Introduction
Parents usually use a simplified register when talking to their infants: a special register commonly referred to as Infant Directed Speech (IDS). IDS is characterized by adaptions in speech and language (for a review, see Soderstrom, Reference Soderstrom2007). IDS exhibits linguistic simplifications, such as shorter and less complex utterances (Kavanaugh & Jirkovsky, Reference Kavanaugh and Jirkovsky1982; Phillips, Reference Phillips1973; Snow, Reference Snow1977) and a simplified lexicon (Phillips, Reference Phillips1973). In addition, caregivers speak with a higher-pitched voice and wider pitch variations and use longer pauses compared to Adult Directed Speech (ADS) (Fernald, Taeschner, Dunn, Papousek, De Boysson-Bardies & Fukui, Reference Fernald, Taeschner, Dunn, Papousek, De Boysson-Bardies and Fukui1989; Fischer & Tokura, Reference Fischer and Tokura1996). Caregivers also speak slower by articulating slower and lengthening their vowels in IDS, particularly utterance-final vowels (Fischer & Tokura, Reference Fischer and Tokura1996; Soderstrom, Reference Soderstrom2007). Similar properties of speech to infants were found across different cultures and languages (Soderstrom, Reference Soderstrom2007).
What is the function of this specific register? In addition to the obvious affective function of IDS (Benders, Reference Benders2013), characteristics of IDS appear to be effective in attracting and maintaining the infant's attention (Wang, Houston & Seidl, Reference Wang, Houston, Seidl, Frühholz and Belin2018). Indeed, studies have found that infants, even newborns and 1-month-old infants, show a preference for IDS over ADS (Cooper & Aslin, Reference Cooper and Aslin1990; Fernald, Reference Fernald, Barkow, Cosmides and Tooby1992). In addition to maintaining the infant's attention, IDS seems to have a didactic function: infants who hear more IDS are more efficient in processing familiar words (Weisleder & Fernald, Reference Weisleder and Fernald2013). Furthermore, the amount of IDS appears to influence the expressive vocabulary: the more IDS children hear, the larger their expressive vocabulary tends to be (Hart & Risley, Reference Hart and Risley1995; Weisleder & Fernald, Reference Weisleder and Fernald2013). Additionally, this may lead to a more rapid growth of vocabulary (Huttenlocher, Haight, Bryk, Seltzer & Lyons, Reference Huttenlocher, Haight, Bryk, Seltzer and Lyons1991). Not only the quantity, but also the quality, of IDS seems to have an influence. The diversity of the vocabulary in caregiver's speech is a predictor of children's vocabulary production (Pan, Rowe, Singer & Snow, Reference Pan, Rowe, Singer and Snow2005).
IDS should not be conceived as a static phenomenon, but as dynamic and responsive to the child: parents fine-tune their speech to the child's (linguistic) interactional functioning. The fine-tuning hypothesis refers to the idea that IDS changes to adapt to the changing linguistic requirements and needs of the child (Snow & Ferguson, Reference Snow and Ferguson1977). Mean length of utterance (MLU) is an example of a characteristic of IDS that seems to be adapted to the changing needs of the child (Ko, Reference Ko2012; Murray, Johnson & Peters, Reference Murray, Johnson and Peters1990; Sherrod, Friedman, Crawley, Drake & Devieux, Reference Sherrod, Friedman, Crawley, Drake and Devieux1977). Much research has shown a decrease in the caregiver's MLU in the second half of the first year of the infant's life. This is probably because infants begin to show some understanding of several common words from around the age of 6 months onwards (Bergelson & Swingley, Reference Bergelson and Swingley2012). In response to this changing level of comprehension, caregivers may begin to fine-tune their utterances by adjusting their MLU (Genovese, Spinelli, Romero Lauro, Aureli, Castelletti & Fasolo, Reference Genovese, Spinelli, Romero Lauro, Aureli, Castelletti and Fasolo2020; Murray et al., Reference Murray, Johnson and Peters1990; Sherrod et al., Reference Sherrod, Friedman, Crawley, Drake and Devieux1977). This general use of shorter utterances in response to the limited abilities of the child is called coarse lexical tuning (Roy, Frank & Roy, Reference Roy, Frank and Roy2009). However, not every study found a change in MLU before and after the onset of lexical items (Kavanaugh & Jirkovsky, Reference Kavanaugh and Jirkovsky1982; Phillips, Reference Phillips1973; Snow, Reference Snow1977).
Roy et al. (Reference Roy, Frank and Roy2009) explored fine lexical tuning (i.e., the adjustment of IDS) at the level of individual lexical items in a case study. They identified for every word in the child's vocabulary its “birth”, i.e., the first time that the word was produced by the child. Then a time-series was created for each word, consisting of the MLU of the utterances containing that particular word in each consecutive month. A systematic decrease in MLU was found preceding the child's first production of that word. After word birth, the MLU did not immediately rise again. This suggests that there appears to be fine lexical tuning in IDS, because parents seem to modify their own utterances based on the child's knowledge of words.
These results imply that IDS should not be seen as a static phenomenon, but as dynamic and responsive to the child's needs. If IDS is dynamic, the evolution of IDS needs to be identified preferably in a longitudinal approach. Previous research often involved children at one particular age or points in time that were quite far apart (e.g., Phillips (Reference Phillips1973) compared IDS to infants of 8, 18 and 28 months). It is difficult to detect a detailed developmental path of IDS under those circumstances. To capture possible changes in the characteristics of IDS, repeated samples need to be taken relatively closely spaced over a sufficiently long period of time. Furthermore, chronological age may not be the most appropriate measure: if IDS is responsive to the child changing needs, then linguistic measures are required instead of chronological age since children of the same chronological age may develop at a different rate. For example, a ten-month-old infant may start producing words, while another one may only be starting to babble at that age. Thus, if IDS is sensitive to when children begin to use words, chronological age is not the most appropriate predicting variable. This is why linguistic milestones and measures of language development derived from a longitudinal corpus of caregiver-child interactions will be used in the present research.
Words in isolation
The fine-tuning hypothesis suggests that the MLU of the caregiver decreases when children start to produce their first words. Shorter MLU's can be the result of more words spoken in isolation. Brent and Siskind (Reference Brent and Siskind2001) found that 9% of IDS are isolated words. In a study by van der Weijer (Reference van der Weijer1998), this percentage was 6.9% when fillers, vocatives and social expressions were excluded. Brent and Siskind (Reference Brent and Siskind2001) also showed that the odds that a child would learn a word increased if that word was uttered frequently in isolation. This was a better predictor than the total frequency of that word. Furthermore, around 40% of the infant productive vocabulary were words that were spoken in isolation in IDS. A similar result was found by Swingley and Humphrey (Reference Swingley and Humphrey2018): 12- and 15-month old children understood and produced more often words that were heard in isolation or in shorter utterances.
In a study on novel word learning, Keren-Portnoy, Vihman, and Lindop Fisher (Reference Keren-Portnoy, Vihman and Lindop Fisher2019) showed that 12-month-olds were better in recognizing words presented in isolation than words that were presented sentence-finally. Isolated words were contrasted with sentence-final words, because parents often use unknown words sentence-finally in IDS. Furthermore, words at the edges of an utterances appear to be easier to segment for infants (Seidl & Johnson, Reference Seidl and Johnson2006). However, when these same words were presented in passages of text, neither words presented in isolation nor words presented sentence-finally were recognized. One reason that words in isolation are easier to learn concerns word segmentation. Infants have difficulties segmenting words in fluent speech, especially from the middle of a multiword utterance (Mattys & Jusczyk, Reference Mattys and Jusczyk2001; Seidl & Johnson, Reference Seidl and Johnson2006).
Not only words in isolation are helpful for learning words, but also shorter utterances. Grimm, Cassani, Gillis, and Daelemans (Reference Grimm, Cassani, Gillis and Daelemans2019) showed that the more a word occurs in short utterances in IDS, the earlier that word is produced by the child. Shorter utterances are better for predicting the time course of when a word is learned than a word's overall frequency. On the other hand, Arnon and Clark (Reference Arnon and Clark2011) showed that children pay attention to the lexical frame around a word. In an elicitation study, 4;6-year-old children performed better in producing irregular plurals when there was a highly correlated lexical frame present. The children performed poorly when they had to produce the words in isolation. Of course, these children were older than the children in the studies discussed before, so this could explain the different results.
Current research
The research question of the current study is: do parents tune their utterances to the emergence of words in the infant? Although there are many studies about coarse lexical tuning (Genovese et al., Reference Genovese, Spinelli, Romero Lauro, Aureli, Castelletti and Fasolo2020; Murray et al., Reference Murray, Johnson and Peters1990; Sherrod et al., Reference Sherrod, Friedman, Crawley, Drake and Devieux1977), there is much less known about fine lexical tuning. The research of Roy et al. (Reference Roy, Frank and Roy2009) on fine lexical tuning was a case study. The findings of Roy et al. (Reference Roy, Frank and Roy2009) indicate that adults tune their utterances to the emergence of words in the infant. Can the results of this study also be found with multiple subjects? In the current study, their research will be expanded to 30 typically developing children and their caregivers. From this corpus, the MLU of the parent's speech interacting with their child over time will be measured. Children will not be compared on their chronological age for this purpose, but on linguistic milestones. Specifically, word births will be used, as did Roy et al. (Reference Roy, Frank and Roy2009).
In order to answer the research question, the MLU in IDS of utterances containing specific words that enter a child's vocabulary at a particular moment will be calculated and compared over time. The expectation is that the MLU exhibits a U-shaped curve, with longer utterances as the child does not yet produce the word, followed by a decrease in MLU immediately before the child's first production of that word, and a corresponding rise afterwards. This expectation is based on the results of the case study by Roy et al. (Reference Roy, Frank and Roy2009). The study will also explore if the number of words a child already knows influences the fine-tuning of utterances. Arnon and Clark (Reference Arnon and Clark2011) showed that older children paid attention to the lexical frames around a word; thus it is expected that the more words a child produces, the higher the MLU of parent's speech around new word births will be.
Method
Participants
The data for the current study were taken from the CLiPS Child Language Corpus (CCLC), a collection of longitudinal recordings. This corpus consists of audio and video recordings of spontaneous speech of 30 Belgian Dutch acquiring monolingual typically developing children and the spontaneous interactions with their primary caretakers. The parents of all children were native speakers of Dutch, normally hearing, and from a mid-to-high SES background The children had no health and developmental problems and were monolingually raised (Molemans, Reference Molemans2011; van den Berg, Reference van den Berg2011; Van Severen, Reference Van Severen2011). Monthly recordings were collected between 6 months and 24 months of age. A total of 570 recordings were available. Recordings lasted on average 64 minutes (median = 63 minutes, range = 33 minutes to 114 minutes).
The current study was approved by the Ethical Committee for the Social Sciences and Humanities of the University of Antwerp (EASHW_17_53). All parents of the children signed an informed consent form.
Data collection and transcription
For the observation sessions, the parents were asked to interact with their children as they normally do in daily routines and free play. Of each original recording a selection of 20 minutes was made for transcription and coding by the researcher who was present at the recording. This selection was made in order to find a pragmatic balance between collecting a reasonable amount of speech material from each recording and keeping the transcription time within reasonable limits (the time-investment per transcription was about 14 hours). The researcher aimed for a selection in which the child was vocally active. Only uninterrupted interactions were selected, and long pauses and noisy parts were avoided. Each selection was transcribed using CHILDES’ CLAN program according to the CHAT conventions (MacWhinney, Reference MacWhinney2000). Children's and parents’ productions were orthographically and phonemically transcribed with stress marking. Target words of children's production were also transcribed phonemically: this meant that the adult equivalent of the word was added.
Ten percent of the corpus was used for an interrater reliability check: another transcriber retranscribed the utterances orthographically. The original and the retranscribed transcriptions were compared on two aspects: the content (identical words) of the utterances and the length of the utterances. The re-annotation of utterance content resulted in a percentage of agreement of 82%. For utterance length, the percentage of agreement was 91%. For intrarater reliability, 5% of the corpus was retranscribed by the original transcriber. The same process as the interrater reliability check was used. This resulted in a percentage of agreement of 88% for the re-annotation of utterance content and 93.5% for utterance length (Molemans, Reference Molemans2011; van den Berg, Reference van den Berg2011; Van Severen, Reference Van Severen2011).
Language measures
First, children's expressive vocabulary was collected using the CLAN software (MacWhinney, Reference MacWhinney2000). A Python script was written to create a cumulative word list and each word's first use and the child's age at that point were marked. Word types were added cumulatively: if a child produced a specific word type in a recording session, this word was added to the child's cumulative vocabulary if it was not already present. This is called a ‘word birth’ (Roy et al., Reference Roy, Frank and Roy2009). The result was a list of word births combined with the age when the word was first produced. Before the script was run, the list was lemmatized. This was done to ensure that inflectional variants of acquired words were not considered as new words. Furthermore, only the content words were kept for the current study.
Another Python script was written to compute the MLU based on number of words in IDS. For each word in the child's cumulative word list, utterances were identified in their parent's speech (main tier *ADU) containing that word. Then, for each month, the MLU was calculated from those utterances by dividing the number of words by the number of utterances. A time series of the MLUs for each word for each parent was the result. The time series were aligned by the time the word was first produced by the child. This means that the month when the word was first produced by the child was set at month zero, and consequently the month before was month -1, and the month after month 0 was set to month + 1, etc.
Statistical analysis
Statistical analyses were done in the software R (R Core Team, 2018) using Multilevel Modeling (MLM) (package lme4) (Bates, Maechler, Bolker & Walker, Reference Bates, Maechler, Bolker and Walker2015). MLM consists of a random and a fixed part. The dependent variable of the model was the MLU of words in IDS. The fixed effects were the linear, quadratic and cubic effects of time, measured as months from the child's first production of a word, the child's cumulative vocabulary, and possible interactions between two effects.
Child and word were introduced as the random effects. The model contained at the level of children and words both random intercepts and random slopes for the linear effect of time. As such, the variation between children and between words was considered, assuming that the effect of time may differ between children and between words.
The models were built step by step first entering random effects and then adding the fixed effects one by one. If a fixed effect improved the model, the effect was kept in the model; otherwise, it was left out of the model. The cut-off level of significance for the analysis was set at p = 0.05. The best fitting model is reported. For the data analyses, the MLU of words from 17 months before word birth, to nine months after word birth, were analyzed. The other months were excluded, because there were too few data points (N < 100).
Results
Cumulative vocabulary
From the transcripts, 5,375 word births for the children were extracted. The mean age of the children was 1;1.16 when they used their first word (median = 1;1.7, range = 0;11.0 to 1;4.2). The children had on average a cumulative vocabulary of 179 words (median = 197, range = 56 to 247) at the age of 24 months. These word births led to 18,684 data points of MLU in infant directed speech.
Figure 1 shows the development of the cumulative vocabulary: per month the mean cumulative number of words is plotted, together with 95% confidence intervals. The individual cumulative vocabulary counts are presented in Appendix 1.
Study children's development of cumulative vocabulary
Development of MLU of utterances with target words
Do parents tune their utterances relative to the emergence of words in infants? A multilevel model was fitted with random effects for each child and each word for the linear effect of time. This resulted in random intercepts and slopes for each child at each month and for each word at each month. The corresponding model can be found in Table 1. The results revealed variations in MLU depending on the number of months from word birth (p < 0.0001). Specifically, the modification of MLU in IDS followed a U-curve according to a significant quadratic and cubic trend (p < 0.05), shown in Figure 2. MLU decreased as word birth approached and increased again shortly after word birth. There was no significant main effect of cumulative vocabulary, meaning that the U-shaped curve occurred irrespective of the child's cumulative vocabulary level. However, parents increased their utterance length more over time as the child acquired a larger vocabulary, shown by the significant interaction between cumulative vocabulary and number of months from word birth (p < 0.05). Age, the quadratic effect of age and other interactions between effects did not improve the model, so these effects were not included in the model reported in Table 1.
Development of MLU in IDS relative to word births (predicted values).
Parameters estimates of parent MLU of utterances with target words
MLU at the time of word birth
Does the child's vocabulary size have an influence on the parents’ MLU at the time of new word births? In order to answer this question, the MLU of the parents at the month a child first produced a word (month 0) was extracted for each word. The cumulative vocabulary of the child at that age was added. A multilevel model was constructed with both random intercepts and random slopes for the linear effect of cumulative vocabulary at the level of children and words. With this addition, it is assumed that each child and each word can have different intercepts and that the effect of cumulative vocabulary can differ for each child and each word. The corresponding model can be found in Table 2.
MLU in IDS at the time of word birth
The results revealed that vocabulary size influenced the MLU of parents at the time of new word births. This is shown by a significant main effect of cumulative vocabulary at month 0 (p < 0.001), as depicted in Figure 3. Parents used longer utterances with a particular target word when the child had a higher cumulative vocabulary at the moment that word was first produced by the child. A quadratic effect of cumulative vocabulary did not improve the model, so it was not included in the model.
Parent MLU of target words in IDS at the time of word birth (predicted values).
Discussion
The current research addressed the question whether parents tune their utterances to the emergence of words in the infant. Our results demonstrated that the length of the utterances with a particular target word decreased as the child's production of that target word came closer. The utterance length increased again afterwards, but not to the level it was at the beginning of the period. This finding is in accordance with the finding reported by Roy et al. (Reference Roy, Frank and Roy2009). The change in MLU suggests that parents use some level of fine lexical tuning in their speech to children in addition to coarse fine tuning, the general use of shorter utterances.
Why do parents shorten their utterances containing words that are on the verge of appearance in their children's spontaneous speech? A possible explanation is that they try to scaffold their children with word learning. Shorter MLU's can mean that there are more words spoken in isolation. A previous study by Brent and Siskind (Reference Brent and Siskind2001) found that children learned words better when they occurred in isolation. The frequency of hearing a word in isolation was a better predictor of word learning than the total frequency of that word's exposure (see also Keren-Portnoy et al., Reference Keren-Portnoy, Vihman and Lindop Fisher2019; Ninio, Reference Ninio2016; Swingley & Humphrey, Reference Swingley and Humphrey2018). Next to words in isolation, shorter utterances are also helpful for learning words (Grimm et al., Reference Grimm, Cassani, Gillis and Daelemans2019; Swingley & Humphrey, Reference Swingley and Humphrey2018). These results are in line with the finding of the present study that parents’ MLU decreases around the time a word is first produced, a finding similar to the one reported by Roy et al. (Reference Roy, Frank and Roy2009) in a case study.
The decrease in MLU is seen prior to word birth. This may be attributed to the well-known phenomenon that young children's comprehension precedes production (Benedict, Reference Benedict1979), and parents possibly act on their child's understanding of a word by shortening their utterances containing that word. However, this hypothesis can only be corroborated by investigating when a word is first understood by the child, and by relating parents’ MLU relative to that point.
However, according to Arnon and Clark (Reference Arnon and Clark2011), children pay attention to the relation between words, and such lexical frames help children acquire words. For instance, “brush your teeth” is more helpful in acquiring “teeth” than that word in isolation. In this respect, the occurrence of a particular word in a specific lexical frame would be more helpful, and not the occurrence of that word in short(er) utterances around word birth per se. This alternative was not explicitly addressed in the current study: however, the mean estimated MLU was never lower than 3. This may indicate that the majority of the words are not spoken in isolation, but in lexical frames. Furthermore, the children in the current study were younger than in the experimental study of Arnon and Clark (Reference Arnon and Clark2011). In the current study, the children were followed from 6 months until 2 years old. The children in the study of Arnon and Clark (Reference Arnon and Clark2011) were 4;6 years old. Possibly, older children, who have a larger cumulative vocabulary, benefit more from framing than from shorter utterances. This is supported by our data, as vocabulary size had an influence on the MLU of parents at the time of new word births: the more words children knew, the higher the MLU of parent's speech around word birth. This could mean that there is less fine-tuning in terms of MLU of a new word when children get older and have larger vocabularies.
There are some limitations of the current study. First, the children were only recorded once a month. Most probably not all the word births in the current study are estimated completely accurately. We looked at the first time a child produced a word in the transcriptions, but it could be that this word was already in the expressive vocabulary of the child. Furthermore, the current study contained less dense data than the study of Roy et al. (Reference Roy, Frank and Roy2009), because they followed a single child every day from birth until the child's third birthday. This makes the current study less precise.
Conclusion
The current study demonstrated that parents modify their utterances relative to the emergence of words in their children. Parents shorten their utterances with a particular word as the word is on the verge of appearing in the child's production. This study replicated the finding of the case study of Roy et al. (Reference Roy, Frank and Roy2009), now with a larger number of subjects. In conclusion, this indicates that parents use fine-lexical tuning when talking to their children: parents use shorter sentences to scaffold children's word learning.
Acknowledgements
We would like to thank the families and infants that participated in the study, and K. Schauwers, I. Molemans, R. van den Berg and L. Van Severen for collecting the CLiPS Child Language Corpus. The research reported in this article was supported by grant G.0235.18 of the Research Foundation Flanders (FWO) to S. Gillis.
Overview of the cumulative vocabulary of children
