Factors impacting word learning in LTs

One explanation for toddlers’ rapidly accelerating vocabulary size is fast-mapping, the process by which one can quickly map or initially associate words with their referents with minimal exposure. This process initiates the encoding of new words and forms memory traces of their representations through brief and incidental mappings to referents (Kucker & Seidler, 2022). LT toddlers have poorer fast-mapping ability than those with typical language development (TLD) (Asadi et al., 2021; Weismer et al., 2013). The difficulties of LTs in learning words could be attributed to their poorer representation in encoding novel words, as compared to their peers (Kucker & Seidler, 2022). This representation is more easily compromised in novel word production tasks than in novel word comprehension tasks (Weismer et al., 2013).

McMurray et al. (2012) proposed a dynamic model to explain how children learn new words in the context of ambiguity; that is, multi-referents. In the past, a constraint approach was used to explain how children solve the problem of referential ambiguity; that is, what the novel word refers. Constraints, such as mutual exclusivity and novel name–nameless principles, help children disambiguate references. However, McMurray et al. (2012) proposed that associative learning is sufficient to explain how children learn new words. According to their model, if children establish a strong and reliable connection between a novel label and its referent during the first encounter, they should be able to locate the referent more quickly in subsequent similar situations where multiple referents are present and they hear the label again. Once again, to find the newly mapped referent can serve as an indicator of the robustness of the newly built representation.

Other factors that are closely associated with word learning are phonological working memory and speech perception abilities. Sufficient phonological working memory is a prerequisite for encoding representations of spoken words (Baddeley et al., 1998), and is often measured by non-word repetition (Moyle et al., 2011). Non-word repetition performance in LTs is not comparable to that of those with TLD (Stokes & Klee, 2009a, 2009b). Speech perception abilities are another foundation for learning words (Chen et al., 2016; Gervain & Mehler, 2010). Mandarin-speaking toddlers must learn how to effectively encode lexical-tones of novel words because lexical-tone carries the meaning of a syllable in Mandarin (Singh & Fu, 2016). For example, /ma/ with Tone 1 means ‘mother’ while /ma/ with Tone 3 means ‘horse’. Poor perception of lexical tones may result in children being less accurate in encoding the phonological representations of newly learned words. Pitch contours of vowels manifest the lexical-tones of syllables in Mandarin, that is, Tone 1 with the level contour and Tone 3 with the dipping contour (Liu et al., 2007). The abilities of those with TLD to perceive lexical tones and vowels could predict their vocabulary acquisition (Lu & Tsao, 2014; Tsao et al., 2004). Mandarin-speaking LTs cannot perceive lexical tones as accurately as their peers (Chen et al., 2016), and difficulties in lexical-tone perception are negatively related to their ability to learn new words (Lu & Tsao, 2014). Based on the aforementioned studies, phonological working memory and lexical-tone perception might serve as predictors for novel word-referent mappings (W-R mappings), particularly among LTs learning a tonal language.

Methods for capturing the word-reference mapping procedure

The looking-while-listening (LWL) eye-tracking paradigm is suitable for detecting subtle real-time processes during word-learning tasks (Ellis et al., 2015; Weighall et al., 2017). Automated eye-tracking technology enables the expansion of the LWL paradigm to include a greater number of objects, while also providing precise monitoring of various eye movement parameters and the assessment of changes in the timing of eye movements. Gaze-contingent stimulus presentation proves particularly valuable for younger participants with fluctuating attention. Fernald and Marchman (2012) used the LWL paradigm to test the processing efficiency of comprehension of familiar words in LTs and those with TLD. They found that LTs with faster reaction time, that is, processing efficiency at 18 months, were more likely to gain greater vocabulary at 30 months. Peter et al. (2019) also found the same trend that fast processing speed positively predicts subsequent vocabulary and syntactic growth. These studies signified that temporal related parameter in the LWL paradigm is a sensitive index for word processing abilities.

Ellis et al.’s (2015) study is a notable contribution that employed eye-tracking techniques to explore the online process of word learning in LTs. To assess LTs’ real-time processing of novel word learning, Ellis and her team employed three indices: target divergence time, the proportion of time spent looking at the target versus the distractor, and the latency of the first fixation. In our study, we built upon the strengths of Ellis et al.’s work, using the same eye-tracking measures. We further employed multilevel models (MLMs) for data analysis. This approach enabled us to gain a more detailed understanding of dynamic changes in these measures over time and to statistically evaluate performance differences between LT and TLD groups.

Additionally, in Ellis et al.’s (2015) study, infants learned two novel word-referent pairs by looking at the object moving back and forth on the screen while the novel word played seven times. After the training, infants had to choose the correct object between the target and the distractor (another newly learned novel object) pictures after hearing the newly learned novel word. Ellis et al.’s method revealed that LTs face challenges in new word learning, but it did not specify the particular stage at which these challenges occur, whether during fast-mapping or later memory and retention phases. Our study focused on evaluating the word-reference mapping ability of LTs, emphasising conditions characterised by reference ambiguity, and specifically concentrating on the early phase of word learning.

The current study

Participants

At 27 months, a total of 72 participants were recruited, with 71 participants completing the data collection: 35 in the LT group and 36 in the TLD group. By 33 months, one participant in the LT group had dropped out, seven did not complete the eye-tracking task due to fussiness, two did not pass the vocabulary checklist for the eye-tracking experiment, and data from three participants were unusable due to the experimental procedural errors. In the TLD group, all participants returned to the lab at 33 months to complete the eye-tracking test, but only those matched with the remaining cohort of LT participants at 27 months were included in the current study sample. Consequently, the sample consisted of 22 Mandarin-speaking LTs [15 boys; Mean (M) = 27.60 months, SD = 2.85] and 22 Mandarin-speaking toddlers with TLD (15 boys; M = 27.70 months, SD = 2.74). These participants were monolingual Mandarin speakers. The mean age of participants did not differ significantly between groups (p = 0.912). All participants were recruited from parenting websites or local paediatric clinics in northern Taiwan. The LT group was defined as having a smaller expressive vocabulary size [percentile rank ≤15 in the Mandarin Chinese version of the Words and Sentences Forms of the MacArthur-Bates Communication Development Inventory-Toddler (MCDI-T; Liu & Tsao, 2010), while participants in the TLD group had a normal expressive vocabulary size with a percentile rank >25. Participants in both groups had normal cognitive and fine-motor abilities (evaluated using the Bayley Scales of Infant and Toddler Development-III; Bayley, 2006), and passed the Modified Checklist for Autism in Toddlers (M-CHAT; Robins et al., 1999), a screening for autism. Each toddler in the LT group was matched with a TLD toddler with the same sex and birth order. The age difference between participants in the LT group and their pairs was within 1 month. According to parental reports, all participants were born after at least 36 weeks of gestation, with birth weights >2500 g, and were free from prenatal and perinatal complications. Participants had no reports of other critical incidents, chronic diseases or sensorimotor deficits.

Table 1 presents the cognitive and fine-motor abilities of the TLD and LT participants. The LT group exhibited lower fine-motor skills than the TLD group [F(1, 42) = 5.16, p = 0.028]. No significant differences in cognitive abilities between the two groups were noted (p = 0.336). Additionally, the maternal educational level of the TLD group was higher than that of the LT group [F(1, 42) = 5.59, p = 0.023]. Vocabulary and syntactic skills, based on parental reports, also were lower in the LT group than in the TLD group [Fs(1, 42) = 141.03 and 94.33, respectively; p < 0.001; η_p²s = 0.77 and 0.69].

Measures at 27 months

Experimental stimuli

The stimuli comprised static images of four animals presented on a white background. The animal images used in the experiment were bright and colourful [mean width: 410.8 pixels (SD = 25.3 pixels), mean height: 349.2 pixels (SD = 11.4 pixels)] with each image placed at the centre of four quadrants on the screen. The sizes of the four images were similar such that the chance of participants looking at the four images was the same under the context without semantic priming.

The word-learning task comprised six experimental trials, with two control trials using familiar words and four experimental trials using novel words. Each trial presented four animal pictures, with one target picture and three distractors. Considering the potential right visual field preference in LTs (Ellis et al., 2015), we designed a multiple-object condition (i.e., four images) to reduce the effect of participants' preference about the target image's location on the interpretation of word-learning process. In the familiar-word trials, the target words were familiar animals such as ‘pig’ (/xiao3-zhu1/) and ‘frog’ (/qing1-wa1/). In the novel word trials, the target words were unfamiliar animal names like ‘rong2-yuan2’ (salamander) and ‘bian3-fu2’ (bat), or nonwords like ‘zi3-xie1’ and ‘chuang2-jia3’. All target words comprised two syllables. To serve as distractors, 18 well-known animals familiar to children aged 2 to 4 years were selected. The positions of the novel animals on the screen were counterbalanced across the trials. Additionally, two carrier phrases (XX zai4-na3-li3? ‘where is the XX?’ and XX bu2-jian4-le ‘XX disappeared’) and six target words were audio recorded and then concatenated into the final audio utterances. The auditory stimuli were recorded by a native Mandarin female speaker, using a mono channel at a sampling rate of 44 kHz with a digital recorder. The mean duration of familiar/novel words was 1157 ms (SD = 19 ms). The mean duration of the audio string was 2 to 3 s.

Apparatus

Eye-tracking data were collected using Tobii T60 (Tobii Technology, Stockholm, Sweden), with Tobii Studio 3.0.0.128 software and a sampling rate of 60 Hz. The visual stimuli were presented on a 17’’ monitor with a screen resolution of 1280 × 768 pixels, and the auditory stimuli were played using loudspeakers (BOSE 240 V AP) at 70 dBA.

Procedures

The experimental procedure of this study is illustrated in Figure 1. Participants visited the laboratory at two time points. Basic information and baseline data were collected at Time 1. Considering the toddlers’ short attention span, data collection at Time 1 was divided into two laboratory visits. During the first visit, participants were administered Bayley-III, and participants’ parents completed the M-CHAT and MCDI-T and provided demographic information. During the second visit, tasks for testing early language processing abilities were performed. Data collection from participants and their parents was conducted in a quiet room. The duration of the laboratory visit was approximately 1−1.5 h.

At Time 2, 6 months after Time 1, participants visited the laboratory again for the eye-tracking experiment. During the experiment, the children sat on their parents’ lap in a quiet room, 60 cm away from the screen. Caregivers wore headphones and heard masking music and were instructed not to look at the screen to prevent interference with participants’ responses. At the beginning of the experiment, a five-point infant calibration within Tobii Studio was used, and participants watched the video clips for approximately 5 min. Additionally, before the eye-tracking experiment began, the parents had to complete a vocabulary checklist to ensure that the children were familiar with the known animals and could not name the novel animals.

Each experimental trial comprised three phases (Figure 2A): the showing phase for attention-getting, W-R mapping phase for target word and its referent mapping, and word-identification test (W-I test) phase for testing W-R mapping outcome. During the showing phase (0–18000 ms from the trial onset), four animals were simultaneously displayed on the screen for the first 5 s. To engage the toddler's attention, the four animals took turns disappearing for 1 s at 5000, 9000, 13000 and 17000 ms, respectively. The order of disappearance was pseudorandomized regardless of the target and distractors. As this phase lasted from 5000 to 18000 ms (i.e., 13 s) after the trial onset, each of the four items in one trial had an equal appearance duration of 12 s and a 1 s disappearance. After that, four animals reappeared on the screen to start the W-R mapping phase (18000–32000 ms from the trial onset), and the children listened to the utterance ‘XX. XX, zai4-na3-li3?’ (XX. XX, where is it?) while looking at four animals on the screen. In the last 6 s of the W-R mapping phase (28000–32000 ms from the trial onset), children listened to the utterance ‘XX, bu2 jian4 le’ (XX, disappeared) paired with the target animal, which then disappeared gradually within the last 4 s. Next, during the W-I test phase (32000–38000 ms from the trial onset), the four animals reappeared on the screen at different positions, and the children listened to the phase with the new word, ‘XX, zai4-na3-li3’ (XX, where is it?) and were asked to look at the corresponding image of the newly learned word. At the end of the trial, the children received cartoon pictures as reinforcers (38000–43000 ms from the trial onset).

To familiarise toddlers with the apparatus and experimental procedures, a familiarisation trial was conducted before the six experimental trials began. The trial used words from the fruit vocabulary typically known to 2-year-old Mandarin-speaking toddlers, along with corresponding images. Before starting this familiarisation trial, the fruit vocabulary list was cross-checked with the parents to ensure the child's familiarity with the words. Similar to the experimental trials, this familiarisation trial also comprised three phases: the showing phase, the W-R mapping phase, and the W-I test phase. The verbal prompts used throughout this trial mirrored those of the experimental trials, including expressions such as ‘XX, XX, zai4-na3-li3?’ (XX, XX, where is it?), ‘XX, bu2 jian4 le’ (XX disappeared), and ‘XX, zai4-na3-li3?’ (XX, where is it?). The eye-tracking trajectories of one participant in a novel-word trial are shown in Figure 2B.

The Research Ethics Committee of National Taiwan University approved this study, and the parents of all participants provided informed, written consent before participating.

Eye movement metrics and data analysis

Data analysis

To answer the first research question “Is there a difference in the expressive and receptive language abilities, phonological memory, and lexical-tone perception between LT and TLD toddlers?”, we compared the performance of participants in receptive and expressive language abilities, lexical-tone perception, and phonological working memory between the LT and TLD groups. Analyses of covariance (ANCOVAs) were performed to examine receptive and expressive language abilities, lexical-tone perception and phonological working memory, with the covariates of participants’ fine-motor ability and maternal educational level.

For the second question, “Do LT toddlers show difficulties with novel W-R mapping in word learning?”, we used the eye-tracking metric that demonstrated preferences toward the target image as shown in the toddlers’ fixations to the target image over time between the TLD and LT groups. The MLMs for change were used to estimate models that include estimates of intra-individual change (i.e., repeated fixation on the target or not, referring to multiple points in times; Level 1) over time, as well as inter-individual variability (i.e., individual characteristics; Level 2). The trajectories of familiar on showing phase, novel on showing phase, familiar on W-R mapping phase, novel on W-R mapping phase, familiar on W-I test phase, and novel on W-I test phase, were examined separately. This is similar to growth curve analyses (Nayar et al., 2022) but is estimated using the MLM framework. Specifically, orthogonal polynomial terms, each representing a different pattern of looking, were added in Level 1: (1) the linear time reflected an increase in fixations to the target images over time in a linear fashion; and (2) the quadratic time reflected the dynamic nature of fixating on the AOIs or switching from the target images to non-target images. The Level 2 model included group (TLD, LT; reference = the TLD group) on intercept, linear time and quadratic time to examine the interaction between the late-talking, linear and quadratic change rate of participants’ fixations to the target, and participants’ fine-motor ability and maternal educational level as adjustments for confounders. Collectively, the set of analyses not only provided information on the dynamic patterns of looking via MLMs but also delineated the difference by which groups diverge from one another. In addition, the total fixation time in the time windows of the showing, W-R mapping and W-I test phases were compared between the LT and TLD groups. If the TLD group noticed and looked at the AOIs sooner, the total fixation time should be significantly longer than that of the LT group.

Finally, to answer the question, “Do expressive and receptive language abilities, phonological memory, and lexical-tone perception predict novel word-referent mapping ability?”, MLMs were also used to estimate the fixations on the AOI over time in the W-I test phase of the novel. The models included estimates of change in intra-individual fixations on the target image over the period (Level 1) and inter-individual variability (Level 2) for receptive and expressive language abilities, lexical-tone perception and phonological working memory and other covariates (fine-motor ability and maternal educational level) in the trajectories. Additionally, we performed a hierarchical multiple regression analysis to investigate the predictors of the total fixation time of the novel images during 500–3350 ms of the W-I test phase (i.e., 32500–35350 ms from the trial onset). The regression model involved seven predictors entered in steps. Step 1 included maternal education and the participants’ cognitive and fine-motor abilities, and Step 2 involved receptive and expressive language abilities, lexical-tone perception and phonological working memory. The results reported the regression statistics (β and p) for the final models (i.e., when all predictors were entered).

The HLM software package, version 7.03 (Scientific Software International Inc., Skokie, IL) (Raudenbush et al., 2016) was used to fit the MLMs. The models were fitted using a restricted maximum likelihood estimation. For the MLMs, a population-averaged model with robust standard errors is described subsequently. Other statistical analyses in this study were performed using IBM SPSS Statistics for Windows, Version 25.0.0 (IBM Corp.), with a significance level (α) of p < 0.05.

Is there a difference in the expressive and receptive language abilities, phonological memory and lexical-tone perception between the LT and TLD toddlers?

Table 1 presents participants’ expressive and receptive language abilities, phonological memory and lexical-tone perception at Time 1. The LT group exhibited lower receptive language abilities [F(1, 40) = 6.53, p = 0.015, η_p² = 0.14] and expressive language abilities [F(1, 40) = 89.26, p < 0.001, η_p² = 0.69] than the TLD group after the maternal educational level and fine-motor ability were considered as covariates. Moreover, results of the ANCOVA showed that the LT group exhibited less accurate perception of lexical tones [F(1, 40) = 15.77, p < 0.001, η_p² = 0.28] and poorer phonological working memory [F(1, 40) = 43.48, p < 0.001, η_p² = 0.52] than the TLD group.

Do LT toddlers show difficulties with novel W-R mapping in word learning?

To understand the difference between the TLD and LT groups in terms of the change in fixation toward the target image during the showing, W-R mapping and W-I test phases, the MLM for change was conducted for an estimation. During the familiar and novel showing phases (i.e., without auditory label), no significance was found in the linear time and the quadratic time (familiar: ps = 0.638 and 0.817; novel: ps = 0.190 and 0.927), and the groups did not differ in their linear and quadratic rate of change (familiar: ps = 0.147 and 0.186; novel: ps = 0.470 and 0.785). As shown in Figure 3A,B, participants’ fixations to the target image did not change over time.

During the familiar and novel W-R mapping phases (i.e., with auditory label), significance was found in the linear time (β = 0.001916 and 0.001622, 95% confidence intervals [CIs] = 0.000176–0.003656 and 0.000076–0.003168, p = 0.037 and 0.046); however, the groups did not differ in their estimated intercept and linear and quadratic rate of change (familiar: p = 0.089, 0.080, 0.574; novel: 0.127, 0.066, 0.313). As shown in Figure 3C,D, participants’ fixations to the target image increased over time, but the change in the fixations on the target over time did not differ between the TLD and LT groups. During the familiar W-I test phases (i.e., with auditory label), significance was found in the linear time (β = 0.004124, 95% CI = 0.002048–0.006200, p < 0.001) and the quadratic time (β = −0.000023, 95% CI = −0.000035 to −0.000011, p < 0.001); however, the groups did not differ in the estimated intercept and linear and quadratic rate of change (p = 0.522, 0.579 and 0.517). As shown in Figure 3E, participants’ fixations to the target image increased over time, but the change in the fixations to the target over time did not differ between the TLD and LT groups.

During the novel W-I test phase, significance was found in the linear time (β = 0.004791, 95% CI = 0.002986–0.006596, p < 0.001) and the quadratic time (β = −0.000024, 95% CI = −0.000034 to −0.000014, p < 0.001). Moreover, the fixations to the target at the onset of the trial did not differ between the TLD and LT groups, as indicated by their estimated intercept (p = 0.715); however, the groups differed in their linear (β = −0.003154, 95% CI = −0.005465 to −0.000843, p = 0.011) and quadratic (β = 0.000018, 95% CI = 0.000004−0.000032, p = 0.007) rate of change in the W-I test phase. The two groups still differed in their linear (adjusted β = −0.003154, 95% CI = −0.005464 to −0.000843, p = 0.011) and quadratic (adjusted β = 0.000018, 95% CI = 0.000004−0.000031, p = 0.007) rate of change in the novel W-I test phase after controlling for maternal educational level and children's fine-motor ability. As shown in Figure 3F, the fixations at the target in the TLD group increased more than that in the LT toddlers. We further compared the total fixation time in the W-I test phase (500–3350 ms) between the LT and TLD groups (LT: M = 383.03 ms, SD = 285.16 ms; TLD: M = 603.15 ms, SD = 435.52 ms), and found that total fixation time in the TLD group was longer than that in the LT group [F(1, 40) = 4.91, p = 0.033, η_p² = 0.11] after the maternal educational level and fine-motor ability were considered as covariates. The TLD group indeed looked at the target image longer than LT group in novel W-I test phase.

Do expressive and receptive language abilities, phonological memory and lexical-tone perception predict novel W-R mapping ability?

The MLM was used to estimate the change in the fixation toward the target image on the W-I test phase of the novel word as associated with expressive and receptive language abilities, phonological memory and lexical-tone perception. Participants with better early expressive language ability could notice and look at the target more quickly 6 months later when they retrieved the novel words (β = 0.000294, 95% CI = 0.000016−0.000572, p = 0.045). The trend remained (adjusted β = 0.000306, 95% CI = 0.000020−0.000592, p = 0.043) after the maternal education, cognitive and fine-motor abilities were considered as covariates.

The bivariate correlations between expressive and receptive language abilities, phonological working memory, lexical-tone perception and total fixation time in the novel image during 500–3350 ms from the onset of the W-I test phase (i.e., 32500–35350 ms from the trial onset) are presented in Table 2. We found that the better the toddlers’ scores for expressive language (r = 0.36, p = 0.016), the longer their total fixation time on the novel image during the W-I test phase. The better the toddlers’ scores for lexical-tone perception (r = 0.49, p = 0.001) and phonological working memory were (r = 0.78, p < 0.001), the better their expressive language ability was.

Hierarchical multiple regression analysis assessed the predictors of total fixation time during 500–3350 ms from the onset of novel's W-I test phase. Maternal educational level and participant's cognitive and fine-motor abilities, added at Step 1, did not make a significant contribution to the prediction of total fixation time (p = 0.573, 0.988 and 0.537). In Step 2, which includes receptive and expressive language, phonological working memory, and lexical-tone perception at 27 months, we observed that neither receptive language, phonological working memory nor lexical-tone perception contributed significantly to the prediction (p = 0.616, 0.484 and 0.900); however, expressive language ability (adjusted β = 0.66, p = 0.025) was a significant predictor of total fixation time during the novel W-I test phase (500–3350 ms). Another hierarchical multiple regression analysis was used to assess the predictors of expressive language ability. Maternal educational level and participant's cognitive and fine-motor abilities, added at Step 1, could significantly contribute to the prediction of expressive language ability (adjusted R² = 0.13, p = 0.036). The addition of phonological working memory and lexical-tone perception at Step 2 significantly contributed to prediction (adjusted R² = 0.69, p < 0.001). In Step 2 of the hierarchical regression model, with all variables included, greater cognitive ability (adjusted β = 0.23, p = 0.030), lexical-tone perception (adjusted β = 0.22, p = 0.030) and phonological working memory (adjusted β = 0.70, p < 0.001) were associated with greater expressive language ability.

Limitations and future research directions

Although the diagnosis of LTs can occur between 18 and 35 months (Hawa & Spanoudis, 2014), we recruited participants older than 2 years to reduce the risk of false diagnosis. The LT group was initially diagnosed at Time 1, and the eye-tracking data were collected at Time 2, which was 6 months after the diagnosis. As we did not reassess their expressive vocabulary size, it is possible that some participants in the LT group could have caught up with their language development by that time. Further research is required to assess whether the rate of growth in expressive vocabulary can predict word learning in toddlers. Second, we increased the number of images from the typical 2 to 4 to minimise participants’ preconceived expectations regarding the location of the target image. Nevertheless, we recognise that the inclusion of extra distractor images could potentially place greater demands on participants’ working memory and cognitive processing.

Lastly, the Bayley-III used in this study was translated from the English version into Mandarin, with modifications to consider Mandarin and Chinese cultural characteristics. Although norms for the Bayley-III have not been established for infants and toddlers in Taiwan, a previous study demonstrated that the Bayley-III is suitable for Taiwanese children (Yu et al., 2013). Many studies in Taiwan have also used English normed scores for analysis (Lin et al., 2020, 2021). Therefore, we derived the scale scores from the norms of the Bayley-III English version and used these scores in analyses. These scores were not used as criteria or cutoff scores to group participants in the study. Instead, they were employed to investigate whether there were any group differences between the LT and TD groups and to assess whether there were any predictors at 27 months of age. Therefore, it is important to note that the results of this study only present the relative association between early holistic language abilities in children and subsequent word learning. Additionally, the Bayley-4 has been translated into Traditional Chinese and normed for Taiwan (Guo, 2023). Future related research can use the Bayley-4 Chinese version to enhance the validity of assessing children's developmental abilities.