The Road to Language Learning Is Not Entirely Iconic: Iconicity, Neighborhood Density, and Frequency Facilitate Acquisition of Sign Language

Abstract

Iconic mappings between words and their meanings are far more prevalent than once estimated and seem to support children’s acquisition of new words, spoken or signed. We asked whether iconicity’s prevalence in sign language overshadows two other factors known to support the acquisition of spoken vocabulary: neighborhood density (the number of lexical items phonologically similar to the target) and lexical frequency. Using mixed-effects logistic regressions, we reanalyzed 58 parental reports of native-signing deaf children’s productive acquisition of 332 signs in American Sign Language (ASL; Anderson & Reilly, 2002) and found that iconicity, neighborhood density, and lexical frequency independently facilitated vocabulary acquisition. Despite differences in iconicity and phonological structure between signed and spoken language, signing children, like children learning a spoken language, track statistical information about lexical items and their phonological properties and leverage this information to expand their vocabulary.

Learning a new word requires mapping information about its form to its meaning. In spoken languages, children use phonological structure to facilitate this mapping; for example, phonological neighborhood density (the number of lexical items that are phonologically related to a target) supports word learning (Carlson, Sonderegger, & Bane, 2014; Coady & Aslin, 2003; Storkel, 2004, 2009). Although lexical forms in both spoken and signed languages have phonological structure, many are also iconically tied to their meaning (see Perniss, Thompson, & Vigliocco, 2010, for a review). Mounting evidence suggests that iconicity may play a modality-general role in supporting the acquisition of both signed and spoken vocabulary (Perry, Perlman, & Lupyan, 2015; Thompson, Vinson, Woll, & Vigliocco, 2012). In the current study, we examined the effects of iconicity, neighborhood density, and lexical frequency on early vocabulary development in a sign language. We asked whether the prevalence of iconicity in sign language means that iconicity alone is sufficient for children to learn new signs, drowning out any benefit that could be derived from phonological or lexical properties, or whether, in addition to iconicity, the frequency of lexical items and their phonological properties remains a powerful predictor of early language development.

Iconicity in Vocabulary Development

Because of the pervasiveness of iconicity in sign languages (Perniss et al., 2010), some researchers have hypothesized that it should be an especially useful tool for acquiring new signs (Brown, 1978). Surprisingly, most research has found no role for iconicity in the acquisition of sign language. In vocabulary production, children make frequent phonological substitutions that diminish rather than enhance iconic features of signs (Meier, Mauk, Cheek, & Moreland, 2008). Further, less than one third of children’s expressive sign vocabulary is transparently iconic; most signs are either metonymic (i.e., the relationship between form and meaning is not transparent to most observers) or arbitrary (i.e., there is no relationship between form and meaning; Orlansky & Bonvillian, 1984). This corresponds to the rate of iconic signs in the American Sign Language (ASL) lexicon more broadly; only 30% of ASL signs are rated as highly iconic (> 4 on a 7-point scale; Caselli, Sevcikova Sehyr, Cohen-Goldberg, & Emmorey, 2016). Children’s first signs include the signs for mommy, cookie, and milk (Anderson & Reilly, 2002), which vary in degree of iconicity in ASL (see Fig. 1). Yet a recent study countered this early work and showed that iconicity predicted parental reports of native deaf children’s acquisition of British Sign Language (BSL; Thompson et al., 2012).

Fig. 1.

The signs for mommy, cookie, and milk in American Sign Language. These signs have iconicity ratings of 1.2, 3.3, and 4.2, respectively, on a 7-point scale.

Notably, iconicity seems to be an important factor in the development of spoken vocabulary. Adults rate the words that children learn earliest as highly iconic (Perry et al., 2015). Further, Japanese 3-year-olds acquire novel iconic, sound-symbolic verbs more reliably than novel non-sound-symbolic verbs, which suggests that they can leverage the mapping between iconic form and meaning to learn novel verbs (Imai, Kita, Nagumo, & Okada, 2008).

The role of iconicity, however, may shift as children develop. A study of 8- to 30-month-old children learning BSL found that the effect of iconicity became more prominent with age (Thompson et al., 2012). This indicates that older children, between the ages of 2 and 3, may have developed cognitive tools to better leverage the mapping between iconic form and meaning (e.g., Gentner & Namy, 2006; Meier, 1982). Further, systematic sound-meaning relationships are stronger among spoken words frequently acquired between the ages of 2 and 6 than among later-acquired words, which suggests an early use of iconicity that is later sacrificed for communicative efficiency (Monaghan, Shillcock, Christiansen, & Kirby, 2014).

Although the findings of Perry et al. (2015) and Thompson et al. (2012) converge to provide tantalizing evidence that iconicity shapes early vocabulary, regardless of modality, these results must be interpreted with caution because neither study controlled for phonological neighborhood density, which robustly predicts spoken-vocabulary development (Coady & Aslin, 2003; Storkel, 2004, 2009). Eliminating this confound is critical for studying acquisition of sign languages because signed phonology and iconicity are interdependent (e.g., Emmorey, 2014; Taub, 2001) and neighborhood density can be correlated with iconicity (Caselli et al., 2016).

Neighborhood Density in Vocabulary Development

Children learning spoken languages use distributional information about the phonological properties of words to build a lexicon (Saffran, Aslin, & Newport, 1996; Thiessen, 2007). Their early vocabularies contain a greater proportion of high-density words than adults’ vocabularies do (Coady & Aslin, 2003; Storkel, 2004, 2009), and children produce high-density words more accurately than low-density words (Sosa & Stoel-Gammon, 2012). However, the role of neighborhood density is not straightforward. Although dense neighborhoods support the configuration of new lexical representations, words from sparse neighborhoods may be more easily identified as novel, and thus trigger word learning (Storkel & Lee, 2011).

Signed languages differ from spoken languages in three important ways that may make children less likely to unpack phonological information. First, iconicity may be so salient in sign language that it circumvents the usefulness of phonology; that is, iconicity alone may be sufficient for sign learning. Second, unlike spoken phonology, which is based on sequences of sounds, signed phonology often includes simultaneously produced features (e.g., hand shapes, locations, movements; Brentari, 1998). In spoken language, children may be sensitive to distributional information about phonology in order to make predictions about upcoming sounds and word boundaries (Saffran et al., 1996). Young signing children may initially ignore phonological structure and treat signs holistically because the simultaneous production of phonological elements may have less predictive value. Third, the absence of many minimal pairs (Brentari, 1998) means that most signs have relatively few phonological neighbors; sparse neighborhoods may not support the configuration of new lexical representations. Even if children attend to phonological structure (e.g., Siedlecki & Bonvillian, 1998), it remains unclear whether the phonology of most sign languages would itself support sign acquisition. For all of these reasons, one might expect to find no effect of neighborhood density on sign acquisition. Alternatively, because adults treat signs compositionally during language processing (e.g., Baus, Gutiérrez-Sigut, Quer, & Carreiras, 2008; Carreiras, Gutiérrez-Sigut, Baquero, & Corina, 2008; Caselli & Cohen-Goldberg, 2014; Mayberry & Witcher, 2005), children may make use of neighborhood density as well.

Lexical Frequency in Vocabulary Development

Children are sensitive to how frequently words appear in their input; common words are acquired more rapidly than uncommon words (Goodman, Dale, & Li, 2008). To date, the only study examining the role of frequency¹ in sign acquisition found that it had no effect (Thompson et al., 2012) either because iconicity is so useful that frequency has little additive value or because the range of frequency values was insufficient to detect an effect.

The Current Project

We examined the structure of the emerging productive signed lexicon in young children by reanalyzing a data set drawn from parental reports of early ASL vocabulary production (Anderson & Reilly, 2002). Our analysis incorporated information from ASL-LEX, a newly available lexical database for ASL that includes the first publicly available measure of neighborhood density in any signed language, as well as measures of iconicity and lexical frequency, for nearly 1,000 signs (Caselli et al., 2016). Thus, for the first time we were able to examine the individual contributions of iconicity, neighborhood density, and lexical frequency to signed vocabulary development.

We considered three competing predictions. First, we investigated whether signing children relied exclusively on iconicity. This prediction would be supported if iconicity positively affected acquisition even when neighborhood density and frequency were controlled, and neighborhood density and frequency exerted no effect. Second, we tested whether signing children relied only on phonological information. If so, then neighborhood density would positively affect sign acquisition, but iconicity would have no effect. Finally, we asked whether signing children use iconicity, phonology, and lexical frequency to learn new signs. In this case, we expected to find independent effects of these three variables.

Method

Participants

This study included one parental report for each of 64 ASL-exposed deaf children from the original ASL-CDI norming study (Anderson & Reilly, 2002).² The ASL-CDI was modeled and named after the MacArthur Communication Development Inventories (CDI; Fenson et al., 1993). One parent of each child was given a list of glosses of ASL signs and reported whether his or her child used each sign. We excluded children from the original data set if they were not deaf (n = 3) or if we did not know their age (n = 2) or their parents’ hearing status (n = 1). This resulted in a final sample of 58 children (mean age = 24 months, Mdn = 26 months, range = 8–35 months). All the children had a deaf mother, and all but 3 had a deaf father.

Materials

Our analyses included only the 373 (57%) ASL-CDI items for which we had neighborhood density, iconicity, and subjective frequency ratings from the ASL-LEX database (Caselli et al., 2016). A subset of 46 items were repeated on the ASL-CDI forms, and only the data for the first presentation of each of these items were included in our analyses. Because the ASL-CDI does not include reference videos, items were matched to ASL-LEX on the basis of an identical or plausibly matching English gloss (e.g. crocodile was matched with alligator). A handful of ASL-LEX items matched multiple ASL-CDI glosses (e.g., bath matched both bath and bathtub), so the set consisted of 361 unique items. These signs included 231 nouns, 51 verbs, 38 adjectives, 10 adverbs, 2 number words, and 29 function words, as defined by ASL-LEX. Because function words are acquired on a very different trajectory than open-class words (e.g., Storkel, 2004), they were excluded from all analyses. This left a total of 332 items (18,404 trials; see the Supplemental Material for a list of the items). Thus, the final data set was more than 7 times the size of the BSL data set (Thompson et al., 2012). As Anderson and Reilly (2002) found in the original study, vocabulary size and age were correlated in this sample (r_s = .86, p < .001; see Fig. 2).

Fig. 2.

Distribution of vocabulary size by age in the current sample. Vocabulary size was calculated over the 332 items used in the current analyses. The rug plot illustrates the marginal distributions of age and vocabulary size.

Three variables were drawn from ASL-LEX (Caselli et al., 2016). For each, we report summary statistics calculated over the 332 items analyzed. Deaf adult signers rated the subjective frequency of each sign in everyday conversation on a scale from 1 to 7 (M = 4.7, SD = 1.2, minimum = 1.8, maximum = 7.0). Hearing nonsigners rated iconicity (how much each sign looked like its referent) on a scale from 1 to 7 (M = 3.4, SD = 1.8, minimum = 1.0, maximum = 7.0).³ Neighborhood density⁴ was defined as the number of signs that shared at least four of the five following phonological properties: selected fingers, major location, flexion, movement, and sign type (M = 34, SD = 27, minimum = 0, maximum = 118). We chose this definition of neighborhood density, called maximal neighborhood density by Caselli et al. (2016), over the other two neighborhood-density estimates available in ASL-LEX because its implicit definition of neighbor is the most similar to definitions used in the literature on spoken language (i.e., neighbors are words that differ by one phoneme). The phonological coding system used to estimate maximal neighborhood density was based on the prosodic model (Brentari, 1998). As illustrated in Figure 3, the distributions of the CDI items and the non-CDI items in ASL-LEX were similar (Caselli et al., 2016). However, Wilcoxon tests indicated that frequency and iconicity were significantly higher for the CDI items than for the non-CDI items (non-CDI items: mean frequency = 4.07, p < .01; mean iconicity = 3.05, p = .01; mean neighborhood density = 32, p = .17). In the current data set, frequency and neighborhood density were correlated (r_s = .13, p = .015), as were frequency and iconicity (r_s = −.20, p < .001) and neighborhood density and iconicity (r_s = .14, p = .009).

Fig. 3.

Comparison of the lexical properties of the CDI and non-CDI items in the ASL-LEX database. From left to right, the graphs show the distribution of signs according to their subjective frequency, iconicity, and neighborhood density. CDI = ASL Communication Development Inventories (Anderson & Reilly, 2002).

Modeling procedure

We ran a mixed-effects logistic regression because it simultaneously accounted for child-specific and item-specific variability, and allowed for generalization beyond both the sample of children and the set of vocabulary items. The dependent variable was acquisition (can produce the sign = 1, cannot produce the sign = 0). The model included two-way interactions between each variable of interest (neighborhood density, iconicity, and subjective frequency) and age (in months), as well as random effects of participants and items. By-participant random slopes for the three variables of interest were also included. All continuous variables were z-transformed (iconicity and subjective frequency were already on a z scale when imported from ASL-LEX). To address issues of collinearity, we confirmed a significant effect of a variable of interest using a log-likelihood test comparing a model containing that variable with a model excluding. This helped us to confirm that each variable had an effect above and beyond the others despite correlations among the variables. The p values reported are from the log-likelihood comparisons.

Results

There were significant positive main effects of age, β = 2.32, SE = 0.18, χ(1) = 85.08, p < .001; subjective frequency, β = 0.57, SE = 0.20, χ(1) = 8.56, p = .003; iconicity, β = 0.36, SE = 0.15, χ(1) = 6.06, p = .014; and neighborhood density, β = 0.26, SE = 0.12, χ(1) = 4.92, p = .026 (see Fig. 4). In order to confirm the direction and size of the effects, which could have been obfuscated by collinearity, we also ran separate models with each of the lexical variables (and its interaction with age) in turn. The direction and magnitude of the effects were qualitatively the same—frequency: β = 0.52, SE = 0.20, χ(1) = 7.53, p < .01; iconicity: β = 0.30, SE = 0.15, χ(1) = 4.42, p = .036; neighborhood density: β = 0.34, SE = 0.12, χ(1) = 8.57, p < .01. These results suggest that, unsurprisingly, the children knew more words as they got older. Further they were more likely to know signs that were more common, signs that were more iconic, and signs that had more neighbors.

Fig. 4.

Predicted probability of a child having acquired a word given the child’s age (upper left) and the sign’s iconicity (upper right), neighborhood density (lower left), and subjective frequency (lower right). These probabilities were calculated using a link inverse function and reflect marginal effects (i.e., all other predictors in a given model were set to their mean).

There was a marginally significant interaction between age and iconicity, β = −0.10, SE = 0.05, χ(1) = 3.36, p = .067. Visual inspection of this interaction (see Fig. 5) suggests that there may have been a ceiling effect, whereby the oldest children knew most of the words and thus exhibited a smaller effect of iconicity. We investigated the age-by-iconicity interaction by dividing the data into the oldest group (≥ 32 months), the age group that appeared to reach a ceiling, and the youngest group (< 32 months). We ran a model separately for each group and found an effect of iconicity in the younger children, β = 0.39, SE = 0.15, χ(1) = 7.30, p = .007, but not the older children, β = −0.77, SE = 0.61, χ(1) = 1.60, p = .21. This finding must be interpreted with caution, in light of the marginally significant interaction. There was no interaction between age and neighborhood density, β = 0.03, SE = 0.03, z = 0.79, p = .43, or between age and frequency, β = 0.11, SE = 0.07, χ(1) = 2.10, p = .15; these results indicate that neighborhood density and frequency had consistent facilitative effects across development.

Fig. 5.

Illustration of the two-way interaction between age and iconicity. Log odds (logit) of sign acquisition is shown as a function of iconicity, separately for each age quartile and the minimum and maximum ages (z-transformed values; translations into raw ages are presented to facilitate interpretation).

Because ASL-LEX contains two other estimates of neighborhood density, we wanted to validate the choice to use maximal neighborhood density. We compared models using this measure with models using parameter-based neighborhood density (the number of signs that share all of the following phonological properties: selected fingers, flexion, major location, and path movement) and minimal neighborhood density (the number of signs that share at least one of the following phonological properties: selected fingers, flexion, major location, path movement, and sign type). The model using maximal neighborhood density had the best fit, as indicated by the lowest Akaike information criterion (AIC)—maximal neighborhood density: AIC = 13,641.11; parameter-based neighborhood density: AIC = 13,642.34; minimal neighborhood density: AIC = 13,643.54. However, the differences between the models were not significant.

Though the distribution of ages was roughly flat, there was a spike in the number of participants between the ages of 32 and 33 months old (see Fig. 2). To rule out the possibility that this spike had a disproportionate effect on the results, we removed 7 randomly chosen participants from this age range and reran the analyses. The results were qualitatively the same: Iconicity, neighborhood density, and frequency facilitated acquisition, and there was an age-by-iconicity interaction.

To examine whether any of the variables of interest had nonlinear effects on vocabulary acquisition, we ran models that were identical to the one reported but in which age, frequency, neighborhood density, and iconicity were replaced, in turn, with their log, square-root, and multiplicative-inverse transformations. Model comparisons indicated that none of these models outperformed the models with untransformed variables.

We conducted a final Monte Carlo analysis to investigate the possibility that the effects of iconicity, neighborhood density, and frequency stemmed from the distribution of the limited set of signs in the CDI. For each participant, we randomly sampled from all of the CDI items the number of signs that child knew (without replacement) and calculated the average iconicity, neighborhood density, and frequency for this randomly generated vocabulary. We did this 10,000 times for each of the 58 children. The distribution of these vocabularies is illustrated in Figure 6. If the observed data were no different from chance, we would expect that for a given variable, the observed means of 5% (~3) of the children would fall above the 95th percentile of their individual distribution of randomly generated means. In this data set, two-tailed one-proportion z tests with the Yates continuity correction revealed that the number of observed averages that fell above the 95th percentile significantly exceeded this expected rate for each of the variables: 17 of the observed average iconicity ratings (29%; z = 8.19, p < .01), 24 of the observed average neighborhood densities (41%; z = 12.21, p < .01), and 22 of the observed average frequency ratings (38%; z = 12.41, p < .01; see Fig. 6). Thus, the effects of iconicity, neighborhood density, and frequency cannot be attributed to their base rates among the CDI items relative to the rest of the lexicon of ASL; rather, these effects indicate that highly iconic, frequent, and dense signs are overrepresented in children’s early productive vocabularies.

Fig. 6.

Results of the Monte Carlo analysis. The circles show the observed individual-level means for iconicity, frequency, and neighborhood density as a function of the child’s vocabulary size. The vertical lines indicate the 95th percentile of the randomly generated vocabulary for each child. Nontransformed values for frequency and iconicity are plotted for interpretability; the results were similar for the transformed and nontransformed measures.

Discussion

We tested three possible routes to sign acquisition: (a) exclusive reliance on iconicity, (b) exclusive reliance on phonology, and (c) reliance on both iconicity and the statistical features of signs. Our results are consistent with the third possibility. Iconicity promoted productive sign acquisition by native-signing children learning ASL. Thus, our study, with a much larger sample, replicated what Thompson et al. (2012) observed in a study with a different sign language, BSL. The effect of iconicity held even when neighborhood density and frequency were controlled. However, iconicity did not trump other factors that have been robustly shown to predict spoken-vocabulary development. Despite the vast differences in the forms of signs and words, sign-exposed children were sensitive to statistical properties of signs: High phonological neighborhood density and lexical frequency promoted sign acquisition, even after controlling for iconicity. This suggests that the road to vocabulary learning is not exclusively iconic; rather, children also leverage statistical information about lexical items and their phonological properties to learn new signs.

The effect of neighborhood density suggests that signing children track sublexical structure and do not treat signs holistically, despite the simultaneity of phonological features and the prevalence of iconicity in signed words. The facilitative effect of neighborhood density is compatible with the idea that dense signed neighborhoods are sufficient to support configuration of new lexical items, even though signs have relatively fewer neighbors than words do (Brentari, 1998). It is unclear whether neighborhood density of signs also triggers word learning (i.e., whether having many neighbors makes it less obvious that a novel sign needs to be learned; e.g., Storkel & Lee, 2011). Crucially, neighborhood density is related to other phonological properties (e.g., phonotactic probability, phonological complexity), and our work cannot specify which phonological properties children attend to. Thompson et al. (2012) observed an interactive effect of age and phonological complexity on sign production: Children under 20 months of age produced significantly less complex signs than older children did. Effects were restricted to young children’s productive sign vocabulary. With our data set, we observed no significant interaction of age and neighborhood density, which indicates that children make use of phonological properties to expand their productive vocabulary throughout their first years.

The iconicity effect suggests that children make use of iconic mappings to learn new words. This finding replicates the BSL work with deaf 11- to 30-month-olds (Thompson et al., 2010) with two main differences. Although Thompson et al. found differences in the effects of iconicity as a function of age when they compared 11- to 20-month-olds with 21- to 30-month-olds, we did not replicate this interaction. We did, however, find a ceiling effect in children older than 32 months, who were not included in the study by Thompson et al. (2012). The way children leverage iconicity is unclear. Its effect may be driven by signs that have a direct mapping between form and meaning (see DeLoache, Kolstad, & Anderson, 1991, and Gentner & Rattermann, 1991, for work on nonlinguistic analogical mappings). For example, children may be sensitive initially to direct iconic mappings (e.g., the hand shape of the sign hammer shows a hand holding a hammer), but ignore distal mappings (e.g., the hand shape of the sign house distally maps to the angle of the roof; see Magid & Pyers, 2017). Related studies have shown that children prefer to produce iconic gestures and signs that represent actions rather than the perceptual features of the referents and tend to interpret such gestures and signs as actions (Marentette & Nicoladis, 2011; Ortega, Sümer, & Özyürek, 2016).

Although our measure of lexical frequency predicts the acquisition of sign language, it has some limitations. Our frequency estimates were based on adult-directed, not child-directed, signing and may not be the ideal measure of input frequency. In our data set, the earliest acquired signs included bath and ball, which are only moderately frequent in adult signing (4.92 and 4.63, respectively, on a 7-point scale), but are likely more common in child-directed signing. Estimates of frequency in child-directed signing will lead to better understanding of the role of frequency in sign acquisition. Indeed, research on spoken languages shows that measures of frequency in child-directed speech are better predictors of children’s vocabulary acquisition than are measures derived from adult-directed speech (Goodman et al., 2008).

Although we found that native sign-language acquisition and native spoken-language acquisition are affected by similar lexical properties, sign acquisition most commonly occurs among deaf children born to parents who do not know sign language, and thus have delayed exposure to language, signed or spoken (e.g., Spencer & Harris, 2005). It may be the case that children acquiring a sign language as their first language later in childhood rely to a greater degree on iconicity than on phonological properties to learn their first signs. Indeed, hearing adults who learn ASL as a second language are highly sensitive to iconicity (Baus, Carreiras, & Emmorey, 2013; Lieberth & Gamble, 1991), and deaf individuals who are late learners of a sign language do not process sign phonology efficiently (Mayberry & Fischer, 1989). Thus, an avenue for future research will be to investigate whether the mechanisms that drive early signed vocabulary development differ depending on the child’s age of exposure to a sign language.

One central aim of the cross-linguistic study of language acquisition is to determine which properties of language acquisition during childhood are relatively invariant, arising similarly across languages and regardless of modality, and which vary across languages and modalities. This works suggests that despite dramatic differences in phonology and iconicity between signed and spoken languages, signing children, like children learning spoken languages, track statistical information about lexical items and their phonological properties and leverage those features to expand their vocabulary.