Temporal Separation of Meanings Facilitates Homophone Learning

Abstract

Theories of word learning often assume that learners adhere to a one-to-one form-to-meaning mapping hypothesis, reducing the number of possible meanings for newly encountered words and permitting learners to assume that known words correspond to their known referents. However, all languages contain homophones, which violate this assumption, raising the question of how children acquire such words. This article presents two studies evaluating whether the timing of homophone meanings relative to one another might play a role in their acquisition. Study 1 examined homophone uses in speech directed to 13 English-learning children from the CHILDES database and determined that multiple meanings of the same word only very rarely occur within a single discourse. When they do, uses with the same meaning are closer together in time than uses with different meanings, indicating the availability of meaning timing as a cue to homophony. Study 2 used the cross-situational word learning paradigm to manipulate the meaning timing of novel homophones. English-learning preschoolers (N=94) were more likely to learn both meanings of a novel homophone when those meanings were separated in time, as compared to when the meanings were heard in an interleaved fashion. Taken together, these studies demonstrate that segregation of homophone meanings across time may support children’s ability to map multiple meanings to a single word.

1. Introduction

In spite of the high degree of referential ambiguity in natural word learning situations, children learn thousands of words in the first several years of life with minimal difficulty. While theories of word learning diverge on the details of how this remarkable feat is accomplished, many of them posit that an assumption of one-to-one form-to-meaning mapping, sometimes called mutual exclusivity, plays a role (Golinkoff, et al., 1992; Markman & Wachtel, 1988; Slobin, 1973). This assumption would allow learners to constrain their guesses at a word’s meaning when multiple referents are present by excluding referents of already known words from contention, but it would also allow them to assume that when they hear a familiar word, it corresponds to its familiar referent. There is extensive research demonstrating children’s adherence to the first half of the previous statement, and the control conditions from some studies of mutual exclusivity also support the second point, with children selecting familiar referents for known words the vast majority of the time (Golinkoff, et al., 1992; Lewis, et al., 2020; Markman, Wasow & Hanson, 2003). One problem with assuming that a familiar word will always refer to a familiar referent, however, is that it’s not true. Natural languages contain homophones, or words with multiple, unrelated meanings, and to be proficient users of their languages, children need to learn that some words have this property and allow those words to violate the one-to-one form-to meaning assumption, while still applying that constraint to the majority of the words that they encounter.

Prior research on children’s acquisition of homophones is somewhat limited and largely focused on the difficulties that they show when it comes to learning such words in lab settings, which is consistent with the use of a one-to-one mapping bias. Much of this work has used a pseudohomophone paradigm, in which participants were taught nonce second meanings for already familiar words. For example, Mazzocco (1997) taught children and adults nonce words as well as new second meanings for familiar words, embedded in rich story contexts and accompanied by illustrations. While learning of both nonce words and pseudohomophones improved with age, pseudohomophone learning was particularly poor compared to nonce word learning in the youngest age group (3-5 years) and developed more slowly than novel word learning. Doherty (2004) noted that Mazzocco’s (1997) study may have presented an unrealistically difficult challenge for learners because the to-be-learned referents of the pseudohomophones were already known to the learners, meaning that they had to accept a pseudosynonym as well as a pseudohomophone. However, Doherty’s own study (2004) showed no difference in pseudohomophone learning when the referent was a novel object versus when it was a known object. Children showed persistent difficulty assigning a new meaning to a familiar word, indicating that homophony represents a meaningful challenge for word learners, perhaps due to difficulty overriding the one-to-one mapping assumption.

In spite of this apparent bias against learning second meanings for familiar words in lab settings, children’s early vocabularies do contain homophones, indicating that such words are learnable in natural settings (Conwell & Morgan, 2012; Doherty, 2000) and presenting an interesting empirical question: what about the natural settings in which children encounter homophones supports the learning of two meanings for a single word? Work using the pseudohomophone paradigm has provided some evidence for the kinds of information that might support homophone learning. Storkel and Maekawa (2005) demonstrated that repetition of the novel word-to-meaning mapping facilitated pseudohomophone learning in preschool-aged children (see also Conwell, et al., 2022), and work by both Casenhiser (2005) and Dautriche and colleagues (2018) has shown that use in a different grammatical category may also support learning of a second meaning for a known word by preschool-aged children and toddlers, respectively. Dautriche and colleagues (2018) also found some support for improved pseudohomophone learning when the new referent was from a different taxonomic class within the same grammatical category. Thus, some pseudohomophone research has provided evidence that the homophone learning problem may be mitigated by properties of learners’ experience (i.e., repetition) and of the structure of language (i.e., semantic and syntactic differentiation).

One aspect of natural experience with homophones that has not yet been explored empirically is the timing of meaning exposure. Speech to young children is “bursty” in terms of content (Slone, et al., 2023), with many references to a single object occurring in rapid succession before switching to a new topic of discussion. When a word has more than one meaning, it is possible that this tendency toward bursts of discussion about single topics could cause the two meanings to occur in completely different contexts, separated by the topic of discourse and, perhaps, significant amounts of time. Such a separation might facilitate learning second meanings for words by allowing each meaning to be learned without direct interference from the other. Alternatively, homophone meanings may be interleaved or used closely together in time, which could set up opportunities for direct exposure to the homophony problem and thereby enhance metalinguistic awareness of homophony as a phenomenon. To date, there has not been a systematic study of the temporal structure of homophone meanings in children’s natural experience. This is a critical omission because the nature of experience inherently constrains potential learning mechanisms. Therefore, one question that the present paper asks is how homophone meanings are distributed in children’s natural experience and how that distribution might support learning two meanings for a single word.

To examine how meaning timing could affect homophone learning, we also need to ask whether the patterns of use observed in natural experience support homophone learning in a controlled setting. While most prior work on homophone learning in lab settings has used the pseudohomophone paradigm, that paradigm may not reflect children’s natural experience with word learning because it tends to involve ostensive teaching of the new word meanings. Some pseudohomophone studies lack referential ambiguity, presenting scenes in which there is only a single possible novel referent for the pseudohomophone (Conwell, et al., 2022; Storkel & Maekawa, 2005), and still others include explicit pointing at the referent (Dautriche, et al., 2018), which has been shown to boost word learning in lab tasks (Dunn, Frost & Monaghan, 2024). Because word learning biases, such as one-to-one mapping, have been posited as a means of reducing referential ambiguity, a lack of ambiguity in experimental tasks may mean that learners do not need to recruit those biases to succeed in the same way that they might in more natural settings. Beyond the limitations of using ostensive contexts to introduce new meanings, the pseudohomophone paradigm may also not so much test homophone learning as it tests learners’ ability to suppress a well-established meaning for a familiar word. Indeed, success in pseudohomophone studies has been shown to fluctuate with the availability of the known meaning during the test phase (Backscheider, et al., 1999), indicating that the pragmatics of the paradigm may limit the extent to which results generalize to more natural learning situations.

While truly natural word learning is not easily replicated in a lab, some word learning paradigms do present words in non-ostensive, referentially ambiguous settings, resulting in somewhat more realistic learning scenarios. In particular, the cross-situational word learning (CSWL) paradigm presents multiple referents for novel words on each learning trial, which requires learners to track, implicitly or explicitly, uses across situations to resolve referential ambiguity and determine the specific referent of a word (Akhtar & Montague, 1999; Smith & Yu, 2008; Yu & Smith, 2007). Additionally, CSWL studies allow precise control over learners’ experience with word-referent pairings, number of exposures, and other factors that may affect word learning. The CSWL paradigm has been used with participants across a wide range of ages, including infants (Smith & Yu, 2008), young children (Suanda, et al., 2014; Vlach & Johnson, 2013), and adults (Yu & Smith, 2007).

CSWL presents an opportunity to study homophone learning under referential ambiguity, which may more closely reflect how homophones are encountered in natural settings and provide better insight into the processes involved in suppressing the one-to-one word-to-meaning bias. Prior research using the CSWL paradigm with children has asked a variety of questions related to how experience may impact word learning. In a study that pioneered the use of this method with young children, Suanda and colleagues (2014) demonstrated that contextual diversity supported word learning in 5- to 7-year-old children. Additional research with this paradigm has shown that both language ability and more general cognitive mechanisms such as memory affect children’s word learning in CSWL tasks (Vlach & DeBrock, 2017). Timing of exposures to word-meaning pairings has been a frequent topic of examination in this work. Although several studies reported no differences in word learning as a function of how timing of exposures was structured, structured exposure did appear to support better word learning. Being clustered very close together in experience and being separated by only a few learning trials both resulted in better learning than completely random presentation of meanings. These prior results suggest that some structure in exposures to meanings is better than none, perhaps due to memory demands and consolidation (Benitez, Zettersten & Wojcik, 2020; Vlach & DeBrock, 2017; Vlach & Johnson, 2013).

The converging evidence that preschool children benefit from structure in the timing of meaning exposure when learning single meaning novel words in CSWL raises the possibility that homophones may additionally benefit from such structure. Given the need to violate the one-to-one mapping assumption to learn homophones, clustering of exposures to each meaning may facilitate learning by bolstering associations, but interleaving of meanings could also support learning if contrastive uses highlight the nature of homophony. To address how the timing of exposure to homophone meanings affects learning, the second study in this paper used a CSWL design with two learning conditions. In one, the two meanings of a novel homophone are interleaved during the learning phase (similar to the “unstructured” condition in Benitez, et al., 2020) and in the other, they are presented in separate blocks. The results of this controlled study will be compared to the patterns of use in the corpus study to address whether learners’ natural experience with homophone timing facilitates homophone learning.

The two studies presented in this article were intended to address the question of the availability and use of temporal distribution as a potential support for learning multiple meanings for one word. Study 1, a corpus analysis, considered how many homophones children really experience in their daily lives as well as how frequently both meanings of those words were used within a single discourse. Given prior work on the “bursty” nature of children’s exposure to words and meanings, we anticipated that exposure to both meanings of a homophone within a single discourse would be rare in children’s natural experience. We predicted, for example, that it would be rare for a caregiver to use bat in the animal sense in a recording and then, in the same recording, to use bat in the sports equipment sense. We further predicted that when both meanings of a homophone were used in the same recording, those uses with the same meaning would occur closer together in time than homophone uses with different meanings. For example, if a caregiver used bat to refer to an animal, we would expect to encounter another use that referred to an animal sooner than we would encounter a use that referred to sports equipment. Study 2, an experimental study using the CSWL paradigm, examined whether spacing or interleaving the meanings of novel homophones better supported 4- and 5-year-old children’s ability to learn two meanings for one word. We predicted that spacing homophone meanings during the learning task would improve the learning of two meanings for a single novel word.

2. Study 1

This study asked about the temporal distribution of meanings of homophones in children’s natural experience. Specifically, we used existing corpus data to examine whether homophone meanings were typically separated in time or whether children were exposed to those meanings in an interleaved fashion. We focused on speech to children under the age of 5 years to ensure that children in the study had not yet started formal education and that, therefore, their experience with homophones would be uninfluenced by differences in spelling.

2.1. Method

2.1.1. Corpora

We examined the child-directed speech from five North American English corpora in the CHILDES database (MacWhinney, 2000; version 2020.1): Bloom (Bloom, Hood & Lightbown, 1974), Braunwald (1971), Providence (Demuth, Culbertson & Alter, 2006), Sachs (1983), and Weist (Weist & Zevenbergen, 2008). The Bloom corpus included three children, the Providence and Weist corpora had six children each, and the Braunwald and Sachs corpora both had one child. These corpora were selected because they contained longitudinal recordings on a regular or quasi-regular schedule and included audio recordings with timestamped transcripts, which would allow us to analyze the relative timing of word meanings. Two children from the Bloom corpus (Eric and Gia) and two from the Weist corpus (Matt and Roman) were excluded from analysis because of missing timestamps in a significant portion of their transcripts, bringing the final sample to 13 target children. Additionally, 736 individual transcripts from the remaining 13 children were excluded for missing timestamps. The final sample included 631 total transcripts. Table 1 shows the ages and the total number of recordings that were included in the final sample for each child.

Table 1.

Age ranges, number of recordings, and recording duration information for transcripts included in the analysis for each child. Ages are given as years; months and durations are in minutes

Corpus	Child	Ages	Total Recordings	Mean Duration (S.D.)	Duration Range (Min-Max)
Bloom	Peter	1;0–3;0	10	217.05 (74.37)	48.38-273.8
Braunwald	Laura	1;5–4;10	158	15.67 (13.4)	0.85-100.53
Providence	Alex	1;4–3;5	51	57.53 (3.68)	49.9-68
Providence	Ethan	0;11–2;11	50	54.72 (10.65)	7.28-63.6
Providence	Lily	1;1–4;0	80	56.98 (5.88)	23.45-67.33
Providence	Naima	0;11–3;10	88	63.68 (13.55)	22.57-82.5
Providence	Violet	1;2–3;11	51	43.1 (7.13)	27.22-59.93
Providence	William	1;4–3;4	44	53.23 (8.7)	20.17-61.98
Sachs	Naomi	1;8–4;9	54	24.75 (12.62)	4.05-64.5
Weist	Benjamin	2;4–3;3	9	29.03 (4.03)	18.57-31.38
Weist	Emily	2;6–3;0	2	36.88 (14.8)	26.4-47.35
Weist	Emma	2;8–4;3	15	32.27 (7.5)	18.2-46.95
Weist	Jillian	2;1–2;10	19	31.18 (0.3)	30.52-31.67

2.1.2. Target Words

A list of 111 ambiguous target words was generated prior to beginning the analysis. To control for possible effects of lexical category, only homophones with at least two, unrelated noun meanings were included. Words were selected if they had two or more noun meanings in the CELEX Database (Baayen, Piepenbrock & Gulikers, 1995) or if they appeared in various literacy resources for early childhood and elementary educators, including picture books about homophones such as Dear Deer (Barretta, 2007) and websites for teachers (e.g., Common Homophone List, 2022). For each target word, a list of potential noun uses was assembled based on CELEX definitions. All potential noun usages of these words were included, resulting in some words with more than two possible meanings (e.g., trunk: tree, elephant, suitcase, car). Some words (e.g., date) had unrelated homophonous meanings (fruit vs. calendar), as well as multiple related meanings (event vs. specific day). Because we were only interested in the spacing of unrelated, homophonous meanings, all related (i.e., polysemous) meanings of a word were annotated with the same broad tag. One word (bow) had some uses that were homographic, but not homophonic, meaning that they had the same spelling but different pronunciations (i.e., /boʊ/ vs /baʊ/) depending on intended meaning. As both pronunciations were homophones, uses for the two pronunciations were analyzed separately to ensure all meanings were truly homophonous. In the event that a homophone was used with a noun meaning that had not been identified as a potential meaning prior to beginning the coding process, the token was tagged with that meaning after confirming that it was a commonly accepted use and not a nonce use of the word.

2.1.3. Procedure

To determine the distribution of homophone meanings in speech to children, we first annotated all uses of the target for intended meaning. All utterances in the selected corpora that contained one or more of the target words were extracted using the childesr package in R (Braginsky, Sanchez, & Yurovsky, 2022). A total of 46,694 tokens of the target words were extracted through this first pass, including tokens produced by the children themselves, which were subsequently removed, as this analysis was only interested in the properties of child-directed speech. To limit the analyses to input sources that children encountered regularly in normal day-to-day experience, the data were also filtered to include only speech from parents, grandparents, or siblings, as opposed to experimenters, tradespeople, etc., who are only in the home occasionally or for very brief periods of time. Exclusion of speech from these adults resulted in the removal of 2,883 tokens. The final dataset included 27,267 tokens for analysis. Eight of the target words had been previously annotated for meaning as part of the ChiSense-12 corpus (Cabiddu, et al., 2022) and the annotations from that corpus were used for this analysis. Heterographic homophones (e.g., yolk/yoke) were annotated based on spelling, but also reviewed to exclude any non-noun uses. Six trained undergraduate research assistants annotated each token of the remaining words by hand. All research assistants had English as their first and primary language. They all completed a brief training on the annotation process in which they were provided with sample tokens, annotated by one of the authors, with explanations of the annotations. They also received written step-by-step instructions for annotation along with the list of possible codes.

The annotation process involved three stages. First, we attempted to identify meanings based solely on the utterance itself. If the homophone’s intended meaning could be determined unambiguously from the content of the utterance alone, that token was tagged with the intended meaning. When a word was used multiple times in a single utterance, each token was individually tagged for its respective meaning. A total of 14,465 tokens were identified at this stage as non-noun, compound noun, or proper noun uses and excluded from further analysis. If a meaning could not be disambiguated based on the utterance context alone, that token was held out for the second stage of annotation. A total of 11,558 noun uses were annotated for meaning on the basis of the utterance content alone.

Uses of target words that could not be definitively annotated using the utterance content alone were re-evaluated in the second stage of annotation. For these tokens, the immediate conversational context, defined as two lines before and two lines after the utterance, was extracted using the childespy library on Python 3.12 (Sanchez et al., 2019) and used as additional disambiguating information. The code for this extraction is available on OSF (https://osf.io/zef2n/overview?view_only=661b1f2d06904cd3a84e10d16b65ca54). Trained undergraduates again hand-coded as many of these uses as they could using only the immediate conversational context. A total of 659 tokens were disambiguated using these surrounding utterances.

Finally, for uses that were still ambiguous given the surrounding transcribed utterances, annotators used the video playback feature on CLAN (MacWhinney, 2000) to disambiguate the meanings of the target words from the Providence corpus (the only corpus with available video recordings). Because Cabiddu and colleagues (2022) did not use extra-linguistic context in their annotation process, tokens of target words from the ChiSense-12 dataset that had been labeled as ambiguous were also re-annotated using video recordings where available. If the intended meaning could not be disambiguated with video, that token was determined to be ambiguous in context and excluded from further analysis. A total of 311 tokens were disambiguated using the video contexts, leaving 274 tokens unannotated due to unresolved ambiguity. The final sample included 12,528 noun tokens of the target homophones annotated for their intended meanings.

To assess annotator reliability, 10% of the tokens of each target word were re-tagged by a second annotator. Words with fewer than 10 total tokens were pooled and 10% of the tokens from that pool were also re-annotated. Disagreements between the first and second coders of each token were resolved by a third independent coder. Inter-annotator agreement for meaning was 95%.

2.1.4. Processing of Annotated Data

To address the question of how homophone meanings are distributed across time, co-occurrence within a transcript and timing of meanings within a discourse were extracted from the dataset. Co-occurrence frequency was defined as the number of transcripts in which two or more noun meanings of a target word occurred. Temporal distribution of competing meanings within a speech context was assessed using only those cases in which more than one meaning of a target word was used in a single transcript. Within a transcript, every homophone use after the first one was coded as same or different, depending on whether its meaning matched or did not match the previous usage, and the distance between neighboring homophones was calculated in seconds. Because timestamps in CHILDES indicate the beginnings and ends of utterances, not the specific timing of each word, the time of the target word production was defined as the midpoint of the utterance. In the corpora, overlapping utterances were typically contained within a single timestamp tied to the first utterance in the sequence of overlap. In these cases, the timestamps of the first utterance in the overlapping sequence were used for all uses within a sequence.

2.2. Results

Four of the original 111 target words (creak/creek, floe/flow, miner/minor, sacks/sax) had no noun uses in any of the corpora and were excluded from further analysis. The remaining 107 target words had 226 potential meanings, 52 of which were never attested in the corpora. Of the 107 target words with noun tokens in the corpora, 45 (42.1%) had only one noun meaning in the dataset and could, therefore, provide no further information about the distribution of homophone meanings relative to one another. Fifty-eight words (54.2%) had two noun meanings attested in these corpora, three words (2.8%) had three noun meanings, and one (0.9%) had four attested noun meanings. A list of all words and the number of attested meanings in the corpora can be found in the supplemental materials on OSF (https://osf.io/zef2n/files/4mn93). The presence of more than one noun meaning in the corpus for a majority of the target homophones indicates that children likely hear both meanings of at least some noun-noun homophones in their natural experience (see also Conwell, 2017), and raises the question of how these meanings are distributed.

First, we asked how often multiple meanings of homophones occurred in the same recording session. The lengths of the recording sessions varied, but few were longer than 1 hour, making a recording session a meaningful first pass at relative temporal proximity for the meanings. Of the 62 homophone types for which multiple meanings were attested in the corpora, only 25 (40.3%) had at least one instance of distinct meanings co-occurring within the same transcript. A complete list of these words and the number of times that they co-occurred within a transcript can be found in the Supplemental Materials on OSF (https://osf.io/zef2n/files/wz6by). Furthermore, none of the transcripts contained more than two meanings for a homophone. This suggests that, broadly, homophone meanings do not tend to co-occur in children’s natural experience. Of the 25 whose meanings did co-occur in at least one transcript, 11 did so only once in the 631 transcripts included in the analysis, further indicating that homophone meaning co-occurrence is rare. Of the remaining 14 types, 5 had two meanings co-occur in 2 transcripts, 5 had 3 transcripts with two meanings, and only 4 had co-occurrences in five or more transcripts (2 in 5, 1 in 11, and 1 in 14). The words that most commonly co-occurred with their homophonic counterpart were flour/flower (5), knight/night (5), letter (11), and son/sun (14).

While co-occurrence of homophone meanings within a single transcript was unusual, there were enough cases to warrant further investigation into whether meanings within each of these discourses occurred in a blocked or interleaved fashion. That is, were like meanings clustered together? To address this question, we considered only those transcripts in which multiple meanings occurred. We asked whether two sequential uses with the same meaning (n=236) were closer to one another than two sequential uses with different meanings (n=94). The average distance between same meanings was 128.6 seconds and the average distance between different meanings was 626.86 seconds, which is a statistically significant difference (t(104.5)=6.26, p<0.001, two-tailed). This indicates that even in the rare cases where more than one meaning of a homophone was used within a single hour of child-directed speech, different meanings were used farther apart in time than were sequential uses of the same meaning. Figure 1 presents a visual representation of the timing of uses from three representative examples.

Figure 1.

Three examples of the distribution of competing homophone meanings within recording sessions. Each vertical bar represents a use of a particular meaning and those that appear thicker represent several uses in quick succession. Top example from Naima, age 1;10 (transcript ID = 42421); middle from Lily, age 1;8 (transcript ID = 42321); bottom from Lily, age 2;10 (transcript ID = 42366).

The findings of this corpus analysis supported the hypothesis that homophone meanings are separated in time. Not only did two meanings of a homophone very rarely occur in the same recording session, when they did, they were farther apart in time than when a single meaning of a homophone was used more than once. These results likely reflect discourse effects on word use and the tendency for homophone meanings to be very distinct from one another. While it is possible to devise hypothetical scenarios in which one might, for example, need to talk about both a flying mammal and a piece of sports equipment (i.e., a bat), such situations are unusual in the natural experience of children. This suggests that separation in time may be a critical cue that allows children to permit exceptions to the one-to-one mapping assumption and learn homophones from their natural experience.

Although separation of homophone meanings in time likely reflects differences in the discourse and referential contexts in which those meanings are used, these differences could impact children’s learning of homophones via two distinct routes. First, semantic differences between meanings that result in separation in time likely also result in separation in terms of reference and discourse contexts. In other words, children may be cued to the existence of a second meaning for a familiar word when the familiar referent is absent from the visual and/or discourse context. On the other hand, children should not assume that a familiar word has a new meaning any time its familiar referent is absent; people frequently speak about things that are not present. Additionally, the findings of Study 1 cannot bear on this point because we did not code for differences in the visual availability of referents, particularly for the availability of the alternate referent.

A second possibility is that the separation in time alone could facilitate homophone learning because it would allow for better memory consolidation and reduced interference between meanings. That is, if a learner has the opportunity to get multiple exposures to a word with one meaning without encountering its homophonic counterpart, they may have a more solid representation of that word-referent pairing that would be less susceptible to interference from the other meaning. Alternatively, the temporal structure of natural exposure to homophones could hinder children’s homophone learning by reducing access to contrastive uses that might highlight the existence of two meanings for one word. To examine how meaning timing affects homophone learning, we designed an experimental study in which we controlled for referential context to look exclusively at the potential role of meaning timing in this process.

3. Study 2

The results of Study 1 provided evidence for the availability of temporal separation of meanings as a potential cue to homophony in children’s natural experience, but the mere availability of a cue to homophony does not indicate that the cue will facilitate learning multiple meanings for a single word. For example, homophone meanings in natural speech have been shown to differ in their token duration as a function of the frequency of the intended meaning (Conwell, 2017; Gahl, 2008) and children themselves have been shown to produce these differences in an elicitation paradigm (Conwell, 2018). However, such differences have not been found to support children’s homophone learning or the disambiguation of intended meanings by either children or adults (Conwell, et al., 2022; Conwell, et al., 2023). Therefore, having established that homophone meanings are separated in time in children’s natural language experience, we must also assess whether that separation supports the learning of two meanings for the same word. Prior research with adult participants has provided some support for this hypothesis (Kachergis, Yu, & Shiffrin, 2012), but whether children show the same patterns as adults remains an open question. Using a cross-situational word learning paradigm, we asked whether separating two meanings for a novel word across time in the experiment would facilitate learning of both meanings when referential context differences are fully controlled for.

3.1. Method

3.1.1. Participants

Participants in this study were 94 monolingual English-learning children between 4 and 6 years of age (41 female; 53 male; mean age 4;9; age range 4;0-6;0). An additional 6 participants completed the study, but had their data excluded from analysis due to significant bilingualism (2 female), selecting the same location on every test trial (2 female; 1 male), or a very high degree of distraction throughout the study (1 male). Another six participants (3 female; 3 male) began the study, but chose to end their participation before the learning phase was completed. Seventy of the children in the final sample participated in a lab setting and were recruited through community networks and events, ads in publications aimed at parents, and messages to listservs. The other 24 participants in the final sample were recruited through their childcare centers and participated in a quiet room in the childcare. A parent or guardian provided written consent for all children to participate, and children provided verbal assent, which they could withdraw at any time. Caregivers were compensated $10 for participating. All methods were approved by the [anonymized] Internal Review Board.

The sample size was determined by a power analysis for a two-condition between-subjects design conducted in G*Power (Faul, et al., 2007), based on an effect size of d=0.52 (the smallest effect size reported in Suanda, et al., 2014) with a target power of 0.8, resulting in an overall sample size of 94 participants. Participants were randomly assigned to a learning condition: 47 (23 female; 24 male) in the spaced condition and 47 (18 female; 29 male) in the interleaved condition. The two groups did not significantly differ in age (μ_spaced=56.98 months, μ_interleaved=56.81 months, t(92)=0.119, p=0.906).

3.1.2. Stimuli

The audio stimuli for this study were 10 disyllabic nonce words consistent with English phonotactics (e.g., daxen, japple), 2 of which were used for the warm-up phase, and 8 of which were to be learned during the experiment. These stimuli were recorded by an adult female native speaker of North American English and presented in isolation.

The visual stimuli for this study were 14 pictures of plush animals that were unlikely to be familiar to young children from the North Midwestern United States (e.g., echidna, okapi). Norming data from a previous study indicated that these animals were unfamiliar to adults from this region (Kuznia & Conwell, 2026). Two of these images were used in the warm-up phase and 12 were used as referents for the to-be-learned nonce words. The images of the plush animals were taken from various sources on the internet. All of them had been photographed against a white background and were re-sized to be 300 x 300 pixels. All stimuli for this study can be found on OSF (https://osf.io/zef2n/overview?view_only=661b1f2d06904cd3a84e10d16b65ca54).

3.1.3. Design

This study used a between-subjects design to test the effects of meaning spacing on the learning of novel homophones in a cross-situational word learning paradigm. Participants were randomly assigned to one of two experimental learning conditions: spaced and interleaved. These conditions differed only in terms of the order in which learning trials were presented. Each study session consisted of 3 phases: warm-up, learning, and test. These phases were based on the procedure described in Suanda, et al. (2014). In the warm-up phase, participants saw an array of all of the visual stimuli at once and heard each of the novel words once with a 500 ms pause between each word. Children were told that these were the new words that they were going to learn. Then, they were introduced to two practice words. On these warm-up learning trials, word-referent pairs were presented one at a time. That is, children saw only one image and heard only one word. They encountered each pairing twice, for a total of four practice learning trials. Then, they saw a practice testing phase that was identical to the one they would encounter at the end of the study, but with only two trials. On each test trial, four animals were presented at the left, right, top, and bottom of the screen, and a single word was played. Children were instructed to point to the animal that went with the word they heard. During the warm-up phase, they were given feedback on whether their choice was accurate to ensure that they understood the task.

Following the warm-up phase, children moved directly to the learning phase. On each learning trial, participants saw two animals, one on the right and one on the left, and heard two novel words with a 1s interstimulus interval. An example of the visual layout of a learning trial is presented in the left half of Figure 2. Four of the novel words were consistently paired with one referent each time they appeared on a learning trial. These were the single meaning words. The other four novel words were paired with one referent on half of the learning trials and a different referent on the other half. These were the dual meaning words. All word-referent pairings occurred four times during learning, which meant that dual meaning words were heard a total of eight times, four with each referent, while single meaning words were each heard four times. There was a total of 24 learning trials.

Figure 2.

Sample trials from the learning phase (left) and the testing phase (right) of Study 2.

All children experienced the same learning trials with the same word-referent pairings. The only difference between conditions was the order of presentation. Children in the spaced condition heard only one meaning for each dual meaning word in the first half of the learning phase, and heard only the other meaning in the second half of the learning phase. They also experienced only half of the single meaning words in the first half of learning, and the other half of those words in the second half of learning. Which specific word-meaning pairs occurred in the first or second half of learning was counterbalanced across participants in the spaced condition. Within each half of the learning phase, the trials themselves were presented in a random order. Children in the interleaved condition encountered all of the same word-referent pairings in the learning phase, but they experienced those trials in a random order.

The testing phase consisted of 12 test trials with the same format as the test trials from the warm-up phase. Participants saw four of the animals from the learning phase on the screen and heard one word. They were asked to point to the animal that went with the word. An example of the visual layout of a test trial is presented in the right half of Figure 2. Test trials were constructed such that dual meaning words never appeared with more than one of their referents and all referents occurred as distractor objects an equal number of times. Target location was balanced across test trials and all test trials were identical in both learning conditions.

3.1.4. Procedure

After caregivers consented to participation, participants were told that they would be learning the names of some animals that they probably didn’t know and asked if that was something that they wanted to do. Upon providing verbal assent, they were led to the testing space. In-lab participants were typically accompanied in the testing space by one experimenter and their caregiver, while childcare center participants were accompanied by two experimenters.

The in-lab testing space was a small, quiet room with a child-height table and chairs. Participants were seated at this table in front of a laptop computer with a 17-inch screen, on which the visual stimuli were displayed. Audio stimuli were played over the built-in laptop speakers. The experimenter sat to the side of the laptop during all training trials and did not look at the screen to avoid accidentally providing eye-gaze information about the intended target. Caregivers were seated behind the child, out of their direct line of sight, again to avoid potential interference. In both of the childcare settings, children were tested in small, quiet rooms adjacent to their primary classrooms. These rooms both had child-height tables and chairs, which were arranged as closely as possible to the way the in-lab room was arranged and the same laptop computer was used to present the stimuli.

The warm-up phase was explained by the experimenter and completed with feedback about accuracy of responses. Then, the child was informed that they would see two animals at a time and hear two words. Children were told to look and listen carefully and to try to decide which words went with each animal, but they were not instructed to repeat words or point to objects. (Some participants did these things spontaneously on some trials.) They were also told that there would be lots of animals to see and to be patient. No feedback was provided during the learning phase and if a participant asked questions or asked for confirmation of a meaning, the experimenter replied, “I don’t know. Let’s keep watching.” To maintain motivation, participants were told that they were doing well halfway through the learning phase, and they were also informed that there were “just a few more” after approximately 20 of the 24 learning trials had been completed.

After completing the learning phase, participants were praised for their patience and told that it was time to find out how many of the animal names they knew. They were reminded verbally that they would see four animals at a time and hear one word and that they should point to the animal that went with the word. Unclear pointing was resolved by asking the child to put their finger “right on” the animal. Continued ambiguity in responding (i.e., a refusal to put their finger directly on a referent) was recorded as a non-response and coded as incorrect. Responses were noted on paper forms by the experimenter. At no time during this testing phase were children given direct feedback about the accuracy of their selections. If a child changed their response, the second response was used in the analysis, although this only occurred five times across all participants. A refusal to respond or an indication that the participant did not believe the referent was present (“there’s no japple”) was recorded as a non-response and coded as incorrect. The entire procedure took approximately 8 minutes for each participant.

3.2. Results

Overall, participants correctly identified the intended referent on 38.9% of all test trials, indicating that this was a very challenging task for the child participants. Participants in the spaced condition correctly identified 42.0% of all referents, while participants in the interleaved condition correctly identified 35.8% of them. All analyses described below were conducted using R (R Core Team, 2022) and mixed models were fit using the lme4 package (Bates, et al., 2015). The data and all analysis code can be found on OSF (https://osf.io/zef2n/overview?view_only=661b1f2d06904cd3a84e10d16b65ca54).

The primary question that this experiment was designed to address was whether timing of meanings of dual meaning words would affect children’s ability to learn more than one meaning for those words. To this end, we filtered the data to include only the dual meaning words and fit a Poisson regression with number of meanings learned (0, 1, 2) as the outcome variable and learning condition (spaced vs. interleaved; dummy coded; reference level: interleaved) as a fixed effect. We also wanted to consider the possibility that age was a meaningful predictor of the ability to learn more than one meaning for a homophone, so age in months (mean-centered) was included as a fixed effect, as was an interaction between age and learning condition. Models containing random effects for participant and random slopes for target word resulted in singular fits, so we report the maximal model that did not have a singular fit, which included only random intercepts by target word. Those results are presented in Table 2. There was a significant effect of learning condition on how many meanings of a dual meaning word participants learned (β=0.265, p=0.032), indicating that separation of word meanings in time facilitated the learning of two meanings for a novel homophone. There was also a significant effect of age on the number of meanings learned for the dual meaning words, with older children showing a slight disadvantage (β=−0.026, p=0.041), but there was no evidence of a significant interaction between the fixed effects (β=−0.007, p=0.694). The average number of meanings learned by condition and word type are presented in Figure 3.

Table 2.

Results of a Poisson mixed effects model of number of word meanings learned by timing condition. The interleaved condition served as the reference level for learning condition.

Fixed Effects	Estimate	Standard Error	95% CI (lower)	95% CI (upper)	z
Intercept	−0.484	0.104	−0.717	−0.269	−4.673***
Learning Condition	0.265	0.124	0.023	0.510	2.139*
Age (in months)	−0.026	0.013	−0.050	−0.001	−2.045*
Condition x Age	0.007	0.017	−0.027	0.041	0.394
Random Effects	Variance	Standard Error
Word	.008	0.044

^*p<0.05,

^**p<0.01,

^***p<0.001

Figure 3.

Word learning results for Study 2. Error bars represent standard error of the mean.

While this analysis was specific to dual meaning words, it may be the case that children in the spaced condition learned both meanings for more dual meaning words because they simply learned better overall. Single meaning words were also more clustered for children in the spaced condition than for those in the interleaved condition, who experienced all word-meaning pairs in a random order. To determine whether learning overall differed by learning condition, we fit a logistic mixed effects model including all words with accuracy as the outcome variable and condition (spaced vs. interleaved; dummy coded; reference level: interleaved) as the fixed effect. Models including random slopes resulted in singular fits, so we report the maximal model that converged without singular fit, which included random intercepts for participant and for target word. Complete results of this model are presented in Table 3. The effect of condition on accuracy was marginally significant (β=0.285, p=0.072), suggesting that segregating word meanings may have improved learning overall, including for single meaning words. However, this result must be interpreted cautiously, as it did not reach statistical significance.

Table 3.

Results of a model of accuracy on all words by learning condition.

Fixed Effects	Estimate	Standard Error	95% CI (lower)	95% CI (upper)	z
Intercept	−0.580	0.194	−0.998	−0.165	−2.988**
Learning Condition	0.285	0.159	−0.028	0.603	1.797
Random Effects	Variance	Standard Error
Participant	0.210	0.047
Word	.0195	0.056

^**p<0.01

Another question that arises from these data is whether there is a cost to learning any meaning of a word when those meanings are in conflict with one another, as suggested by the lower overall accuracy in the interleaved learning condition. That is, are dual meaning words harder to learn in the interleaved condition because the two meanings are interfering with one another. To assess whether there was a penalty to learning even one word meaning when a word had a homophone and whether this was affected by learning condition, data were coded based on whether any meaning of the word was learned. Single meaning words were coded as 1 if their referent was correctly identified during the final test, while dual meaning words were coded as 1 if either or both referents were correctly identified on the final test. Then, we fit a logistic mixed effects regression with word type (single meaning vs. dual meaning; reference: single meaning) and learning condition (spaced vs. interleaved; reference: interleaved) as fixed effects, including the interaction of these variables. Both fixed effects were dummy coded. Learning of at least one meaning was the outcome variable. Models including random slopes resulted in singular fits, so we report the maximal model that did not result in a singular fit, which only included by-participant and by-item random intercepts. The complete output of this model is in Table 4.

Table 4.

Results of a logit mixed effects model of learning of at least one word meaning by learning condition and word type.

Fixed Effects	Estimate	Standard Error	95% CI (lower)	95% CI (upper)	z
Intercept	0.092	0.265	−0.473	0.660	0.348
Word Type	−0.310	0.361	−1.095	0.464	−0.859
Learning Condition	0.572	0.243	0.096	1.057	2.352*
Word Type x Learning Condition	−0.595	0.310	−1.206	0.012	−1.919
Random Effects	Variance	Standard Error
Participant	0.252	0.052
Word	0.167	0.144

^*p<0.05

While there was no evidence for a difference in the likelihood of learning at least one meaning for a word as a function of whether the word had one meaning or two (β=−0.310, p=0.390), there was a significant main effect of learning condition, with participants in the spaced condition showing higher rates of learning overall (β=0.572, p=0.019). Additionally, there was a marginal interaction of learning condition and number of word meanings (β=−0.595, p=0.055), which indicated that participants in the spaced condition showed lower performance on the single meaning words than the participants in the interleaved condition. This interaction should be interpreted cautiously because it is only marginally significant and because this analysis was exploratory and may have been underpowered for the interaction. However, the interaction may have been the result of a primacy or recency effect because the single meaning words in the spaced condition were also clustered across time. To evaluate this possibility, we looked at only the participants in the spaced condition and asked whether they performed better with words that they encountered in the first half of learning or in the second half of learning and whether that interacted with the number of meanings a word had. We fit a logistic mixed effects model with accuracy as the outcome variable and word type (single vs. dual meaning; reference: single) and learning block (first vs. second; reference: first), as well as the interaction of those two, as fixed effects (both dummy coded). A fully maximal model resulted in singular fit, so we report the maximal model that converged without singular fit, which included random intercepts for target word and by-participant random slopes for learning block. The complete results of this model are presented in Table 5. There was no evidence of a significant main effect of word type (β=−0.226, p=0.597), suggesting that children did not have more or less difficulty learning a word’s meaning as a function of whether it was a single or dual meaning word. However, there was a marginal main effect of learning block (β=−0.444, p=0.061) with words in the second half of the learning phase showing reduced accuracy. Additionally, there was a statistically significant interaction of word type and learning block (β=0.766, p=0.046).

Table 5.

Results of a logit mixed effects model of accuracy by learning block and number of potential word meanings.

Fixed Effects	Estimate	Standard Error	95% CI (lower)	95% CI (upper)	z
Intercept	−0.188	0.291	−0.830	0.451	−0.647
Learning Block	−0.444	0.237	−0.922	0.026	−1.872
Word Type	−0.226	0.426	−1.168	0.687	−0.529
Learning Block x Word Type	0.766	0.384	0.014	1.527	1.993*
Random Effects	Variance	Standard Error
Participant	0.364	0.088
Participant x Second Block	0.324	0.083
Word	0.213	0.163

^*p<0.05

To further explore this significant interaction, we used the emmeans packages (Lenth & Piaskowski, 2026) in R to estimate marginal means for each condition and to conduct simple effects analyses. The marginal means are presented in Table 6. The simple effects analyses revealed no statistically significant contrasts, although the simple effect of learning block on dual meaning words was marginal (z=1.872, p=0.061; all other z<1.275, all other p>0.2). Children performed marginally better on the dual meaning words that they encountered in the first half of learning, which might mean that dual meaning words in the second half of learning were subject to interference from previously learned meanings, consistent with the mutual exclusivity constraint. While suggestive, this must be interpreted carefully. The sample size for this study was determined based on the condition-level comparison, not for any analysis involving an interaction, which means that it may be underpowered. Further work on this topic will require a larger sample to meaningfully examine how timing of meaning exposure and interference affect homophone learning.

Table 6.

Estimated marginal means for the learning block by word type interaction. Estimates are odds ratios.

Learning Block	Word Type	Estimate	Standard Error	95% CI (lower)	95% CI (upper)
First	Single meaning	0.398	0.080	0.255	0.561
Second	Single meaning	0.477	0.081	0.326	0.632
First	Dual meaning	0.453	0.072	0.319	0.595
Second	Dual meaning	0.347	0.065	0.232	0.483

Overall, the findings of Study 2 suggest that learning more than one meaning for a novel word in a cross-situational word learning task presented a significant challenge for child participants. Most children were unable to learn both meanings of any dual meaning word. This reinforces the idea that homophones represent especially hard cases in word learning, particularly in lab settings. However, these results also provide support for our hypothesis that clustering homophone exposures by meaning, even in a very brief task, improves children’s ability to assign two meanings to fully novel words, as indicated by the statistically significant difference in the number of dual word meanings learned between conditions. Given that children’s natural experience with homophones consists of clustered meanings, these data support the hypothesis that spacing across time facilitates children’s learning of multiple meanings for a single word. The exploratory analyses related to whether learning in general was impacted by experiencing a particular word or meaning in the first or second block of learning indicate that single meaning words may have benefited from a recency effect, while dual meaning words may have been subject to interference from the already-learned meaning. This provides some preliminary evidence regarding how separation of meanings in time impacts the learning of homophones, but additional research specifically targeting interference effects is needed to fully understand that process.

4. Discussion

The two studies in this article examined the issue of timing of meanings in homophone learning by young children. Study 1 asked how homophone meanings are distributed in children’s natural experience and demonstrated that multiple meanings for a single word only rarely occur in the same discourse and, when they do, different meanings of a homophone are farther apart in time than matching meanings of a homophone. Study 2 asked whether separation in time supported children’s ability to map two meanings to a single word in a lab task and demonstrated that, indeed, blocking participants’ exposure to each meaning increased the number of meanings that children were able to learn for dual meaning words. This work adds to the literature on word learning by providing evidence for a cue to homophone learning that is both present in children’s natural experience and apparently used by children in a controlled setting.

The finding that homophone meanings tend to be separated in time is consistent with a number of prior studies on the distribution of referents in natural speech to children. Natural discourse does not jump from topic to topic randomly, but rather dwells on a specific topic before moving on to the next, which would favor clustered uses of words. Prior research on single meaning words in child-directed speech has consistently demonstrated that word uses are “bursty,” characterized by periods of dense usage interspersed with long stretches without any tokens (Sloane, et al., 2023), so it is unsurprising that the unrelated meanings of noun-noun homophones only rarely occur in the same discourse. While one previous study demonstrated that adults produced metalinguistic disambiguating information when using homophones around young children (Kohn & Landau, 1990), that study used an elicitation paradigm in which the homophone meanings were both contextually available rather than data from natural speech. Study 1 provides the first clear empirical demonstration that children’s experience with homophones typically involves separation of the meanings across time.

Separation of meanings across time in natural experience to some extent conflates two possible supports for homophone learning. Homophone meanings are unrelated, which suggests that they are separated in time because they are used in different referential and discourse contexts. Therefore, it is possible that any benefit to learning might be based on occurrence in different referential contexts rather than mere separation in time. Study 2 partially addressed this possibility by using the same learning trials in both timing conditions, thus manipulating the timing of meanings, but not the referential contexts in which they were experienced. Referential context was controlled for in Study 2 and children still showed better learning of both meanings of homophones when those meanings were separated in time. The benefit of temporal separation of meanings for learning in the CSWL paradigm suggests that children can acquire a word that violates the one-to-one mapping constraint by learning one meaning first and then incorporating a second meaning, but that they have difficulty maintaining two distinct, partially learned meanings. These results indicate that temporal separation of homophone meanings likely bolstered homophone learning because it allowed for stronger associations between the word and each of its referents, not because of referential context differences providing semantic differentiation.

This is further supported by the finding that children in the spaced condition showed marginally better learning of the homophone meanings that were introduced in the first learning block. These words could be, effectively, treated as single meaning words when that first meaning was being learned; homophony only had to be accommodated after one meaning had been established. Knowing one meaning well may have given children the opportunity to recognize the absence of the familiar meaning from the referential context in the second half of learning, producing an effective cue to the existence of two meanings. However, this effect must be interpreted with caution, as the effect of learning block was not the original intent of this study and the interaction analysis may have been underpowered. That this was possible in the very short learning period with only a few exposures to either word-referent pair and approximately 3 minutes of separation between the meanings during the task further bolsters the potential utility of timing of meanings as a cue to homophony in natural experience, where uses are typically much farther apart. Of course, in real life experience, meaning timing is likely confounded with referential context differences. Future work should examine how these factors interact in controlled learning settings.

Another consideration in the interpretation of the results of Study 2 is whether the difficulty that children experienced in learning overall may have been due to the high degree of similarity of the visual referents. All of the visual stimuli were plush animals, so they were all of the same semantic class, but many of them were also a similar grayish-brown color and most of them were quadruped mammals. Holding the semantic class constant was intentional because we wanted to isolate timing as a potential cue to homophony, and taxonomic category has also been proposed as a potential cue to support homophone learning (Dautriche, et al., 2018). However, the high degree of similarity between some of these animals may have resulted in visual interference effects that reduced learning overall. Indeed, some data from CSWL with adults has shown that using novel artifacts, such as those in the NOUN database (Horst & Hout, 2016), as referents resulted in higher levels of learning as compared to using the animal stimuli, potentially due to easier discriminability (Conwell & Machida, in preparation). Future work could further examine effects of visual distinctiveness in CSWL, both generally and in the context of homophone learning.

Study 2 was designed to use the CSWL paradigm as a tool for investigating homophone learning, rather than to evaluate the mechanisms underlying CSWL itself. However, these findings may bear on how CSWL proceeds in children, particularly because prior research on CSWL has found conflicting effects of timing on word learning. The design of the interleaved condition in Study 2 most closely replicates the “unstructured” condition and the spaced condition is a hybrid of the “interleaved” and “massed” conditions of Benitez and colleagues’ (2020) study. That study found poor learning overall when the timing of learning trials was unstructured and improved learning when exposures were relatively close together in time, regardless of whether they were separated by only 0-2 trials or by 2-4 trials. The findings of Study 2 add to the growing body of evidence that children benefit from clustered experience with word-referent pairings, perhaps due to their more limited memory capacities (Benitez, et al., 2020; Vlach & DeBrock, 2017).

The studies in this article present evidence that timing of homophone meanings is both available to children in their natural experience and used by children in a lab-based learning task. Timing of meanings is not the only potential cue to homophony that may be both available to and used by children. Cross-linguistic corpus research on adult-directed language suggests that homophone meanings tend to be semantically and syntactically distinct in ways that might support learning (Casenhiser, 2005; Dautriche, et al., 2018), although there is not yet direct evidence that the homophone meanings that children experience follow this pattern and learning from those cues has only been tested in ostensive pseudohomophone studies. Whether such information is utilized by child learners when referential contexts are ambiguous remains an open question. Future work could additionally address whether cues such as semantic or syntactic distinctiveness function in tandem with timing information to further facilitate homophone learning. Children’s natural language experience is exceptionally rich with information that could support word learning, but it is also rife with ambiguity at all levels. Expanding our understanding of how all of this information interacts will allow us to expand our theories of word learning to capture words like homophones, which do not fit neatly into simple accounts, but which children learn anyway.