A Changing Role for Transitional Probabilities in Word Learning During the Transition to Toddlerhood?

Abstract

Infants’ sensitivity to transitional probabilities (TPs) supports language development by facilitating mapping high-TP (HTP) words to meaning, at least up to 18 months of age. Here we tested whether this HTP advantage holds as lexical development progresses, and infants become better at forming word-referent mappings. Two groups of 24-month-olds (N=64 and all White, tested in the USA) first listened to Italian sentences containing HTP and low-TP (LTP) words. We then used HTP and LTP words, and sequences that violated these statistics, in a mapping task. Infants learned HTP and LTP words equally well. They also learned LTP violations as well as LTP words, but learned HTP words better than HTP violations. Thus, by 2 years of age sensitivity to TPs does not lead to an HTP advantage but rather to poor mapping of violations of HTP word forms.

Spoken language unfolds relatively continuously, with few transparent cues to word boundaries, making word segmentation a challenge for novice language learners. Infants and adults readily track the distribution of co-occurrence relationships across syllables, such as transitional probabilities (TPs; Saffran et al., 1996a, b). TPs tend to dip at word boundaries, and thus provide information about which syllable sequences form a word, and which span a word boundary (Saksida et al., 2017; Swingley, 1999; 2005). There is also evidence that TPs play a facilitative role in learning word meanings, at least in 17-month-olds (Graf Estes et al., 2007; Hay et al., 2011), such that high-TP (HTP) words are more readily mapped to referents than words with lower statistical coherence, or low-TPs (LTPs). Infants’ sensitivity to statistical structure is not likely to play a static role in language learning (Forest et. al., 2023), and thus in the current study we asked whether the role that TPs play in word learning changes around the transition to toddlerhood, as their language skills improve. Before introducing our specific research questions, we discuss what is currently known about this developmental process, and what is yet unknown.

Transitional Probabilities and Lexical Development in Infants

In a seminal study, Saffran and colleagues (1996a) showed that 8-month-old infants are sensitive to the frequency with which syllables co-occur in continuous speech. They did so by familiarizing infants with a simple artificial language that contained four trisyllabic HTP sequences whose syllables always co-occurred, leading to TPs of 1.0. The four HTP sequences themselves, however, were randomly ordered, so that the TP across syllables spanning the boundaries of HTP sequences was only .33. By using synthesized speech, the authors were able to remove potentially confounding cues to word boundaries (e.g., pauses, syllable lengthening, stress). Infants were subsequently able to discriminate between the HTP sequences (referred to as “words” in the original study) and unfamiliar sequences (syllables from the corpus that had never co-occurred, which had a TP of 0). They were also able to discriminate between HTP sequences and less frequent syllable sequences with relatively low TPs of .33, or those that crossed the boundaries of an HTP sequence. These sequences were referred to as “part-words” in the original work, but we refer to them as LTP part-sequences here. Subsequent studies have revealed that infants can also differentiate between equally frequent HTP sequences and LTP part-sequences based on their TPs (i.e., Aslin et al., 1998). For clarity, Table 1 contains a breakdown of the kinds of items used in several seminal studies investigating TP learning, as well as the terminology we use for these items in this paper.

Table 1.

Type of items used in studies investigating TP learning

Test Item Type	Definition	Examples of Familiarization Stimuli and Tests Items	Language Material Type & Citation
HTP sequence	Syllable sequences that always co-occur during familiarization	babupu bupadadutabapatubipidabututibu e.g., babupu, TP = 1.0	Artificial Language Safran, Newport & Aslin (1996)
LTP part-sequence	Syllables sequences that span HTP sequence boundaries	babupubupadadutabapatubipidabututibu e.g., pidata, TP = ~.33	Artificial Language Safran, Newport & Aslin (1996)
Unfamiliar sequence	Syllables sequence composed of syllables from the corpus that never co-occurred	babupubupadadutabapatubipidabututibu e.g., pubati, TP = 0	Artificial Language Safran, Newport & Aslin (1996)
Unfamiliar word	Words and syllables that compose the words are not presented during familiarization	Torno a casa con le bici cariche di frutta in bilico sulla sella. La zia Carola si è esibita in una fuga colla bici verde. e.g., FUga, MElo, CAsa, BIci, TP = 0	Natural Language Hay, Pelucchi, Estes & Saffran (2011)
HTP word	Syllable sequences that always occur during familiarization, and never occur anywhere else in the corpus	Spesso Lisa capita in FUga nella CAsa dove giaci gracile e tesa. Se cadi col la BIci prima del bivio del MElo cavo ti do dieci bigoli e una biro. e.g., FUga, MElo, TP = 1.0	Natural Language Hay, Pelucchi, Estes & Saffran (2011)
LTP word	Syllables that co-occur regularly, but which also occur in many other words throughout the corpus	Spesso Lisa capita in FUga nella CAsa dove giaci gracile e tesa. Se cadi col la BIci prima del bivio del MElo cavo ti do dieci bigoli e una biro. e.g., CAsa, BIci, TP = .33	Natural Language Hay, Pelucchi, Estes & Saffran (2011)
Violated-HTP word	Words formed by combining the first syllable of one HTP word with the second syllable of the other HTP word	Spesso Lisa capita in FUga nella CAsa dove giaci gracile e tesa. Se cadi col la BIci prima del bivio del MElo cavo ti do dieci bigoli e una biro. e.g., FUlo, MEga, TP = 0	Natural Language Current Paper
Violated-LTP word	Words formed by combining the first syllable of one LTP word with the second syllable of the other LTP word	Spesso Lisa capita in FUga nella CAsa dove giaci gracile e tesa. Se cadi col la BIci prima del bivio del MElo cavo ti do dieci bigoli e una biro. e.g., CAci, BIsa, TP = 0	Natural Language Current Paper

Note. To facilitate comparison across studies investigating TP learning that have used different terminology, we have created a set of terms that highlight the features relevant for the current work. In studies using artificial language materials, syllable sequences with high TPs are referred to as HTP sequences, and sequences that span the boundaries of an HTP sequence are referred to as LTP part-sequences. Note that these LTP part-sequences violate the parsing suggested by HTPs. Sequences that contain unfamiliar syllables, and also sequences that contain syllables that had never co-occurred, are referred to as Unfamiliar sequences. In studies using natural language materials, there are HTP words, referred to as such because they are real words in the language. Likewise, the LTP words are real words, but they have lower TPs. Critically, in these studies, the LTP words do not contain any of the same syllables as HTP words, and thus they do not break up the HTP words as the LTP part-sequences used in the artificial language studies do. There are also Unfamiliar words, which are words that had never occurred, nor had their component syllables. Highlighted lines show the item types used in the current study.

These initial studies used relatively simple synthesized speech materials which consisted solely of four HTP sequences that were repeated numerous times across a few minutes, with no phonemic or prosodic variability. However, subsequent studies have shown that 8-month-old infants can track TPs when listening to an unfamiliar natural language. For example, English-learning infants can discriminate between real Italian words based on their TPs (i.e., distinguishing between HTP and LTP words) once familiarized to a set of naturally-spoken sentences (Pelucchi et al., 2009a, b; Karaman & Hay, 2018; see Table 1). Note that in these studies using Italian, the HTP and LTP words do not contain any overlapping syllables, and thus the LTP words were not formed from parts of HTP words. Instead, the LTP words were actual words. These words have LTPs because their syllables also occurred in other, untested words rather than because they span word boundaries (see Table 1; the Appendix also contains examples of these Italian stimuli). In contrast, in the artificial language studies the LTP part-sequences typically consist of the final syllable(s) of one HTP word followed by the initial syllable(s) of another, thereby violating the parsing suggested by the TPs. This difference in language design, and specifically between the LTP words used in the Italian materials and the LTP part-sequences used in the artificial languages, will become relevant in a later section.

Altogether these findings, which have been replicated in younger infants (e.g., Fló et al., 2019), as well as with infants learning languages other than American English (e.g., Hoareau et al., 2019), suggest that infants are able to track TPs relevant for word segmentation. There is also evidence that infants’ ability to track TPs relevant to segmentation is related to their native language development. For example, infants can discriminate between syllable sequences that frequently vs. infrequently co-occur in their native language (Ngon et al., 2013). Moreover, several studies have linked infants’ ability to track TPs in artificial languages to their vocabulary size and lexical processing efficiency (e.g., Frost et al., 2020; Hoareau et al., 2019; Lany et al., 2018).

In addition to correlational evidence that tracking TPs in speech supports lexical development, there is evidence from experimental tasks that TPs influence how readily infants learn speech sequences as labels in mapping tasks (Graf Estes et al., 2007; Hay et al., 2011). For example, in Hay et al. (2011) 17-month-olds were familiarized with Italian speech containing 2 HTP and 2 LTP words (see Table 1). They were next habituated to pairings between either the 2 HTP words and referents, or the 2 LTP words and referents (e.g., an isolated token of the HTP melo was paired with one object, and a token of fuga was paired with another). Infants were then tested using the Switch paradigm, such that the word-object pairings were switched during the test (Stager & Werker, 1997). Hay et al. found an “HTP advantage”, with only HTP words learned as labels. This effect appears to reflect an advantage for mapping HTP sequences relative to mapping sequences that are less coherent or familiar, because in the same paradigm, 17-month-olds also failed to map unfamiliar Italian words to referents (i.e., they failed to map words that did not appear in the familiarization corpus; Hay et al., 2011). In other words, infants mapped HTP words, but did not map either familiar LTP words or totally unfamiliar words. If the “HTP advantage” were actually an “LTP disadvantage”, then LTP words should have been learned worse than both HTP words and unfamiliar words. In an artificial language paradigm, Graf Estes et al. (2007) also found that infants are better able to map HTP sequences than LTP part-sequences or unfamiliar words, consistent with the existence of an HTP advantage at this age.

A plausible interpretation of these findings is that at 17 months of age, strong TPs provide important support in a word-referent mapping task because high statistical coherence leads to better encoding and recognition of word forms. LTP words, lacking this coherence, might be more poorly encoded or recognized, which would make it difficult to learn an association between that word form and a referent. This would lead to the observed pattern of successfully mapping HTP labels to referents, and failure map LTP labels to referents. This interpretation is consistent with evidence that 18-month-olds are better able to learn word-referent mappings when the word forms themselves are more familiar (e.g., Swingley, 2007).

Transitional Probabilities and Lexical Development in Children and Adults

TPs are highly imperfect cues to word boundaries in natural languages (Saksida et al., 2017), and there is substantial evidence that language learners rely heavily on other, language-specific cues, such as stress, prosody, and phonotactics, to segment native language speech by adulthood (Gout et al., 2004; Mattys et al., 2005). Nonetheless, in the same year as their seminal infant study was published, Saffran et al. (1996b) provided evidence that adults can track TPs in a similar artificial language, discriminating HTP sequences from unfamiliar sequences and LTP part-sequences. The finding that adults can track TPs in continuous speech has been widely replicated (e.g., Frank et al., 2010; Kurumada et al., 2013; Hartshorne et al., 2019). Moreover, in one study adults and 6-year-old children showed equivalent sensitivity to TPs when trained and tested using identical materials and procedures (Saffran et al., 1997), discriminating HTP sequences from unfamiliar sequences. Likewise, Raviv and Arnon (2016) found that sensitivity to TPs across syllables in an artificial language is relatively constant across 6 to 12 years of age. Altogether, these findings suggest that humans can track TPs in novel speech sequences across the lifespan.

Given that infants, children, and adults can track TPs in speech, it may be tempting to assume that the HTP advantage in word learning also holds across development. For example, it is plausible that sequences with high statistical coherence would always have an advantage over those with lower coherence, even in relatively skilled word learners who excel at forming arbitrary mappings, due to being more predictable, easier to process, or to being represented more strongly. However, few studies have directly tested how TPs affect mapping speech sequences to referents beyond infancy. Critically, Mirman et al. (2008) found that adults familiarized with an artificial language did not subsequently learn HTP sequences as object labels better than novel sequences that were composed of unfamiliar syllables. Adults were nonetheless affected by the TPs in their familiarization, as they learned mappings between LTP part-sequences and referents more slowly than both HTP sequences and novel sequences. These results suggest that by adulthood, learners may no longer benefit noticeably from HTPs when forming word-object mappings. Instead, sequences that are inconsistent with groupings suggested by HTPs (e.g., LTP part-sequences which contain the last two syllables of an HTP sequence and the first of another) may be disadvantaged, or resistant to mapping.

There is indirect evidence that the HTP advantage may start to diminish as infants approach age 2, or as they transition to toddlerhood. This evidence comes largely from several studies testing whether TPs in naturally spoken unfamiliar speech influence word learning. As described above, English-learning 17-month-olds benefit from HTPs when learning words, such that they map HTP Italian words to objects better than they map equally frequent LTP words or unfamiliar words (Hay et al., 2011). In a study using the same Italian language materials, Shoaib et al. (2018) found that the advantage of HTP over LTP words no longer held at 20 months of age. The design of Shoaib et al. was related to that of Hay et al., but was more challenging, in that infants were trained on more words and were tested using a non-habituation-based task traditionally used to assess word learning in older infants and toddlers. The infants in Shoaib et al. did not learn the HTP mappings better than the LTP mappings. However, infants also showed little evidence of having learned any mappings at all, potentially because they struggled to learn 4 words simultaneously (Graf-Estes et al., 2011). Thus, although these infants did not show an HTP advantage, it is not clear that they showed the adult-like pattern from Mirman et al., (2008), in which word forms are readily learned as labels regardless of their TPs, because they did not learn HTP or LTP words.

Additional indirect evidence that the role of TPs in word learning changes across development can be found in a related study that focused on 24-month-old infants’ ability to remember HTP and LTP mappings across a brief delay (Karaman et al., under review). In that study, infants were familiarized to the Italian corpus from Shoaib et al. (2018) and Hay et al., (2011), and given a mapping task that was designed to facilitate forming word-referent mappings. Specifically, during the mapping phase, the HTP and LTP words were embedded in English sentence frames (“Look, it’s a melo!”), and interspersed with trials using familiar English words (e.g., “Look, it’s a shoe!”). These are some of the most common sentence contexts in which nouns occur in English child-directed-speech (Cameron-Faulkner et al., 2003). Thus, using the Italian words in these English sentences during the mapping phase provided infants with “referential support”, or cues that the HTP and LTP words were intended to serve as labels. Several studies suggest that providing referential support makes the mapping task easier for infants, even supporting them in mapping word forms that contain non-native phonemes like clicks or phonotactic violations (Bijeljac-Babic et al., 2009; May & Werker, 2014; Vukatana et al., 2016). Indeed, when the 24-month-olds in Karaman et al. were trained on the word-referent mappings immediately after familiarization with the Italian corpus and tested on those mappings after a 10-minute delay, they showed evidence of having mapped both HTP and LTP words, and to a similar degree. This pattern of results could suggest that the HTP advantage has diminished by 24 months. However, because Karaman et al. tested how well the HTP and LTP words were learned after a delay, it is possible that there were differences in how well HTP and LTP words were learned, but that those differences diminished across the delay. Specifically, mappings between HTP words and referents may have been better recognized than those between LTP words and referents in an immediate test (if one had been given), even if they performed equivalently on HTP and LTP words after a delay.

The Current Study

To summarize what is known about the effects of TPs on word-referent mapping across development, at 17 months there is an HTP advantage, such that infants map HTP words and sequences better than LTP words, LTP part-sequences, and unfamiliar words. For these novice word learners, who can struggle to encode and remember word forms, HTP sequences may be more robustly represented than LTP sequences or unfamiliar ones. This could lead only HTP sequences to be successfully mapped to referents. By adulthood, however, there is no evidence that HTP sequences are learned better than unfamiliar ones. Instead, adults are slow to map items that contain syllable sequences that are inconsistent with the parsing suggested by transitional probabilities in their familiarization (e.g., items that contain the final syllable of one HTP sequence and the initial syllable of a different HTP sequence; Mirman et al., 2008; see Table 1). It is possible that the HTP advantage is fading by age two, but the evidence for this is indirect, and there is no evidence regarding whether resistance to mapping HTP violations has taken hold by this age.

Determining whether and when this shift occurs is theoretically important, as it will yield a better understanding of how sensitivity to TPs shapes the word learning process. In the domain of word-form segmentation, infants’ sensitivity to TPs in speech appears to play a changing role across development. For example, 7-month-old infants appear to rely more strongly on TPs than on language-specific cues, such as lexical stress relevant for segmentation, but by 9 months they rely more on language-specific cues such as lexical stress (Thiessen & Saffran, 2003; Hay & Saffran, 2012). In fact, infants can learn language-specific cues, such as which phonemes are likely to occur at word onsets, by first segmenting words using TPs (Sahni et al., 2010). These results suggest that relatively domain-general statistical learning mechanisms interact with domain-specific language processing mechanisms, thus providing a window onto how language-learning mechanisms become tuned to the specific language being heard, at least in the domain of word segmentation. We suggest that the role of sensitivity to TPs in word-referent mapping is also likely to change across development, and infants become better able to rapidly form and remember mappings between potential word forms and their referents. Specifically, we suggest that TPs will not lead to a strong HTP advantage, but will instead influence whether word forms are resistant to mapping.

Thus, in the current study we tested the role of TPs in word-referent mapping at 24-months of age, and specifically whether infants fail to show an HTP advantage in a word-learning task (i.e., successfully mapping both HTP and LTP words, and equally well). We also tested whether they fail to map sequences that violate HTPs. To that end, we familiarized two groups of English-learning infants to naturally spoken Italian sentences containing equally frequent HTP and LTP words, (i.e., casa, bici, fuga, and melo). One group of infants, those in the Intact-TPs condition, was tested on how well they learned to map the HTP and LTP words from their familiarization. We can compare their performance on HTP and LTP words to test whether 24-month-olds learn HTP words better than LTP words, as younger infants presented with these materials do (Hay et al., 2011), or whether this HTP advantage has begun to fade. Note that the LTP words are not violations of attested HTP sequences in these language materials, and thus this comparison tells us whether words with relatively high TPs are learned better than words with relatively low TPs.

Infants in a second group, the Violated-TPs condition, were familiarized to the identical Italian sentences, and then trained and immediately tested on their ability to map sequences that involved violations of the HTP and LTP words. In these sequences, the syllables from the HTP and LTP words were recombined such that their TPs were zero. For example, if the HTP words were casa and bici, the violated-HTP words were caci and bisa. The two LTP violations were formed in the same way (e.g., fuga and melo became fulo and mega). This allowed us to test whether infants learned the mappings between HTP-violations more poorly than they learned the HTP words, similar to the adults in Mirman et al. (2008) who resisted learning LTP part-sequences.

Note that we did not assess mapping of HTP words and HTP violations in the same infants because training on casa and bici (the HTP words), as well as on caci and bisa (the HTP violations), would lead to substantial syllable overlap across the word forms we directly taught them, and therefore potentially to unwanted confusability or competition among them. This could lead both HTP word and HTP violations to be learned very poorly. We instead assessed the extent to which violations of HTP words were learned in the Violated-TPs group, and were specifically interested in the between-participant comparison between the HTP and violated-HTP words. Importantly, the Familiarization phase was identical for infants in the Intact- and Violated-TP conditions, and the Referent Training phase was nearly identical: The same syllables were used in the labels heard in both groups, but for infants in the Intact-TPs condition, the syllables were combined into familiar sequences, namely the HTP and LTP words they had heard during familiarization. In contrast, for infants in the Violated-TPs condition, these sequences of familiar syllables were combined into sequences they had never heard before. Thus, for infants in the Violated-TPs condition we intended for there to be a potential conflict between the word forms they heard during Familiarization (e.g., casa, bici) and Referent Training (e.g., caci, bisa).

Based on the findings reviewed above (Karaman et al., under review; Mirman et al., 2008; Shoaib et al., 2018), we predicted that 24-month-olds would show an adult-like pattern. More specifically, we predicted that infants would map HTP and LTP word forms to referents equally well. We also predicted that they would resist mapping sequences that contain violations of HTP words, relative to HTP words. We considered two alternative predictions for performance on the LTP violations. On the one hand, infants might fail to learn all word forms that contain TP violations, in which case we would see a main effect of group (Intact vs. Violated TPs). On the other hand, in comparison to LTP words, HTP words may be easier to encode or have stronger representations due to their greater internal coherence, and learning of violated-HTP words may therefore suffer more. In this case, we would predict an interaction, with infants learning HTP words better than HTP violations, but learning LTP words and LTP violations equally well.

Note that this design tests a fuller range of item-types than is typically tested in studies of infant statistical learning. For example, many studies have examined just two of the item types presented in Table 1, such as HTP and LTP part-sequences (which are violations of the HTPs). Critically, including a larger range of items is necessary to evaluate the extent to which infants both show an HTP advantage relative to actual words that have LTPs, and resist learning sequences that represent part of a HTP word. Specifically, we can test whether infants show an HTP advantage by testing whether they are better able to map words with high statistical coherence to referents than words with low coherence, but which nonetheless do not represent part of an HTP word (i.e., LTP words). We can also test whether infants resist learning sequences that violate HTP word forms by testing whether HTP violations are learned more poorly than HTP word forms. Note, however, that the advantage of including all four kinds of sequences is not just methodological. In natural languages, some words have higher TPs than others, like our HTP and LTP words. Likewise, infants also encounter sequences that contain the same syllables as word forms that they have already encountered, like our HTP and LTP violations.

Finally, we tested the hypothesis that changes in the role of TPs in word learning are related to vocabulary size. We hypothesized that the HTP advantage diminishes as infants build their vocabularies. We reasoned that the infants with better word learning skills (and larger vocabularies) would be less dependent on strong TPs to map words to referents successfully. We also tested the more exploratory hypothesis that infants with larger vocabularies would show greater resistance to learning HTP violations than infants with smaller vocabularies. Resistance to learning word forms as labels based on phonotactic patterns is linked to vocabulary size (Graf Estes et al., 2011; May & Werker, 2014) such that the more words an infant knows the stronger their sensitivity to the native-language phonotactic patterns is likely to be. This sensitivity may lead infants with larger vocabularies to be more resistant to learning new words that contain violations of those patterns (Storkel, 2001). Although knowing more English words is not likely to lead infants to be more attuned to violations of the TPs of HTP Italian words, it may be that the inhibitory mechanisms that lead to resistance to learning some word forms depend on developments in lexical organization. For example, Mani and Plunkett (2011) suggest that the emergence of inhibitory effects of lexical neighbors on recognition tasks across 18 to 24 months of age may be due to rapid increases in the size of the lexicon; whereas before 18 months there are too few words in infants’ vocabularies to create a competitive environment. Thus, we reasoned that resistance to learning word forms containing violations of HTP Italian words may likewise increase as infants develop larger lexicons.

Method

All of the procedures used in this experiment complied with the ethical standards for the treatment of human research participants from the American Psychological Association, and were approved by the IRB of the institution at which data were collected.

Participants

Sixty-four 22- to 24-month-old English-learning monolingual infants (Mage = 22.91 months, range = 22.01 – 24.18, 35 females, 29 males; all White) participated in the study. Infants were randomly assigned to one of the two conditions: Intact TPs or Violated TPs. Within each condition, infants were assigned to one of two counterbalanced languages, described below. All infants had a gestational age of greater than 36 weeks, and had no hearing or vision problems, according to parental reports. Infants were recruited through a participant database maintained in the Department of Psychology at the host university. Our sample size was informed by effect sizes reported in previous work investigating infants’ ability to map high TP sequences to novel objects (Hay et al., 2011; d = .64), and 85% power. This yielded a minimum recommended sample size of 24 infants. We chose to increase the sample size by 33% because our task was more demanding; here infants were taught 4 novel label-object associations instead of 2 and we used the Looking-While-Listening (LWL) procedure (Fernald et al., 2008) instead of the Switch Paradigm (Werker et al., 1998). Data from 40 additional infants were not included in the analysis due to: fussiness (11), failure to provide usable data on at least half of both HTP and LTP trials (14), not paying attention (7), parental interference (4), and experimental error (4).

Italian Language Materials

Familiarization Phase

In the familiarization phase, all infants listened to a corpus of 12 grammatically-correct and semantically-meaningful Italian sentences that were produced by a native female speaker. There were two counterbalanced languages, which were re-recordings of the same sentences comprising Languages 2A and 2B from Hay et al., 2011 (see Appendix for the full set of familiarization materials). The speaker produced the sentences in a lively manner. Four disyllabic trochaic (strong-weak) target words: CAsa, BIci, FUga, and MElo were embedded in the sentences (we use capitalized letters to mark stressed syllables). Each of the target words appeared equally often in the corpus, but their internal TPs differed. The HTP words (TP=1.0) contained syllables that never appeared elsewhere in the corpus. Both the first (stressed) syllables and the second (unstressed) syllables of the LTP words (TP = .33) appeared 12 additional times throughout the corpus (e.g., the stressed syllable of the LTP target words was also stressed in all additional occurrences, and likewise for the unstressed syllables).

Each target word appeared 6 times across the 12 sentences and each sentence was repeated 3 times during familiarization, which lasted 2 min 30 s. Thus, infants heard each target word 18 times throughout familiarization. To avoid systematic bias in the results due to arbitrary preferences for particular target words based on specific phonemes or phonotactic regularities (including the potential influence of English TPs), we used two counterbalanced languages. Specifically, words that were HTP in Language A were LTP in Language B and vice versa.

Although the Italian corpus likely sounds non-native to English-learning infants (Mehler et al., 1988; Shoaib et al., 2018), it shares some key similarities with English: all of the Italian target words were trochees (i.e., strong-weak stress, a pattern that is common in English), and all of the target words were phonotactically legal in English. These features make it likely that infants will be able to track the syllable-level TPs, and evidence from prior studies with these materials suggests that 8- to 24-month-olds do so (e.g., Hay et al., 2011; Karaman et al., under review; Pelucchi et al., 2009a; Shoaib et al., 2018).

Referent Training and Test Phases

Intact-TP Condition.

During the referent training and test phases, intact target labels (i.e., the HTP and LTP words CAsa, BIci, FUga, and MElo) were presented both in isolation and embedded in English sentences (e.g., referent training → “See the [target]! It’s a [target]! [Target]!”; test → “Find the [target]! [Target]! Do you see it?”, or “Where’s the [target]? [Target]! Do you like it?”). The referent training phase also included four trials containing familiar English words (baby, doggie, shoe, and book) to orient the infants to the format of the label-object mapping task. Target labels were matched in intensity (~65 dB SPL) and approximate length (range 750 ms to 850 ms). These phrases were produced in a lively manner in accented English by the same native Italian speaker who produced the familiarization corpus materials.

The training and test trials incorporated colorful images of novel objects, matched in size and brightness, for use as referents (see Figure 1). Four images of the familiar objects (i.e., a baby, doggie, shoe, and book) were used in familiar-word trials. In referent training trials, the image of a single object moved slowly across a white box that appeared on either the bottom right or bottom left corner of the display screen (object visual angle ~10°). The movement of the objects was not synchronized with the timing of object labeling. On test trials, two stationary objects were presented in a white box in the bottom right and left corners of the display screen (visual angle between objects ~30°). Objects were yoked based on trial type. For example, in HTP test trials, the two objects that appeared together on the display screen had both previously been paired with HTP words. The first syllable of the word was always sufficient to determine which object was referred to, as the 2 HTP words had different onsets (e.g., CAsa and BIci), as did the 2 LTP words (e.g., FUga and MElo). Familiar objects were yoked based on animacy (shoe-book, baby-doggie).

Figure 1.

The Experimental procedure involved 3 phases; 1) Familiarization, in which infants listened to the Italian corpus containing HTP and LTP words; 2) Training, in which infants were training on mappings between wither words tight Intact or Violate TPs and referents; and 3) Test, in which infants were tested on how well they had learned the trained mappings using the Looking-While-Listening procedure.

Violated-TP Condition.

The referent training and test stimuli contained the same novel objects used in the Intact-TP condition, but the labels we used (Violated-HTP and Violated-LTP labels) were modifications of the HTP and LTP words. Specifically, violated-HTP labels were created by pairing the first syllable of each HTP word with the second syllable of the other HTP word (e.g., CAsa/BIci → CAci/BIsa). The same process was used to create the violated-LTP labels (e.g., FUga/MElo → FUlo/MEga). Thus, the TP for both violated-HTP and violated-LTP labels was 0 (i.e., infants had never heard those syllable combinations). However, all of the words (intact- and violated-TPs) followed a strong-weak stress pattern, and the syllables used in the HTP words and violated-HTP words were equally frequent, as was the case for the syllables that comprised the LTP words and LTP-violations.

Infants were trained on mappings between these violated-TP labels and objects, and then tested on how well they learned them using the identical procedure as that in the Intact-TP conditions (described below). Importantly, the first syllable of the violated-HTP label unambiguously indicated the referent (e.g., CAci vs BIsa), and the same was true for violated-LTP words (e.g., FUlo and MEga). Note that if infants simply mapped the first syllable of a HTP word from Familiarization to a referent in this phase, they should learn the Intact and Violated-HTP words equally well, because both contain initial syllables consistent with the HTP words from Familiarization.

Procedure

The procedures were identical in the Intact-TP and Violated-TP conditions. There were three phases: Familiarization, Referent Training, and Test. Infants were first familiarized with the Italian corpus containing HTP and LTP words, and then were immediately trained and tested on mappings between words and objects. During all three phases, infants were seated on their caregiver’s lap approximately 1 m from a 42-inch flat-screen television, which was used to present the visual stimuli. Caregivers listened to masking music via headphones throughout the experiment. During Familiarization, infants in both the Intact- and Violated-TP conditions were presented with the Italian sentences, with language version counterbalanced across participants. To maintain infants’ interest throughout this phase, they were shown a silent cartoon video (Winnie-the Pooh) on the monitor.

During the Referent Training phase, infants were presented with four novel label-object pairs (2 HTP and 2 LTP labels in the Intact-TP condition, or 2 violated-HTP and 2 violated-LTP labels in the Violated-TP condition) and four familiar word-object pairs. On each referent training trial, a single moving object was paired with a corresponding label. On a given trial, labels were presented 4 times each, twice in a carrier phrase, and twice in isolation. There were four trials for each of the novel label-object pairs, for a total of 16 referent training trials. Each of the four familiar word-object pairs was also presented once, for a total of 20 training trials. The training phase began with two familiar label-object trials. The remaining 18 training trials were presented in one of four quasi-random orders.

Infants’ word learning performance was tested using a Looking-While-Listening (LWL) procedure (Fernald et al., 2008). On each test trial, two stationary objects appeared on the screen for 500 ms before the onset of an English carrier phrase, in which either a familiar or a novel object label was embedded. This target label occurred 2 seconds after trial onset. An additional repetition of the isolated target word was presented 1.5 seconds after the first target label onset. Finally, 500 ms after the second repetition of the target word, infants heard another English phrase (e.g., “Do you like it?” or “Do you see it?”), and then the trial ended. Each test trial lasted 8 seconds. Infants were tested on the words that they had been trained on, and thus infants in the Intact-TP condition were tested on the HTP and LTP words, and infants in the Violated-TPs condition were tested on the violated-HTP and violated-LTP words. The test phase began with 2 familiar label-object trials to help infants get accustomed to the structure of the LWL procedure (Fernald et al., 2008). The remaining trials were presented in four quasi-random testing orders for a total of 32 test trials. The target object appeared on both the left and right sides of the display screen an equal number of times. No labels occurred twice in succession. The entire experiment lasted about 10 minutes.

Vocabulary Measures

Caregivers completed a demographic information questionnaire and the MacArthur-Bates Communicative Developmental Inventory (MCDI; Fenson et al., 2000). We used the toddler short form (Level II, Form B, for 16- to 30-month-olds), which contains a 100-word vocabulary production checklist. Infants in Intact-TP and Violated-TP conditions were comparable in their expressive vocabulary scores (Intact-TP: M = 52.5, SD = 28; Violated-TP: M = 47.9, SD = 27.3, t (62) = .669, p = .506, d = .17).

Data Coding & Analysis

Infants’ eye gaze was video-recorded at a rate of 30 frames per second and coded offline by trained coders using iCoder software (Fernald et al., 2008). On each frame, the coder indicated whether the infant was looking to the object on the left, to the object on the right, shifting between objects, or whether the infant was looking away/off task (see Fernald et al., 2008). Word learning was assessed using an accuracy score that was based on the proportion of time spent looking at the target object divided by the total looking time to the target and the distracter object for each trial during the 300–2000 ms post-naming window.

A second coder re-coded a random selection of 25% of the infants. We obtained two reliability scores: a frame agreement score based on the percentage of frames on which two coders’ judgments agree overall, and a shift agreement score which focuses only on sequences of frames where the shifts occurred (Fernald et al., 2008). Frame and shift agreements between coders were greater than 98%.

This study was not preregistered. The data that support the findings of this study are openly available in OSF at https://osf.io/mfcyj/?view_only=9421b968ebc8465f8046be52b9df98d0.

Results

We performed linear mixed-effects modeling to examine infants’ ability to learn the trained mappings. Analyses were carried out through the lme4 package in R (Bates, Mächler, Bolker, & Walker, 2015; R Development Core Team, 2019). The model was fitted with the maximal random effects structure, including a by-subject random intercept and slope for Label Type (i.e., HTP and LTP) and a by-item random intercept and slope for Condition (i.e., Intact TP and Violated TP; Barr, Levy, Scheepers, & Tily, 2013) After pruning of random effect structure for singularity, the final model included a by-subject random intercept and a by-item random slope for condition. An Interaction between Condition and Label Type was also added to the model. We used dummy coding to represent the fixed effects of Condition and Label Type: The reference level for Condition was Violated TP, while the reference level for Label Type was LTP.

The intercept in our model was statistically significant, β = .595, SE = .023, p < .001, 95% CI = [.549 – .640]. The main effects of Condition, β = −.044, SE = .025, p = .084, 95% CI = [−.094 – .006] and Label Type were not significant, β = −.001, SE = .019, p = .978, 95% CI = [−.039 – .038]. However, there was a significant interaction between Condition and Label Type, β = .078, SE = .039, p = .042, 95% CI = [.003 – .154] (see Table 2 for model parameters and results). Figure 2 depicts performance across the 4 conditions.

Table 2.

Model parameters and results

	β	SE	CI	p
Multilevel Model
Intercept	.595	.023	.549 – .640	<.001
Condition (Violated-TP)	−.044	.025	−.094 – .006	.084
Label Type (LTP)	−.001	.019	−.039 – .038	.978
Condition (Violated-TP) * Label Type (LTP)	.078	.039	.003 – .154	.042
Intact TP Linear Model (LTP)
HTP vs LTP	−.038	.027	−.090 – .014	.156
Violated-TP Linear Model (vLTP)
vHTP vs vLTP	.044	.028	−.012 – .099	.122
HTP Linear Model (vHTP)
HTP vs vHTP	−.083	.029	−.140 – −.026	.004
LTP Linear Model (vLTP)
LTP vs vLTP	−.004	.033	−.070 – .062	.909

Figure 2.

Bar plots of the mean proportion of time looking to target objects on HTP, LTP, violated-HTP, and violated-LTP test trials averaged across the analysis window (300–2000 ms). Error bars represent 95 % confidence intervals. Circles represent individual infant means.

Following up on the interaction effect between Condition and Label Type, we tested the specific predictions that there would be no difference in performance on HTP vs. LTP trials, but that performance on violated-HTP trials would be worse than performance on HTP trials. To assess the effects of Condition and Label type, separate mixed effects models were fit for each comparison. All models included random intercepts for both Subject and Item. We found no significant difference in performance on the HTP vs. the LTP trials [HTP: (M = %61, SD = %12) vs. LTP: (M = 57%, SD = 17%)], β = −.038, SE = .027, p = .156, 95% CI = [−.090 – .014]. We also found that performance on the HTP trials was significantly better (M = %61, SD = %12) than on the violated-HTP trials, (M = %53, SD = %11), β = −.083, SE = .029, p = .004, 95% CI = [−.140 – −.026] (see Figure 3). We did not have a specific prediction about performance on the LTP Violation trials, but the results provided no evidence that violating the internal statistics of the LTP words diminished performance, relative to that on LTP trials [LTP: (M = 57%, SD = 17%) vs. violated-LTP: (M = 57%, SD = 12%)], β = −.004, SE = .033, p = .91, 95% CI = [−.070 – .062] (see Figure 4). Thus, the interaction reflects the fact that violating HTPs significantly diminished performance but violating LTPs did not. Note that we found no significant differences in accuracy between violated-HTP and violated-LTP trials [violated-HTP: (M = 53%, SD = 11%) vs violated-LTP: (M = 57%, SD = 12%)] β = .044, SE = .028, p = .122, 95% CI = [−.012 – .099]. We address this point in the Discussion.

Figure 3.

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

Figure 4.

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

Although the model intercept was significant, indicating that infants performed above chance on violated-LTP trials, the significant interaction suggests that performance was not equivalent in all conditions. Thus, we tested whether infants’ accuracy was significantly greater than chance (or .5) for each label type using planned one sample t-tests (see Table 3 for means). We used a Bonferroni correction, such that alpha was set to p < .0125, and interpret the pattern of statistical significance alongside estimates of effect sizes (here Cohen’s d) and confidence intervals (Cumming, 2012). Figure 2 shows the means and their 95% CIs for performance on each of the label types.

Table 3.

Descriptive statistics

	N	Mean	Median	SD	SE
Intact TP
HTP	32	.613	.605	.121	.021
LTP	32	.572	.576	.171	.030
Violated TP
HTP	32	.529	.518	.107	.019
LTP	32	.566	.580	.116	.020
Karaman et al. (under review) - Exp2
Unfamiliar	32	.565	.566	.098	.017

The results of the one-sample t-tests indicated that infants successfully learned HTP (M = 61%, SD = 12%), t (31) = 5.29, p < .001, 95% CI = [.07 – .16], d = .94 as well as violated-LTP labels (M = %57, SD = %12), t (31) = 3.24, p = .003, 95% CI = [.02 – .11], d = .57. Note that these were large and medium effect sizes respectively, and the 95% CIs did not include a mean difference of zero. Using the stringent Bonferroni criterion, performance on the LTP labels was not significantly better than chance; (M = 57%, SD = 17%), t (31) = 2.37, p = .024. However, the effect size was medium (d = .42), and like in the HTP and violated-LTP conditions, the 95% CI of the mean difference did not include zero (95% CI = [.01 – .13]). Performance on the violated-HTP labels did not differ from chance (M =53%, SD = %11), t (31) = 1.52, p = .138. Furthermore, the effect size (d = .27) was small, and the 95% CI of the mean difference includes zero (95% CI = [−.01 – .07]). Taken together, these results suggest that the HTP advantage is diminishing at 24-months of age, but that TPs nonetheless affect mapping by leading word forms containing violations of HTPs to be challenging to learn.

We next tested our prediction that the HTP advantage (i.e., performance on HTP words relative to the LTP words) would be smaller in toddlers with larger native-language vocabularies. To test this, we subtracted performance on the LTP trials from that of the HTP trials in infants in the Intact-TPs condition. There was a negative relationship between this measure of the HTP advantage and vocabulary size, such that infants with smaller vocabularies demonstrated an HTP advantage, but infants with larger vocabularies tended to learn HTP and LTP words more similarly, r (32) = −.358, p = .0449, 95% CI = [−.628 – .010]. This relation was primarily driven by performance on the LTP trials, such that infants with larger native-language vocabularies were better able to learn LTP labels than infants with smaller vocabularies, r (32) = .51, p = .003, 95% CI = [.195 – .729] but performance on HTP trials was not related to vocabulary size, r (32) = .18, p = .33, 95% CI = [−.182 – .496] (see Figure 5). This is consistent with our hypothesis that the HTP advantage decreases as a function of lexical development due to increases in the ability to learn words with relatively weak internal TPs. However, we consider alternative explanations in the Discussion.

Figure 5.

Correlations between vocabulary size and accuracy performance on HTP and LTP test trials in the Intact TP condition.

We also tested the more exploratory hypothesis that infants with larger vocabularies would be more resistant to learning violations of HTP words, relative to infants with smaller vocabularies. However, there was no relation between vocabulary size and learning violations of HTP words, r (32) = .06, p = .76, 95% CI = [−.298 – .397]. Instead, infants with larger vocabularies tended to learn the LTP violations better than those with smaller vocabularies, r (32) = .50, p = .004, 95% CI = [.184 – .723] (see Figure 6).

Figure 6.

Correlations between vocabulary size and accuracy performance on violated HTP and violated LTP test trials in the Violated TP Condition.

In sum, infants failed to show evidence of a clear HTP advantage in this experiment. Infants learned HTP words significantly better than violated HTP words, but there was no evidence for a difference between learning of LTP words and LTP violations. Furthermore, infants with larger vocabularies learned LTP words, and words that violate them, better than infants with relatively small vocabularies, but performance on HTP words and HTP violations was unrelated to vocabulary size.

Discussion

By 17 months of age, sequences with HTPs appear to be advantaged in word-referent mapping tasks, relative to unfamiliar words as well as sequences with lower internal coherence (Graf Estes et al., 2007; Hay et al., 2011). Although few studies have tested developmental changes in the role of TPs in word learning, one study with adults suggests that the HTP advantage may eventually be supplanted by resistance to mapping sequences that break up HTP words (Mirman et al., 2008). Here we tested whether these changes in the role of TPs in mapping word forms to meanings are emerging as early as the transition to toddlerhood. Specifically, we tested the hypothesis that 24-month-old infants will map both HTP and LTP words to referents, and that they will resist mapping word forms that break up HTP words (i.e., HTP violations). We also tested whether these changes are related to gains in infants’ word-learning abilities, as indexed by vocabulary size. To that end, we familiarized two groups of 24-month-olds learning English as their native language with fluent Italian speech that contained HTP and LTP words. We then tested how well one group of infants mapped the HTP and LTP words to referents, and how well a second group mapped sequences that violated the statistics of the HTP and LTP words.

We found that infants show a largely adult-like pattern by 24 months of age. More specifically, we found no clear evidence of an HTP advantage at 24 months of age: Infants mapped HTP and LTP words to a similar degree. Note that the highest mean performance was 62% (on HTP trials), and thus none of the words were learned extremely well. However, the size of the difference from chance was large for HTP words, while the size of the difference from chance on LTP trials was only medium in comparison, suggesting that the HTP advantage, while not statistically reliable in this study, may not be entirely absent. Thus, we interpret these results to suggest that the HTP advantage is diminishing by 24-months of age, and specifically that infants do not get – or in fact need – a boost in mapping tasks from experience.

However, one might wonder whether the HTP advantage is truly diminishing, such that 24-month-olds did not benefit from strong TPs when learning word-referent mappings, or whether infants instead benefitted from the familiarity of both HTP and LTP words. If infants did not benefit from strong statistical coherence when mapping word forms to referents, then infants would not only learn HTP and LTP words equivalently, but they also would not learn the HTP and LTP words any better than unfamiliar sequences. The finding that infants successfully learned LTP violations, which contained unfamiliar syllable sequences, and no evidence that they learned them more poorly than they learned LTP words, suggests that infants’ do not necessarily learn familiar sequences better than unfamiliar ones. In fact, we also found no evidence that they learned HTP words significantly better than violated-LTP words (HTP vs violated-LTP, β = −.037, SE = .029, p = .206, 95% CI = [−.095 – .021]). This suggests that successful mapping of the HTP and LTP words does not hinge on their familiarity, for if it did, infants should have performed worse on violated-LTP words than the familiar HTP and LTP words. Thus, in contrast to 17-month-olds familiarized with this same language, who learn HTP words better than LTP words and unfamiliar ones, 24-month-olds are relatively good at learning all 3 kinds of word forms.

We also found that relative to HTP word, 24-month-olds resisted learning sequences that violated HTP words, similar to the finding of Mirman et al., (2008) that adults learned sequences that violated HTP sequences more slowly than both HTP sequences and unfamiliar sequences. Interestingly, infants mapped LTP violations successfully, and there was no evidence that they learned them worse than LTP words. The finding that infants showed resistance to mapping only violations of the HTP words suggests that infants’ representations of HTP and LTP words are not equivalent. Critically, this suggests that infants were sensitive to the statistics in the Familiarization stream, and tracked them well enough to differentiate between the HTP and LTP words. We propose that this difference is key to explaining why only HTP violations were poorly learned. One possibility is that after hearing a perfectly predictive relationship between the initial and final syllables of HTP words during Familiarization, the HTP violations violate something like a “rule” for infants. If so, then the HTP violations might be perceived as poor word-form candidates, and thus mapped more poorly. In contrast, the initial syllables of LTP words are followed by several different syllables. Thus, the LTP violations are less likely to be perceived as violations, and more likely to serve as good word-form candidates. A related possibility is that LTP violations are easier to accurately encode and process than HTP Violations for this same reason.

Another related possibility is that HTP words are encoded or represented more robustly than LTP words due to their higher statistical coherence, even though infants learned them as labels equivalently when they were used in ostensive labeling phrases. By this age, when a novel word form is encountered, similar known words are co-activated, and interfere with mapping the novel word form to a referent (Mani & Plunkett, 2011; Swingley & Aslin, 2007). Furthermore, word forms with more robust representations are more strongly activated when hearing similar sounding words (Swingley, 2007). Thus, if infants form relatively strong representations of HTP words during familiarization, the HTP words should be strongly activated during the referent training phase when the HTP violations are presented. As a result, the violated HTP words would be poorly encoded due to competition or inhibitory activation from the HTP words. This would lead to weaker learning of the violated HTP mappings, relative to intact HTP mappings, which is what we observed. If the LTP word forms are not as strongly represented as HTP word forms, due to their weaker statistics, hearing violated LTP words during referent training should lead to relatively weak activation of LTP words, and to weak competition with the LTP violations. This would also lead to less inhibition of LTP violations from LTP words, relative to inhibition of HTP violations from HTP words, and would ultimately lead to the relatively spared mapping that we observed for the LTP words. These findings should be replicated in future studies, but provide potentially important insights into how TPs may influence whether or not word forms are successfully mapped to referents by 24-months of age.

We predicted that changes in the role of TPs in word learning would be correlated with infants’ vocabulary size. However, the results that we obtained were only partially consistent with our predictions. In line with our hypothesis, the HTP advantage, which is present in younger infants (Hay et al., 2011), and here in 24-month-olds with smaller vocabularies, appears to diminish as infants’ vocabularies grow. However, we predicted that infants with larger vocabularies would also show more resistance to learning HTP violations than infants with larger vocabularies, but we found no evidence for such a relation. If the relatively poor learning of violated HTP words results from strong lexical competition from HTP words, given that learning HTP words was not related to vocabulary size, then we would not expect resistance to learning violations to be related to vocabulary size either.

Interestingly, we found that infants with larger vocabularies, who tended to learn LTP words better than infants with smaller vocabularies, also tended to learn LTP violations better. On the one hand, this result suggests that infants with larger vocabularies, relative to those with smaller vocabularies, are better able to learn words that do not have strong internal TP statistics (i.e., LTP words and violated-LTP words), as long as they do not violate HTP words. This advantage may be due to a better ability to form associations between novel or weakly represented word forms and referents. However, there was no relation between how well infants learned unfamiliar Italian words and their vocabulary size in Karaman et al. (under review). Thus, the correlation we obtained may not suggest that infants with relatively large vs. small vocabularies were better able to learn forms that do not have strong internal TPs. Instead, it may reflect differences in how infants with smaller versus larger vocabularies are impacted by syllable frequency. The LTP words and LTP violations contained syllables that were 3 times more frequent than the syllables in the HTP words in the initial familiarization phase, even though the HTP and LTP words themselves were equally frequent. Benefits from high frequency of occurrence are well-attested in many aspects of language acquisition (Ambridge et al., 2015). Thus, the correlations we observed may suggest that infants with larger vocabularies are more likely to benefit from high syllable frequency in word learning. Additional research is necessary to determine why infants with larger English vocabularies showed stronger mapping of the LTP and violated-LTP word forms, and whether they are better able to map word forms regardless of familiarity, or whether they are more likely to benefit from high syllable frequency in forming mappings.

On a related point, it is important to note that some of our results could be specific to the materials we used, and to an American-English-learning sample. For example, it is not clear that we would obtain the same results if we used a language either more or less similar to English than Italian in its phonology and prosody, as TP tracking can be impacted by phonological and phonotactic cues (Finn & Kam, 2008; Thiessen & Saffran, 2003). Also, as mentioned above, the LTP words and LTP violations contained very frequent syllables, which may have played a role in how they were learned, or in the relations between their learning and vocabulary size. Furthermore, given that Mirman et al. (2008) found that LTP part-sequences were eventually learned just as well as HTP sequences, infants might also learn HTP violations if they encountered them often enough. Likewise, resistance to learning HTP violations may diminish over time, as memories for newly encoded statistical distributions and potential word forms decay (Karaman & Hay, 2018; Simon et al., 2017). Nonetheless, our findings are in line with results from adults using a task with substantially different materials (Mirman et al., 2008), suggesting that the key pattern of results reflects a real developmental change in the role of TPs in lexical development.

Altogether, these results suggest that the role of TPs in learning word-referent mappings changes across development. At 17-months of age, word forms with relatively strong TPs are better mapped to referents than those with relatively weak TPs, but by 24 months there is no evidence for such an advantage. This pattern should be replicated in a study directly comparing infants of different ages. However, they are suggestive of a developmental change that is likely to reflect a combination of factors. First, older infants are better able to learn word-referent mappings due to increases in the ability to rapidly encode novel word forms, and to establish and retain connections between these words and their referents. Thus, relatively novel word forms’ statistical coherence is less likely to determine whether 24-month-olds succeed or fail to learn a new mapping (see Swingley, 2007, for a similar finding). Second, TPs may carry less weight in identifying which sequences are learned as labels as language development progresses. Specifically, TPs may be less influential than more language-specific cues like phonotactics, stress etc., or than use in ostensive labeling phrases, in identifying whether an auditory sequence is a discrete word form that should be mapped to a referent. Here the use of ostensive labeling phrases from infants’ native language may have led them to form strong mappings, regardless of whether the word forms themselves contained high TPs. This suggestion is consistent with theories suggesting that statistical learning mechanisms are especially important to lexical development before infants can utilize relevant language-specific cues (e.g., Thiessen & Saffran, 2003).

Nonetheless, our results clearly suggest that infants continue to track TPs at 24 months of age, and that sensitivity to TPs continues to have an impact on word-referent mapping, albeit a different one. Thus, an important implication of these results is that the ability to track TPs in speech may be related to language proficiency in different ways across development. During infancy, tasks assessing the ability to distinguish between HTP and LTP sequences may be positively related to lexical development, as this ability may support vocabulary growth by facilitating the formation of word-referent mappings when learners would otherwise struggle to do so (Lany et al., 2018; Frost et al., 2020; Hoareau et al., 2019). However, the ability to distinguish between HTP and LTP sequences may not be related to measures of lexical development later on in toddlerhood or adulthood, as discriminating between HTP and LTP sequences may not support learning mappings between words and referents beyond infancy, but may instead lead infants to resist learning similar-sounding word forms. Thus, these results may help researchers to make better predictions about how individual differences in sensitivity to TPs in speech at different points in development are related to measures of language proficiency.

In sum, while 17-month-old infants typically learn HTP words better than LTP words (Graf Estes et al., 2007; Hay et al., 2011), our study suggests that 24-month-olds can learn both HTP and LTP words, especially those infants with larger native-language vocabularies. At this age, there may be a small HTP advantage as a function of the familiarity and strong coherence of HTP words relative to unfamiliar ones, but that advantage is modest at best. There is evidence that HTP words are nonetheless represented differently than LTP words, as infants only resist mapping sequences that violate HTPs. This effect potentially results from stronger interference from HTP words than LTP words when hearing the violated versions of these word forms. These results confirm that the ability to track TPs in speech is highly relevant to infants’ language development. They also suggest that the role of TPs changes with development. In future work, it will be important to probe the mechanism by which infants develop resistance to learning sequences that violate strong TPs.

Public Significance Statement.

Learning words is a fundamental aspect of early language development. This experiment sheds light on how the mechanisms that support word learning change across time, and specifically suggest that experience with patterns in speech relevant to finding word forms play an important role in mapping word forms to meaning.

Appendix

Familiarization Languages

Language A

HTP words: fuga, melo LTP words: casa, bici

Spesso Lisa capita in fuga nella casa dove giaci gracile e tesa.

Se cadi con la bici prima del bivio del melo cavo ti do dieci bigoli e una biro.

Gli amici della cavia Bida poggiano le bici in bilico presso il melo per difesa dalla biscia.

Sovente carico la spesa nel vicinato dopo una fuga con la bici nuova.

Carola si è esibita in una fuga verso il melo perché offesa dagli amici scortesi.

Se vai a casa in bici ti debiliti ma cali e non sei più obesa.

Dietro la casa del capo ho sprecato i ceci sotto al melo ombroso.

Se cuci subito sulla divisa bigia il distintivo col melo vado in casa a dormire.

Teresa si abitua alla fuga da casa con la vecchia bici senza luci posteriori.

Taci sulla fuga di Marisa con il caro lattaio.

Il bel melo sta tra la casa dei Greci e la chiesa arcana dove hai giocato con le bilie.

I soci della ditta Musa si danno alla fuga con la bici della maglia rosa.

Language B

HTP words: casa, bici LTP words: fuga, melo

Roméro fu coinvolto in una futile fuga in bici verso il profumo del mélo ombroso.

Il collega di Paolo Fusi trovò la bici per la fuga presso la casa del molo.

La maga tiene in casa almeno un fuco, uno squalo e una tartaruga del Nilo.

Il fuco procede parallelo alla casa sulla riga tracciata dalla cometa.

Il gattone Refuso medita sul mélo presso casa ascoltando una fuga di Verdi.

Il fu Medo Rossi ruppe la braga nella bici il mese scorso durante la gara.

Giga ogni mese paga con zelo l’affitto per la casa con il melo in fiore.

Meco prega il cielo che ogni fuga da casa termini sotto melo ombroso.

Il delfino beluga si dimena tutto solo nella fuga verso il Nilo azzurro.

Un pezzo di filo si è infilato nella bici appoggiata al melo dietro la méscita.

Vi fu un tempo in cui la bici in lega non temeva il gelo del rifugio della Futa.

La strega del melo fu vista in fuga sulla bici con un chilo di rametti.

Data Availability Statement

The data that support the findings of this study are openly available in OSF at https://osf.io/mfcyj/?view_only=9421b968ebc8465f8046be52b9df98d0