, 2009a, 2009b). Thus, tracking statistical distributions may be one mechanism by which infants distinguish potential words from syllable sequences that span word boundaries (Saffran et al., 1996).

Not only does tracking statistical distributions help infants to find word forms in fluent speech, but high-TP sequences are also learned better as object labels than their equally frequent low-TP counterparts at 17 months of age (Graf Estes, Evans, Alibali & Saffran, 2007; Hay, Pelucchi, Graf Estes, & Saffran, 2011). Hay and colleagues (2011) showed this by presenting 17-month-old monolingual English-learning infants with naturally spoken Italian sentences that contained four target word forms (Hay et al., 2011; Experiment 3). Two-word forms were high TP (HTP; TP=1) because their syllables never appeared elsewhere in the corpus. The two other word forms were low TP (LTP; TP=.33) because both their first and second syllables appeared in other words throughout the corpus. Immediately following familiarization, infants were trained on mappings between novel objects and either HTP or LTP words. Infants readily learned mappings between HTP words and objects, but they failed to learn LTP mappings (see Graf Estes et al., 2007 for a similar finding using artificial language materials). These results suggest that statistical learning may contribute to infants’ ability to solve a fundamental problem in language learning – identifying word forms within continuous speech – and that the outcome of this process may play a role in solving yet another challenging task – mapping word forms to referents.

However, the availability of potentially informative statistical regularities in speech does not mean that learners are always influenced by those regularities when learning word meanings, or that these regularities always exert the same influence on learners. In fact, the role of TPs in mapping word forms to referents appears to change across the lifespan. For example, although adults distinguish between HTP and LTP sequences from an artificial language (Saffran, Newport & Aslin, 1996), they are ultimately able to learn both as labels (Mirman et al., 2008). Further, corpus studies indicate that within-word TPs in natural languages tend to be far lower than 1 (Saksida, Langus, & Nespor, 2017; Swingley, 1999). These findings suggest that learners are capable of incorporating words with lower TPs into their vocabularies. Interestingly the role of TPs in word-referent mapping appears to change by the time infants are approaching age two. For instance, immediately following familiarization with an Italian corpus, 21- to 23-month-old English-learning infants map HTP and LTP word forms to referents to a similar degree, despite evidence that they tracked the statistics (Shoaib et al., 2018; Lany et al., 2024). Taken together, these results suggest that for 17-month-olds, who are less experienced word learners, TPs can make a critical difference in whether a word form is mapped (Hay et al., 2011). However, for older infants who have more experience learning words, labels with relatively strong and weak statistics can be mapped equivalently, at least when there is no delay between exposure to the corpus and training on the word-referent mappings.

Even though HTP and LTP word forms are mapped equally well when there is no delay at 23 months (Lany et al., 2024), it would be premature to conclude that there is no difference in how well these infants represent these word forms or their mappings. In particular, in all of the studies we have reviewed, infants are trained and tested on word-referent mappings immediately after familiarization, and thus they may not provide an especially sensitive test of how robust these representations are. It is possible that HTP word forms are more robustly encoded and represented than LTP word forms, due to their relatively strong statistical coherence. Because strength of encoding is related to retention (Craik & Tulving, 1975), HTP word forms could be retained better across a delay. If so, they might be more readily mapped to referents at that point. Likewise, HTP word-referent mappings may be better remembered than LTP mappings across delays. Here, we tested whether HTP word forms are mapped to referents better than LTP words following a 10-minute delay, and whether memory for HTP and LTP mappings withstand delays equally well. These questions are highly relevant, as infants are often exposed to statistical distributions relevant for segmenting speech in the absence of immediate exposure to visual referents that they map onto, and have to retain word-referent mappings once formed.

Although it is well-established that infants can remember familiar words across short and long delays (e.g., Jusczyk and Hohne, 1997; Houston and Jusczyk, 2003), these studies do not tell us whether some words are more easily remembered than others, and specifically whether TPs impact how well infants remember word forms. A couple of studies have tested whether infants show differential sensitivity to HTP and LTP sequences when there is a delay between initial experience with them and the test. These studies suggest that although infants can distinguish between HTP and LTP words when tested immediately, 8- to 9-month-olds’ sensitivity to statistical distributions appears to be fairly time-limited (Karaman and Hay, 2018; Simon et al., 2017). Critically, little is known about how well older, more proficient word learners retain statistical distributions, nor how delays impact subsequent word-referent mappings. Likewise, although 2-year-olds can retain word-referent mappings for days (Woodward et al., 1994; Wojcik, 2017), the extent to which the statistical properties of labels can impact memory for such mappings has not been tested. Even though both HTP and LTP words can be mapped immediately at this age (Lany et al., 2024), once learned, HTP mappings may be better remembered than LTP mappings. Despite showing similar levels of performance on HTP and LTP mappings immediately after being trained on them, the strong statistics of HTP word forms may nevertheless support a more robust mapping. In this scenario, this more robust mapping does not appear to help infants to immediately find the referent when it is paired with a dissimilar distractor, but rather may allow the mapping to persist over a longer time period.

We tested these questions across 3 experiments. In Experiment 1, we tested whether infants map HTP and LTP word forms equivalently when there is a 10-minute delay between familiarization with the TPs in speech and training on the word-referent mappings. Specifically, we exposed 23-month-old English-learning infants to a naturally-produced corpus of Italian sentences that contained HTP and LTP word forms. Following a 10-minute delay, infants were trained and tested on HTP and LTP word-referent mappings. Infants only showed evidence of mapping the HTP word forms, not the LTP word forms. In Experiment 2, we probed whether results from Experiment 1 suggest that LTP word forms were retained more poorly than HTP word forms (i.e., whether LTP word forms seemed relatively unfamiliar after a delay). Here, 23-month-olds were trained and tested on mappings between unfamiliarized Italian words and referents. These unfamiliar word forms were mapped as well as the HTP word forms, and better than the LTP word forms in Experiment 1, suggesting that infants had not simply remembered the HTP word forms and forgotten the LTP word forms across the delay. In Experiment 3, we tested whether infants retain HTP and LTP word-referent mappings equivalently across a delay. To explore this question, HTP and LTP word forms were immediately trained as labels after familiarization with the Italian corpus, and infants were tested on the mappings 10 minutes later. Infants showed evidence of remembering both mappings. Altogether these experiments suggest that 2-year-olds can learn HTP and LTP word forms as labels immediately and retain them both across a delay. They also successfully map HTP word forms to referents after a delay, but fail to do so for LTP word forms, although this difference does not reflect straightforward forgetting of the LTP word forms. These results suggest that statistical structure in speech influences word learning in 23-month-olds, as it does for younger infants, although the underlying mechanism may be different. We consider this possibility in the General Discussion.

2. Experiment 1

At 23 months, infants map HTP and LTP word forms equally well when trained and tested immediately after they encounter them in fluent speech (Lany et al., 2024). In Experiment 1 we tested how a 10-minute delay after hearing the HTP and LTP word forms in speech impacts mapping them to referents.

2.1. Methods

The experiment consisted of three phases: Familiarization, Referent Training, and Test (see Figure 1). To give a brief overview, infants first listened to a corpus of Italian sentences containing HTP and LTP words (Familiarization). After a 10-minute delay those words were trained as labels (Referent Training) and tested (Test) using the Looking-While-Listening (LWL) paradigm (Fernald et al., 2008). The specific stimuli and procedures used in each phase are described in detail in the following sections.

Figure 1.

Overview of experimental procedures for Experiments 1, 2, and 3.

2.1.1. Participants

Our final sample for Experiment 1 consisted of thirty-two 22- to 24-month-old monolingual infants (Mage = 23.16 months, range = 22.31 – 24.0, 15 females, 17 males) whose primary language was American English. All infants had a gestational age of greater than 36 weeks and no hearing or vision problems, according to parental reports. Infants were recruited through a participant database maintained in the Department of Psychology at the University of Tennessee, Knoxville. The target sample size was calculated using effect sizes reported in previous work investigating infants’ ability to map HTP sequences to novel objects (Hay et al., 2011; d = .64) and an 85% power threshold. This yielded a minimum recommended sample size of 24 infants. We increased the sample size to 32 participants because our task was more demanding. Specifically, infants were taught four novel label-object mappings instead of just two as in Hay et al. (2011). Data from 19 additional infants were not included in the analysis due to fussiness (5), failure to provide usable data on at least half of both HTP and LTP trials (7), not paying attention (3), parental interference (2), and experimental error (2). The demographics of the sample were 30 White, 1 Black/African-American, and 1 Native Hawaiian or other Pacific Islander.

2.1.2. Stimuli

In the Familiarization Phase, infants listened to a corpus of 12 grammatically correct and semantically meaningful Italian sentences that were produced by a native female speaker. The speaker was instructed to read the stimuli with a lively voice, pretending to be in front of a baby, in a manner consistent with Infant-Directed Speech (IDS). These were re-recordings of Languages 2A and 2B from Hay et al., 2011 (see the Supplementary Document for the full set of familiarization materials). Four disyllabic target words, casa, bici, fuga, and melo, were embedded in the sentences. Each of the target words appeared equally often in the corpus, but the internal TPs of these words differed. The two HTP words (TP=1.0) contained syllables that never appeared anywhere else in the corpus. Both the first and the second syllables of the two LTP words (TP = .33) appeared 12 additional times throughout the corpus. Each target word appeared 6 times across the 12 sentences and each sentence was repeated 3 times during familiarization, which lasted 2 min 30 s. Thus, infants heard each target word 18 times during Familiarization.

To avoid bias in the results due to arbitrary preferences for particular target words based on specific phonemes, or phonotactic regularities (including the potential influence of English TPs), we used two counterbalanced languages. Specifically, words that were HTP in Language A were LTP in Language B, and vice versa. Thus, the HTP words were casa and bici in Language A, and the LTP words were fuga and melo. In Language B the LTP words were casa and bici, and the HTP words were fuga and melo. Although the Italian corpus likely sounded non-native to our English-learning participants (Mehler et al., 1988), it shared some key similarities with English: all of the Italian target words were trochees (i.e., strong-weak stress, a pattern that is common in English), and all of the target words were phonotactically legal in English. These features made it likely that infants would be able to track the syllable-level TPs, and previous studies using these languages suggest that 8- to 23-month-olds readily do so (e.g., Hay et al., 2011; Pelucchi et al., 2009a; Shoaib et al., 2018; Lany et al., 2024).

During the Referent Training and Test Phases, depicted in Figure 1, HTP and LTP target words were presented in isolation and also embedded in English carrier phrases (e.g., Referent Training → “[Target] See the [target]! It’s a [target]! [Target]!”; Test → “Find the [target]! [Target]! Do you like it?”, or “Where’s the [target]? [Target]! Do you see it?”). Short pauses were introduced between the sentences to allow time to process each sentence. In addition to the HTP and LTP target-word trials, the Referent Training Phase also included four trials containing familiar English words (baby, doggie, shoe, and book) presented along with images of the corresponding objects. This was done to orient the infants to the format of the label-object mapping task. These familiar words are commonly used in studies testing infants’ comprehension of early-learned English words, and are highly likely to be known at 23 months (Fernald et al., 2008). Both target words and familiar words were matched in intensity (~65 dBSPL) and approximate length (range 750 ms to 850 ms). These phrases were produced in a lively manner in accented English by the same native Italian speaker who produced the familiarization corpus materials.

To maintain interest throughout the Familiarization Phase, infants watched a silent cartoon video (Winnie-the Pooh) while listening to the Italian corpus. On HTP and LTP target-word trials, we used colorful images of novel objects, matched in size and brightness, as referents (see Figure 1). Four familiar object images (baby, doggie, shoe, and book) were used on familiar-word trials. In all Referent Training trials, the image of a single object moved slowly across a white box that appeared on either the bottom right or bottom left corner of the display screen (object visual angle ~10°). The movement of the object was not synchronized with the timing of object labeling. On test trials, two stationary objects were presented, each in a white box, in the bottom right and left corners of the display screen (visual angle between objects ~30°). Objects were yoked based on test trial type. For example, in HTP test trials, the two objects that appeared together on the display screen had both previously been paired with HTP words. Familiar objects were yoked based on animacy (shoe-book, baby-doggie). All stimuli, data, and analysis scripts are available through OSF https://osf.io/pfymu/?view_only=228c0d84fc1f4ce592ed5aabaf51ad32.

2.1.3. Procedure

Infants were seated on their caregiver’s lap approximately 1 m from a 42-inch flat-screen television, which was used to present the visual stimuli. Caregivers listened to masking music via headphones throughout the experiment. Infants were first familiarized with one of the Italian languages, counterbalanced across participants. After the Familiarization Phase, infants were given a 10-minute break. We chose a 10-minute break because it is a retention interval that is commonly experienced by infants in their everyday lives when hearing a word form in the absence of a referent and then reencountering it with a referent (Slone et al., 2023). It is also a sufficiently long delay to ensure that any influence of previously encountered word forms is based on retrieval from long-term memory (Karaman & Hay, 2018). Additionally, a 10-min delay was sufficiently brief to ensure that infants did not become tired or bored during the lab visit. During the delay, a research assistant played quietly with the infant in the waiting area outside of the testing booth. After the delay, infants were trained and tested on word-referent mappings.

During the Referent Training Phase, infants were presented with four novel word-referent pairs (i.e., 2 HTP and 2 LTP labels with their referents) and four familiar word-referent pairs. On each Referent Training trial, a single moving object was simultaneously paired with speech containing its label. On a given trial, the target word was presented 4 times, twice in the context of a carrier phrase, and twice in isolation (e.g. “Melo! See the melo! It’s a melo! Melo!”). The first presentation of the target word occurred one second after the onset of the trial. However, due to natural variations in the duration of the target words and the phases used, the remaining tokens occurred at slightly different absolute times over the course of the Referent Training trial. Each referent training trial lasted 10 seconds. There were four trials for each of the HTP and LTP word-referent pairs, for a total of 16 Referent Training trials. Each of the four familiar word-referent pairs was also presented once, for a total of 20 Referent Training trials. The Referent Training Phase began with two familiar word-referent trials. The remaining 18 training trials were presented in one of four quasi-random orders.

Infants’ ability to map the HTP and LTP words to referents was tested using the Looking-While-Listening (LWL) procedure (Fernald, Zangl, Portillo, & Marchman, 2008). On each test trial, two stationary objects appeared on the screen for 500 ms before the onset of the English carrier phrase, in which either a familiar or a novel target word was embedded. This target word occurred 2 seconds after trial onset. An additional repetition of the isolated target word was presented 1.5 seconds after the onset of the first one. Finally, 500 ms after the second repetition of the target word, infants heard another English phrase (e.g. “Do you like it?” or “Do you see it?”), and then the trial ended (see Figure 2 for a timeline). Each test trial lasted 8 seconds. The test phase began with 2 familiar word-referent trials to help infants get accustomed to the structure of the LWL procedure (Fernald et al., 2008). After the first two trials, trial type was counterbalanced in four quasi-random testing orders for a total of 32 test trials (8 familiar, 12 HTP, and 12 LTP word trials). The target object appeared on both the left and right sides of the display screen an equal number of times. No labels occurred twice in succession. The entire experiment lasted about 20 minutes, including the 10-minute delay.

Figure 2.

Timeline of a test trial.

2.1.4. Data Coding & Analysis

Infants’ eye gaze was video-recorded at a rate of 30 frames per second. In a first pass, a trained coder pre-screened all of the test trials by watching the video back at regular speed. We excluded individual trials from the analysis if the infant or caregiver was talking, if the infants’ eyes were not visible, or if the infant was fussy or appeared to be off-task (e.g., looking at their shoes, or got off the parents’ lap) for more than half of the trial. These trials were not coded or included in the analysis. The remainder of the trials were coded frame-by-frame using iCoder software (Fernald et al., 2008). On each frame, the coder indicated whether the infant was looking to the object on the left, to the object on the right, shifting between objects, or whether the infant was looking away/off task (see Fernald et al., 2008). A second coder re-coded a random selection of 25% of the infants. We obtained two reliability scores: a frame agreement score which is based on the percentage of frames on which two coders’ judgments agree overall and a shift agreement score which focuses only on sequences of frames where the shifts occurred (Fernald et al., 2008). Frame and shift agreements between coders were greater than 98%.

Infants’ word learning was assessed using an accuracy measure reflecting the proportion of time spent looking at the target object in a window beginning just after the first label was presented. Specifically, mean accuracy for each HTP and LTP trial was calculated using an analysis window beginning at 300 ms after the onset of the first target label and extending to 2000 ms post-label onset, consistent with the extant LWL literature (Fernald, Perfors, & Marchman, 2006; Fernald et al., 2008; Swingley & Aslin, 2000; Swingley, Pinto, & Aslin, 1999). Note that the analysis window included the onset of the second presentation of the target label (see Figure 2). The accuracy score reflects the proportion of time spent looking at the target object following label onset divided by the total time spent looking at the target and distractor objects. Individual trials were automatically excluded from further analysis if the infant looked somewhere other than to the target or distractor object for more than 15 consecutive frames (i.e., 500 ms) during the analysis window (i.e., insufficient looking). Seven infants who failed to provide enough usable trials (i.e., at least half, or at least 6 HTP and 6 LTP trials) were excluded from our final analysis. For the remaining 32 infants in our sample, 18% of trials (139 out of 768 total possible HTP and LTP test trials) were excluded from the analysis based on prescreening or insufficient looking. A total of 629 trials (HTP trials: 313, LTP trials = 316) were included in the analyses, with each participant contributing an average of 20 trials.

2.2. Results and Discussion

To examine infants’ mapping of HTP and LTP word forms to referents following a delay, we performed linear mixed-effects modeling through the lme4 package in R (Bates, Maächler, Bolker, & Walker, 2015; R Development Core Team, 2019). The model was fitted with a maximal random effects structure, which included a by-subject random intercept and slope for Label Type (i.e., HTP and LTP), as well as a by-item random intercept (Barr, Levy, Scheepers, & Tily, 2013). After pruning the random effect structure to address singularity issues, the final model retained a by-subject random intercept and a by-item random intercept. Note that the maximal model, which included a by-subject random slope for Label Type, produced equivalent results to the final model with only by-subject and by-item random intercepts. In the analysis, dummy coding was used to represent the fixed effects of Label Type and Counterbalanced Language. The reference level for Label Type was LTP, while the reference level for Language was B. All significance tests used a two-tailed alpha set to .05.

Results revealed no significant main effect of Counterbalanced Language on overall accuracy, β = .024, SE = .034, p = .482, 95% CI = [−.043 – .090]. However, there was a significant effect of Label Type, β = −.087, SE = .025, p < .001, 95% CI = [−.137 – −.037], with better performance on HTP than LTP trials (see Table 1). Interestingly, planned one-sample t-tests comparing performance to chance revealed that although infants successfully mapped the HTP words to novel objects (M = 60%, SD = 10%), t(31) = 5.74, p < .001, 95% CI = [.065 – .138], d=1.01, they failed to map the LTP words (M = 51%, SD = 13%), t(31) = .25, p = .80, 95% CI = [−.041 – .053], d = .04 (see Figure 3). As is traditional in the literature (e.g., Fernald et al., 2008), and in order to visually inspect looking patterns across time, we plotted infants’ eye gaze data as a proportion of looking to the target across our analysis window (see Figure 4). The figure suggests that infants may have fixated on the LTP referent starting at the end of our analysis window, between 1400 and 2300 ms. However, exploratory analysis revealed that even in this later window, looking did not significantly differ from chance on LTP trials (M = 54%, SD = 15%), t(31) = 1.62, p = .11, 95% CI = [−.011 – .099], d = .29. Further, infants performed better on HTP than LTP trials in this later window, β = −.092, SE = .021, p < .001, 95% CI = [−.133 – −.050], just as they did in the original analysis window.

Table 1.

Model parameters and results

		β	SE	CI	p
Experiment 1
	(Intercept)	.587	.031	.525 – .648	<.001
	Language [B]	.024	.034	−.043 – .090	.482
	Label Type [LTP]	−.087	.025	−.137 – −.037	< .001
Experiment 2
	(Intercept)	.511	.022	.468 – .553	<.001
	Unfamiliar Word	.057	.028	.001 – .112	.044
Experiment 3
	(Intercept)	.572	.029	.516 – .629	<.001
	Language [B]	.032	.027	−.021 – .084	.234
	Label Type [LTP]	−.010	.026	−.062 – .042	.695

Note. Language was not included as a fixed effect variable in Experiment 2, given that only a single familiarization language was used.

Figure 3.

Bar plots display the mean proportion of time spent looking at target objects during HTP, LTP, and Unfamiliar word test trials. Performance was assessed across the analysis window of 300–2000 ms for Experiments 1, 2, and 3. Error bars represent ± 1 SE.* p < .01.

Figure 4.

Mean proportion of time looking to target objects on HTP and LTP word test trials as a function of time in Experiment 1. Dashed lines represented the analysis window. The ribbon around the lines indicates ± 1 SE.

In sum, when there was a 10-minute delay between Familiarization and Referent Training, infants learned HTP but not LTP word forms as labels, and performance on these trial types was significantly different. These results clearly suggest that 23-month-old infants track TPs in the speech stream during Familiarization. Further, these results provide evidence that experience with a label’s TP influences how well that label is mapped to a referent after a delay. This supports the idea that, at this age, although the statistical coherence of word forms does not measurably influence how well word forms are mapped to referents when there is no need to retain these word forms over time (Lany et al., 2024), it does impact how well word forms are learned as labels when they are re-encountered after a 10-minute delay.

There are two potential explanations for our results. Given that infants learn both HTP and LTP word forms as labels immediately (Lany et al., 2024), one explanation is that at 23 months, infants remember HTP word forms across the delay but forget LTP ones, and therefore only learn mappings involving HTP word forms. Note that this explanation assumes that previous experience with the word forms in the speech stream supports – and in fact, is critical to – learning both HTP and LTP word forms as labels in immediate mapping tasks at this age. In this scenario, delays between experience with TPs in fluent speech and word learning opportunities selectively confer a mapping advantage for word forms with strong statistics, which could be more salient and thus better retained than those with weaker statistics. If HTP and LTP word forms were mapped differently following a delay because HTP word forms were retained and LTP word forms were forgotten, infants should also fail to map unfamiliar word forms.

A second possibility is that differential retention of HTP versus LTP word forms is not what leads to a mapping advantage for HTP words. Instead, infants may become resistant to learning LTP word forms as labels across the delay. In this scenario, the high internal coherence of HTP word forms may not in fact lead them to be privileged relative to unfamiliar word forms at this age, but rather the lower internal coherence of the LTP words may lead to a mapping disadvantage across the delay. Experiment 2 is designed to differentiate these alternative possibilities by testing whether infants readily map unfamiliar Italian word forms to referents at this age.

3. Experiment 2

If infants’ poor mapping of LTP word forms in Experiment 1 results from forgetting them across the 10-minute delay, then infants should learn them equivalently to unfamiliar word forms. Likewise, if infants’ relatively strong mapping of HTP word forms is because they are better remembered than LTP word forms (i.e., if LTP word forms are perceived as unfamiliar), then they should learn HTP word forms better than unfamiliar ones. In contrast, if infants’ poor mapping of LTP word forms is due to resistance, then they should learn them worse than unfamiliar word forms. Thus, in Experiment 2, we tested how well 23-month-olds learn unfamiliar Italian words when they have no prior experience with them, such that their TPs are zero. By comparing their performance on these trials to their performance on HTP and LTP trials from Experiment 1, we can determine whether HTP word forms are mapped better, and LTP word forms more poorly, than unfamiliar words. As in Experiment 1, infants were exposed to a corpus of Italian sentences during a Familiarization Phase. Critically, the words used as labels during the subsequent Referent Training Phase, and their component syllables, were not present in the corpus. Then, with no delay, infants were trained and tested on word-referent mappings using the Referent Training procedure from Experiment 1. Importantly, the same labels (casa, bici, fuga, melo) were used in both Experiments 1 and 2. For infants in Experiment 2 the labels were unfamiliar, while for infants in Experiment 1 the labels had been encountered 10 minutes beforehand.

3.1. Methods

3.1.1. Participants

Our final sample consisted of thirty-two 22- to 24-month-old infants (Mage = 22.95 months, range = 22.05 – 23.85, 16 females, 16 males). All other participant characteristics and recruitment methods were identical to Experiment 1. Data from 20 additional infants were not included in the analysis due to fussiness (12), failure to provide usable data on at least half of the trials (2), not paying attention (3), parental interference (1), and experimental error (2). The sample consisted of 30 White infants, 1 Black/African-American infant, and 1 Asian infant.

3.1.2. Stimuli

Familiarization stimuli were a corpus of Italian sentences that were similar to those used in Experiment 1, but the target words (casa, bici, fuga, melo) were not present in it (i.e., they were Unfamiliar). The target words and visual stimuli used in the Referent Training and Test Phases were identical to those used in Experiment 1.

3.1.3. Procedures

Procedures were identical to those in Experiment 1 except that infants were trained and tested on word-referent mappings immediately after familiarization with the Italian corpus.

3.1.4. Data Coding & Analysis

The data were pre-screened and coded as in Experiment 1, and the same exclusion criteria were applied. As in Experiment 1, the accuracy score reflects the proportion of time spent looking at the target object following label onset divided by the total looking time at the target and distractor objects. The same analysis window was used as in Experiment 1. Once again, data from 25% of the infants was re-coded by a second coder. Frame and shift agreements between coders were 99% and 96% respectively. Approximately 18% (141 out of 768 total possible HTP and LTP test trials) were excluded from the analyses based on prescreening and insufficient looking.

3.2. Results and Discussion

Infants successfully learned the Unfamiliar labels, as accuracy scores were significantly greater than chance (M = 57%, SD = 10%), t(31) = 3.75, p < .001, 95% CI = [.030 – .101] d= .66 (see Figures 1 and 5). The analysis using a linear mixed-effects model revealed that infants performed significantly worse on the LTP trials from Experiment 1 than on the Unfamiliar trials tested in Experiment 2, β = .057, SE = .028, p = .044, 95% CI = [.001 – .112]. However, there was no significant difference between performance on Unfamiliar trials and performance on HTP trials from Experiment 1, β = −.033, SE = .026, p = .208, 95% CI = [−.084 – .018].

Figure 5.

Mean proportion of time looking to target objects on Unfamiliar word trials as a function of time in Experiment 2. Dashed lines represented the analysis window. The ribbon around the lines indicates ± 1 SE.

These data suggest that at 23 months, infants can map word forms to referents whether or not the word forms have been encountered previously. Moreover, at the point of mapping, infants in Experiment 1 did not appear to benefit from the high coherence of the HTP word forms that had been heard 10 minutes previously, as performance on HTP and Unfamiliar word forms was comparable. Critically, infants mapped Unfamiliar word forms better than LTP word forms. These results suggest that infants did not simply forget the LTP word forms across the delay. If they had forgotten the LTP word forms, then they should have learned them equivalently to the Unfamiliar words. In the General Discussion, we consider potential explanations for why infants more readily mapped Unfamiliar words than LTP words.

4. Experiment 3

Just as infants need to retain sensitivity to statistical distributions for them to impact word learning (either by supporting or interfering with forming word-referent mappings), learning the meanings of words ultimately requires that infants retain mappings over time. By 2.5 years of age, infants can retain mappings between novel word forms and referents in lab-based tasks, similar to this one, for up to a week (Wojcik, 2017). However, newly formed mappings can also be forgotten quickly (Horst & Samuelson, 2008), depending on task parameters. Thus, in Experiment 3, we tested whether infants are able to retain HTP and LTP word-referent mappings across a 10-minute delay. Although 23-month-old infants learn HTP and LTP word forms as labels when trained and then tested immediately following familiarization with a corpus (Lany et al., 2024), Experiment 1 shows that infants differentially map HTP and LTP word forms to referents when there is a delay between encountering the forms and Referent Training. It is possible that the robustness of the HTP and LTP word-referent mappings may also vary as a function of the statistical coherence of the labels, even if we cannot observe these differences when infants are tested immediately. In this case, infants might retain HTP mappings better than LTP ones across a delay. Alternatively, experience mapping HTP and LTP words to referent immediately after familiarization may lead infants to form robust representations of both types of mappings, thus leading infants to perform equivalently on them even when tested after a delay. In Experiment 3, we tested these alternatives by modifying the procedure used in Experiment 1, such that we introduced a 10-minute delay between Referent Training and Test, rather than between Familiarization and Referent Training. Thus, in Experiment 3, infants were familiarized with the same corpus of Italian sentences used in Experiment 1 and were then trained immediately on mappings between the HTP and LTP word forms and novel objects. Infants’ memory for the word-referent associations was then tested after a 10-minute delay. Thus, unlike in Experiment 1, in which we tested the impact of the internal coherence of the word forms on infants’ ability to form word-referent mappings after a delay, in Experiment 3 we tested how the internal coherence of the word forms impacts infants’ memories for already formed word-referent associations.

4.1. Methods

4.1.1. Participants

Our final sample consisted of thirty-two 22- to 24-month-old English-learning monolingual infants (Mage = 23.03 months, range = 22.01 – 23.98, 16 females, 16 males, all White). All other participant characteristics and recruitment methods were identical to Experiments 1 and 2. Data from 12 additional infants were not included in the analysis due to: fussiness (5), failure to provide usable data on at least half of the HTP and LTP trials (5), not paying attention (1), and experimental error (1).

4.1.2. Stimuli

Familiarization stimuli and visual stimuli were identical to those used in Experiment 1.

4.1.3. Procedure

Procedures were identical to those in Experiment 1, except that infants were trained on word-referent mappings immediately following familiarization and were then tested on those mappings after a 10-minute delay. During the delay, a research assistant played quietly with the infants in the waiting area outside of the testing booth.

4.1.4. Data Coding & Analysis

The data were pre-screened and coded as in Experiments 1 and 2, and the same exclusion criteria were applied. The same analysis window was used as in Experiments 1 and 2. Data from 25% of the infants was re-coded by a second coder. As in Experiment 1, frame and shift agreements between coders were greater than 98%. Five infants who failed to provide data on at least half of HTP and LTP trials were excluded from our final analysis. For the remaining 32 infants in our sample, approximately 22% of trials (170 out of 768 total possible HTP and LTP test trials) were excluded from the analysis based on pre-screening and insufficient looking. A total of 598 trials (HTP trials: 293, LTP trials = 305) were included in the analyses, with each participant contributing an average of 19 trials.

4.2. Results and Discussion

In Experiment 3, we employed a similar model structure to Experiment 1 to examine whether infants are able to retain HTP and LTP word-referent mappings following a delay. We performed linear mixed-effects modeling in which we regressed infants’ trial-level accuracy as a function of Label Type (HTP vs. LTP) and Counterbalanced Language (A vs. B). The final model included a by-subject random intercept and a by-item random intercept.

Results revealed no significant main effects of Counterbalanced Language, β = .032, SE = .027, p = .234, 95% CI = [−.021 – .084] or Label Type, β = −.010, SE = .026, p = .695, 95% CI = [−.062 – .042], on overall accuracy (see Table 1). Planned one-sample t-tests comparing performance to chance revealed that infants successfully learned both the HTP (M = 58%, SD = 12%), t(31) = 3.92, p < .001, %95 CI = [.039 – .124], d = .69 and LTP labels (M = 58%, SD = 10%), t(31) = 4.42, p < .001, %95 CI = [.043 – .118], d = .78 (see Figure 3 and Figure 6). These data suggest that when infants were trained on word-referent mappings immediately after familiarization with a corpus, they retained both HTP and LTP mappings across a 10-minute delay.

Figure 6.

Mean proportion of time looking to target objects on HTP and LTP word trials as a function of time in Experiment 3. Dashed lines represented the analysis window. The ribbon around the lines indicates ± 1 SE.

Interestingly, in a recent study by Lany and colleagues (Lany et al., 2024), in which 23-month-old infants’ were trained and tested HTP and LTP mappings immediately following familiarization, performance (i.e., HTP: M = 61%, SD = 12%, LTP: M=57%, SD = 17%) was very similar to that of infants in our third experiment (HTP: M = 58%, SD = 12%, LTP: M=58%, SD=10%). Comparing the effect sizes in Lany et al.’s immediate condition (HTP: d = .94, LTP: d = .42) and Experiment 3 (HTP: d = .69, LTP: d = .78) suggests that the delay we included had little impact on the strength of the connections between labels and referents. Together, these data suggests that once mappings between labels and referents are formed, they appear to be remembered quite well regardless of the internal statistics of the labels.

5. General Discussion

Infants’ sensitivity to statistical distributions in their linguistic environment influences both segmenting potential word forms (e.g., Saffran et al., 1996, Pelucchi et al., 2009a) and mapping those word forms to referents (e.g., Graf Estes et al., 2007; Hay et al., 2011). Although word forms with strong statistics (e.g., high TPs between syllables) are better learned as labels at 17 months (Graf Estes et al., 2007; Hay et al., 2011), by 21 to 23 months HTP word forms are no longer advantaged in mapping tasks, at least when there is no need to remember word forms over time (Shoaib et al., 2018; Lany et al., 2024). Nevertheless, in Experiment 1, we found that 23-month-old infants are influenced by word forms’ TPs when there is a delay between encountering the word forms in speech and opportunities to map them to referents. Specifically, we found that infants successfully mapped HTP word forms to referents but failed to do so for LTP word forms.

Importantly, the results from Experiment 2 suggest that infants did not simply remember the HTP word forms across the delay and forget the LTP word forms, as LTP word forms were actually learned worse than Italian words with which infants had no prior experience. The same word forms (casa, bici, fuga, and melo) were used as HTP, LTP, and Unfamiliar labels in Experiments 1 and 2, respectively. The only difference was that in Experiment 2, the word forms had not been encountered prior to the Referent Training Phase, whereas in Experiment 1, infants heard a corpus in which these word forms had high and low TPs 10 minutes prior to the Referent Training Phase. Thus, if LTP word forms had been forgotten, they should have been learned equivalently to Unfamiliar words. Likewise, if HTP word forms were mapped significantly better than LTP word forms because they were better remembered, they should have been learned better than Unfamiliar word forms. However, infants in Experiment 2 mapped Unfamiliar word forms as well as infants in Experiment 1 mapped HTP word forms, and better than LTP word forms.

In contrast to Experiment 1, in Experiment 3 we tested whether mappings between HTP and LTP word forms and referents that are trained immediately can be retained across a 10-minute delay. Previous research by Lany and colleagues (Lany et al., 2024) suggests that 23-month-olds readily learn both HTP and LTP word forms as labels when they are trained on the mappings immediately after familiarization, and in the current study we found that these newly-formed word-referent mappings were remembered across a 10-minute delay, regardless of the word forms’ TPs. Altogether, the findings from the three experiments suggest that, although 23-month-old infants do not need strong statistics to map word forms to referents immediately, or to remember those mappings across a delay, infants are nevertheless sensitive to these statistics in the speech stream, and they influence mapping after a delay.

The findings from the current study are consistent with evidence that the role of statistical coherence in mapping word forms to referents changes across development. Earlier work showed that 17-month-olds readily learned HTP Italian word forms as labels but failed to map both LTP and unfamiliar Italian word forms (Hay et al., 2011). Those data suggest that at 17 months, strong statistics support word learning by facilitating the formation of a connection between a word form and a specific referent. The current results, consistent with previous findings (Shoaib, et al., 2018; Lany et al., 2024), suggest that although infants continue to track statistics in fluent speech at 23 months, the obvious boost they get from them in immediate mapping tasks has disappeared. Indeed, infants appear to successfully map HTP, LTP, and even Unfamiliar word forms to referents. Older infants’ success may result from a better ability to rapidly learn associations between word forms and referents without prior experience with those forms. The current study adds to these findings by suggesting that mappings between HTP and LTP word forms and referents, once learned, are remembered well across a 10-minute delay, regardless of the statistical coherence of the word forms.

Critically, when there is a delay between exposure to statistical cues relevant to word-form segmentation and opportunities to map those word forms to referents, the influence of those statistics becomes apparent at 23 months of age. Specifically, word forms with strong statistical coherence are mapped after a 10-minute delay, but word forms with low statistical coherence are not. Given that the data from Experiment 2 suggests that this is not because HTP word forms are remembered while LTP word forms are forgotten, it is important to ask why LTP word forms are not mapped after a delay. We suggest that it is because of lexical competition effects that emerge across the 10-minute delay period. A brief period of quiet rest can facilitate the consolidation of recently encountered information, serving to consolidate or stabilize memory for it (see Wamsley, 2019 and 2022 for reviews). This consolidation period can involve stabilizing memory for specific newly encountered items, such as word forms, but it can also lead to integrating across newly learned items. For example, after exposure to nonsense words like cathedruke that are neighbors to known words (e.g., cathedral), competition effects emerge across delay periods as this newly encountered information is consolidated and integrated. Specifically, participants are slower to recognize cathedral if they heard cathedruke prior to the delay, presumably because activation of cathedral is inhibited by competing activation from the new neighbor cathedruke (Lindsay & Gaskell, 2013). These competition effects are not apparent immediately after familiarization, instead only emerging after delays, especially delays that contain sleep (Dumay & Gaskell, 2007), but also after brief waking delays when neighbors (or related items) are both presented or activated, such as when familiarization with cathedruke is interleaved with instances of cathedral (Lindsay & Gaskell, 2013).

We suggest that in Experiment 1, initial consolidation and integration processes for memory traces of syllable sequences in the Familiarization Phase began during the 10-minute delay period. The syllables in the HTP sequences never occurred elsewhere in the Familiarization corpus, and thus as those sequences were consolidated and integrated, there were no activated neighbors to create competition. In contrast, the syllables in the LTP words occurred in other sequences (for example, in the case of the LTP word casa, they also heard caro, cadi, and cavo). All of these sequences would then undergo consolidation and integration. Lexical competition effects are evident by this age, even for newly-learned word forms (Swingley 2007), and the fact that LTP word forms and their neighbors were interleaved makes it especially plausible that competition effects between them could emerge during the delay (Lindsay & Gaskell, 201).

In sum, across a period of consolidation, HTP word forms may become relatively stable proto-lexical entries because TP cues suggest they are good words, and there are no neighboring word forms to compete with them. Across the same period of consolidation, LTP words may become less stable proto-lexical entries because of competition emerging during the delay. This in turn may lead them to be more poorly mapped during the subsequent Referent Training Phase. This explanation is reminiscent of ideas in the fast- (and slow-) mapping literature (e.g., Kucker et al., 2015), which suggests that although mappings for various word forms can be formed in the moments right after familiarization, this does not mean that all mappings will fare the same over longer timescales. In our paradigm, both HTP and LTP word forms can be mapped immediately, but their differential overlap with similar words in the corpus appears to lead to different patterns of mapping after a delay, potentially as a function of consolidation processes.

In sum, statistics appear to play a role in mapping previously encountered word forms following a delay, specifically by downgrading words with low statistical coherence. However, when HTP and LTP word forms are used as labels immediately, as in Experiment 3, both are readily mapped, suggesting that using HTP and LTP word forms as labels may stabilize their representations. In particular, the LTP word forms may be protected from weakening via competition once they have been mapped to referents. Future research should examine how the statistical coherence of word forms influences the stabilization of representations over longer delays. It would be particularly interesting to explore a potential role for sleep in this process, considering its well-established role in memory consolidation (Axelsson, Williams & Horst, 2016).

More broadly, the results of Experiment 3 suggest that by age 2, infants are able to retain at least four newly-learned mappings between word forms and referents across a 10-minute delay. Previous research shows that in the second year, infants appear to readily learn word-referent mappings across a range of conditions, but also that these mappings can rapidly fade when the conditions under which they are learned are challenging (Horst & Samuelson, 2008; for a comprehensive review see Kucker et al., 2018 and Wojcik, 2013). However, there is evidence that previous experience with the objects improves memory for mappings in fast-mapping (Kucker & Samuelson, 2012). In future work, it will be important to determine how previous experience with the word forms in fluent speech, as we gave infants in this study, impacts retention of these newly formed mappings. For example, infants may retain the mappings between recently segmented word forms and referents in Experiment 3 better than they would retain the mappings involving Unfamiliar word forms in Experiment 2.

While previous demonstrations, focusing on younger infants, have clearly documented the facilitatory effect that sensitivity to TPs has in earlier stages of lexical development, here we demonstrate that sensitivity to TPs may eventually suppress the mapping of word forms with relatively weak statistics. Our findings are generally consistent with evidence that adults learn LTP sequences as labels more slowly than HTP sequences (Mirman et al., 2008). Developing resistance to mapping LTP word forms in older infants may be just as important as the facilitatory effects that TPs have on forming mappings for younger infants (see Graf Estes et al., 2007; Hay et al., 2011, 2015; Graf Estes & Hay, 2015; Shoaib et al., 2018 for additional evidence of increasing specialization in what infants will accept as word forms across the second year of life).

In sum, our findings suggest that statistical learning may play a more nuanced role in language development than earlier work has suggested. Much work has focused on the possibility that statistical word segmentation allows novice word learners to find word forms that can be more easily mapped to referents than those with low coherence, thereby facilitating lexical development, the current results suggest that this is not the only role that these statistics play across development. Specifically, the current work suggests 1) that by 23 months infants learn and remember mappings between HTP and LTP words and referents equally well, and 2) that tracking these statistics may influence more skilled word learners by leading some potential word forms to be downgraded, or more poorly mapped to referents, especially across a delay. In future research, it will be important not only to replicate these findings, but also to determine the mechanism by which LTP word forms become downgraded as potential object labels across time.