The present work examined the discovery of linguistic cues during a word segmentation task. Whereas previous studies have focused on sensitivity to individual cues, this study addresses how individual cues may be used to discover additional, correlated cues. Twenty‐four 9‐month‐old infants were familiarized with a speech stream in which syllable‐level transitional probabilities and an overlapping novel cue served as cues to word boundaries. Infants' behavior at test indicated that they were able to discover the novel cue. Additional experiments showed that infants did not have a preexisting preference for specific test items and that transitional probability information was necessary to acquire the novel cue. Results suggest one way learners can discover relevant linguistic structure amid the multiple overlapping properties of natural language.
Natural languages exhibit structure at multiple levels in parallel (e.g., phonological, lexical, morphological, syntactic, and discourse). For the adult listener, this complexity creates temporary ambiguities that must be resolved for speech to be understood. Individual bits of information are imprecise, such as the meaning of words like bow, colon, saw, and wave. Such ambiguities are resolved via a constraint satisfaction process that exploits correlations among different types of information ([
While studies of adult language have investigated how constraints are combined to resolve ambiguities, studies of language acquisition have examined how children use statistical cues to learn their native language. These are complementary issues, the "constraints" that are relevant to adult listeners are the "cues" by which the child acquires language ([
Experiments investigating language acquisition via statistical learning have typically focused on infants' abilities to use one statistical cue at a time (e.g., [
Alternatively, the constraint satisfaction approach suggests that the complexity of natural language provides a rich system for learning mechanisms to exploit. Linguistic regularities reinforce each other across levels, allowing statistical learning mechanisms to capitalize on multiple cues and redundancies. For example, lexical stress patterns are found in numerous languages of the world. These patterns consist of a specific ordering of strong and weak syllables that occur frequently and can help identify word boundaries or classify groups of words. In English, many words have a trochaic, or strong–weak, stress pattern, as in the words BAby and MOmmy ([
Psycholinguistic studies support the hypothesis that infants are sensitive to the conjunction of multiple probabilistic cues ([
Despite this progress, it remains to be determined how language learners isolate and combine cues given the complexity of human language. In considering this problem, it is helpful to distinguish between two classes of potentially useful cues: language‐general cues and language‐specific cues. Measures of co‐occurrence or predictability between syllables are language‐general cues, in that they operate in similar ways across natural languages. For example, transitional probabilities are not specific to any given language (though the units over which these computations are performed are).
Other cues may or may not be useful in any given language and are thus language specific, and must be learned. For example, languages have different lexical stress patterns (iambic vs. trochaic), and in some languages stress patterns do not mark boundaries or help individual units cohere. By 9 months of age, infants typically show sensitivity to a range of language‐specific cues (for a recent review, see [
In the domain of word segmentation, previous work suggests that younger infants tend to use language‐general cues and later shift to language‐specific cues ([
How might infants discover these language‐specific cues? One potential explanation is that language‐general cues provide a basis for discovering overlapping or co‐occurring language‐specific cues. For example, in word segmentation, infants may use their sensitivity to TP cues, which is present early in life ([
The present work tested the hypothesis that infants can discover novel cues by exploiting redundancies between language‐general and language‐specific cues. Nine‐month‐old infants were exposed to a fluent speech stream that contained two overlapping cues to word boundaries: a language‐general cue (TPs) and a language‐specific cue (/t/‐onsets). TPs are known to be salient to 9‐month‐old infants. The second cue was specific to the artificial language and therefore novel: Each word in the speech stream began with /t/. Experiment 1 was thus designed to test the hypothesis that infants can use the language‐general TP cue to discover the overlapping language‐specific /t/‐onset cue. The /t/‐initial syllables are only informative as a cue to word boundaries due to their overlap with the TP cue; the TP cue positions the /t/‐initial syllable at the onset of each word. Consequently, the only way infants can extract this pattern is to use its overlap with the TP cue. We tested infants using items that were all novel relative to the exposure corpus but that varied in their use of the /t/‐onset cue. The question of interest was whether infants would be sensitive to the presence of /t/‐onsets in the test items. If so, this would provide evidence that infants can isolate individual cues by using redundancies in the speech stream.
To examine whether infants can use a language‐general segmentation cue to discover an overlapping novel language‐specific cue, infants heard a fluent speech stream that contained two overlapping cues to word boundaries: (a) dips in TPs at word boundaries and (b) /t/‐initial syllables at word onsets. To determine whether infants acquired the novel /t/‐onset pattern, test items either adhered to the pattern (began with a /t/‐syllable) or violated it (contained a medial /t/). Crucially, these items were previously unheard combinations of syllables from the speech stream (i.e., TP = 0). Therefore, TP information would not allow infants to distinguish between the two types of test items. Instead, successful discrimination hinged on discovery of the /t/‐initial pattern present in the speech stream played during familiarization.
Participants. Twenty‐four 9.5‐month‐old monolingual English‐learning infants (mean age = 9.5 months, range = 9.0–10.0) participated in this experiment. All infants were born full‐term and had fewer than four prior ear infections and no history of hearing or vision impairments. Data from an additional 8 infants were excluded due to fussiness (
Stimuli. A fluent stream of speech was created from recordings of a female native English speaker who was blind to the structure of the artificial language. The language contained six bisyllabic words: tohsigh, teemay, tiepu, tukee, tayla, and tafo. A pseudosynthesis technique was used to create the speech stream, which allowed for use of naturally produced syllables while permitting control over coarticulation, duration, pitch, and volume of all syllables in the language. All three‐syllable sequences that occurred in the language, both within and between word boundaries (e.g., tohsightee, sighteemay), were recorded in a monotone, isochronous register. Medial syllables were spliced out of the three‐syllable sequences and concatenated together with no silence between syllables. By using these medial syllables, coarticulation within each syllable and between every pair of syllables in the language was maintained. Syllables were edited prior to concatenation to have the same duration, pitch, and volume. The stream contained 40 repetitions (2 min 17 s) of each word in a pseudorandom order with no word appearing twice in succession (see the Appendix for a transcript of the complete familiarization language). Each within‐word syllable pair had a TP of 1.0; between‐word syllable pairs had a TP between.1 and.25 (M = .20). The speech stream thus contained two overlapping and completely redundant cues to word boundaries: dips in TPs and /t/‐onsets.
Four novel test items were constructed from syllables in the artificial language. Two of these items began with /t/ (tiemay, tohla), and two contained a medial /t/ (fota, keetu). Test items were created from recordings of each bisyllabic item spoken in isolation. Duration, pitch, and volume were edited so that all test items were essentially equivalent.
Procedure. During familiarization, infants listened to the speech stream at a comfortable volume, played over speakers mounted on each sidewall while viewing an unrelated Baby Einstein video. An experimenter then entered the booth, covered the monitor that displayed the video, and placed headphones playing masking music on the caregiver. The test phase began with two practice trials (a recording of piano tones), designed to help the infants learn the contingency between their head‐turns and the lights and sounds. The practice trials were followed by 12 test trials, three blocks of each of the four test items (tiemay, tohla, fota, and keetu). Infants' ability to discriminate the test items was assessed using the Headturn Preference Procedure ([
We tested infants' ability to discriminate /t/‐initial from /t/‐medial test items over the three blocks of testing with a 2 (test item type: /t/‐initial vs. /t/‐medial) × 3 (test block: 1, 2, 3) repeated‐measures analysis of variance (ANOVA; means shown in Table 1). The main effect of test item type (/t/‐initial vs. /t/‐medial) was not significant, F(
1 Mean and Standard Errors for Looking Times
Experiment 1 Experiment 2 Experiment 3 / / / / / / Block 1 8.69 (0.80) 7.54 (0.70) 8.44 (0.88) 8.75 (0.70) 8.30 (0.96) 7.75 (0.66) Block 2 7.45 (0.71) 6.02 (0.75) 7.00 (0.57) 7.56 (0.70) 6.77 (0.56) 6.95 (0.68) Block 3 4.58 (0.39) 5.94 (0.56) 6.79 (0.68) 6.32 (0.79) 7.18 (0.85) 6.77 (0.73) Blocks 1 and 2 (averaged) 8.07 (0.58) 6.78 (0.59) 7.57 (0.60) 8.18 (0.53) 7.53 (0.58) 7.35 (0.55) All trials (averaged) 6.92 (0.46) 6.53 (0.46) 7.40 (0.51) 7.52 (0.48) 7.31 (0.39) 7.09 (0.50)
Graph: 1 Experiment 1: Mean looking times to /t/‐initial and /t/‐medial test items for each test block.
Subsequent analyses focused on the first two blocks (eight test trials), as these looking times are more proximal to the familiarization phase and thus most likely to reflect learning from the fluent speech. A one‐way (test item type: /t/‐initial vs. /t/‐medial) repeated‐measures ANOVA revealed a significant difference in looking times to the two types of items, F(
Graph: 2 The average looking time for /t/‐initial and /t/‐medial items across test Blocks 1 and 2 for all three experiments.
The results of Experiment 1 suggest that infants were able to exploit the /t/‐initial pattern, successfully discriminating novel test items that followed the pattern from those that did not. This pattern was not immediately obvious in the input; the speech stream consisted of syllables beginning with /t/ alternating with syllables that began with other sounds. In order to discover that the /t/ segment signaled word onsets, infants presumably capitalized on the TP cues in the speech stream, which also provided cues to word boundaries. On this view, infants discovered the language‐specific /t/‐onset cue by capitalizing on the language‐general TP cue.
One interesting feature of these data is that infants' looking behavior changed over the course of testing. Familiarity to novelty preference switches is not uncommon in infant behavioral studies, though the factors responsible for the shift may vary ([
Graph: 3 The difference between looking times to /t/‐initial and /t/‐medial test items, for each block for all three experiments. Note. Positive values indicate a familiarity preference; negative values indicate a novelty preference.
The findings from Experiment 1 support the hypothesis that infants used low TPs at word boundaries to acquire the overlapping but novel /t/‐onset pattern. However, there is an alternate explanation for these results: It is possible that infants' behavior reflected preexisting preferences for individual items. A counterbalanced language composed of items that all contain a medial /t/ would clarify this issue. However, previous work ([
This experiment was designed to test the hypothesis that infants in Experiment 1 listened longer to the /t/‐initial test items due to an a priori preference for these particular items. Infants in Experiment 2 were not exposed to the familiarization speech stream, participating only in the testing procedure used in Experiment 1.
Participants. Twenty‐four 9.5‐month‐old monolingual English‐learning infants (mean age = 9.5 months, range = 9.0–9.9) participated. Data from an additional 5 infants were excluded from the analyses because of experimenter error (
Stimuli. The test items were the same as those used in Experiment 1.
Procedure. There was no exposure phase. The testing procedure was identical to Experiment 1, with 2 practice trials followed by 12 test trials.
As in Experiment 1, a 2 (test item type: /t/‐initial vs. /t/‐medial) × 3 (test block: 1, 2, 3) repeated‐measures ANOVA was conducted. There was no significant effect of test item type (/t/‐initial vs. /t/‐medial), F(
We next conducted a 2 (test item type: /t/‐initial vs. /t/‐medial) × 2 (group: Experiment 1, Experiment 2) repeated‐measures ANOVA over the data from the first two test blocks from Experiments 1 and 2. This analysis was intended to determine whether the behavior of infants differed reliably across the two experiments. The interaction between test item type and group was significant, F(
This between‐group analysis, coupled with the within‐subject analysis showing no effect of test item type, indicates that infants in Experiment 2 did not have an a priori preference for the /t/‐initial items relative to the /t/‐medial items. We can therefore attribute infants' behavior in Experiment 1 to familiarization with the fluent speech stream. Nevertheless, it is still unclear which aspects of the familiarization stimuli elicited infants' successful discrimination between /t/‐initial and /t/‐medial test items. It is possible that, as originally hypothesized, low TPs at word boundaries anchor the alternating /t/ syllables, allowing infants to extract the /t/‐onset cue and generalize it to the novel test items. Another possibility is that infants are extracting the extremely regular alternating /t/ syllable pattern (created because each of the bisyllabic words begins with a /t/) and uniformly mapping the /t/ syllable to word onsets. According to this alternative hypothesis, infants could capitalize on a systematic pattern (/t/‐onsets) without the aid of another cue. On this account, they detect the regular alternation and map it onto onsets, potentially because onsets are privileged perceptually and/or lexically (e.g., [
To explore this hypothesis, we designed a new speech stream to determine whether infants could extract the novel /t/‐onset pattern without the aid of another cue for bootstrapping. The resulting speech stream did not have low TPs at word boundaries but still contained the /t/‐onset pattern. If infants can extract the /t/‐onset pattern without an overlapping cue, infants in Experiment 3 should show a significant difference in looking time to /t/‐initial items compared to /t/‐medial items. However, if TPs played a critical role in the discovery of the novel pattern via bootstrapping, infants should not show a significant difference in looking time to /t/‐initial items compared to /t/‐medial items in the absence of TP cues.
This experiment was designed to determine whether the overlapping cue from Experiment 1, low TPs at word boundaries, was necessary for infants to acquire the novel /t/‐onset pattern. Infants were exposed to a new fluent speech stream that did not have low TPs at word boundaries but still contained the novel /t/‐onset pattern. The procedure and test items were identical to Experiment 1.
Participants. Twenty‐four 9.5‐month‐old monolingual English‐learning infants (mean age = 9.5 months, range = 9.1–10.0) participated in this experiment. Data from an additional 17 infants were excluded from the analyses because of parental interference (
Stimuli. A fluent speech stream was created using the procedure and words (tohsigh, teemay, tiepu, tukee, tayla, and tafo) from Experiment 1. Again, all syllables in the language were measured and edited such that duration, pitch, and volume were equivalent for all syllables. Unlike Experiment 1, in which words were repeated in a random order, the words in this fluent speech stream were repeated in exactly the same order 40 times (teemaytieputukeetafotaylatohsighteemaytieputukeetafotaylatohsigh. . .; see [
Procedure. The experimental procedure was the same as Experiment 1.
As in the previous experiments, we first ran a 2 (test item type: /t/‐initial vs. /t/‐medial) × 3 (test block: 1, 2, 3) repeated‐measures ANOVA. There was no significant effect of test item type, F(
Though the between‐group analysis is inconclusive, infants' failure to discriminate the test items in Experiment 3 is consistent with the hypothesis that infants were influenced by the presence of TP cues in Experiment 1. Without the low TPs at word boundaries to anchor the alternating /t/ syllables to segment onsets, infants seem unable to extract the novel pattern.
The results of these experiments indicate that infants are able to use a language‐general regularity (dips in TPs at word boundaries) to discover a second, language‐specific regularity (/t/‐onsets). Moreover, infants generalized this newly learned cue to novel items, as demonstrated by their test performance. These findings suggest one possible class of solutions to the learning problem described earlier: How do infants discover relevant linguistic cues when there are many way to analyze speech and a single bit of information can be informative at multiple levels? Just as adults use correlations among cues to resolve ambiguities when using language, infants are able to use such correlations to acquire language, as suggested by the constraint satisfaction approach. Thus, infants' learning capacities, such as the ability to encode correlations across different types of information ([
Research over the past decade has shown that infants are sensitive to many different patterns that can be informative for language learning. It remains unclear, however, how infants find individual cues and combine different types of information to understand the complex structure of their language. Behavioral research on multiple cue usage in this domain has typically taken the form of cue‐conflict studies, examining relative reliance on different types of information over time (e.g., [
The process observed in the present work is reminiscent of bootstrapping: Partial information about one element of language provides evidence about another element, which in turn provides further evidence for the first element ([
This work shows that infants can use the overlapping nature of speech to isolate cues. Language‐general regularities, such as TPs, may support the discovery of language‐specific cues. Critically, the present research also expands the scope of statistical learning mechanisms. Not only can infants use such mechanisms to exploit the structure of their language, but they can also use statistical learning to discover the structure of their native language.
teemaytohsighteemaytukeetafotieputaylatohsightayla tukeeteemaytieputaylateemaytukeeteemaytohsightu keetafoteemaytukeetohsighteemaytohsighteemaytoh sightafoteemaytafotohsightukeetieputukeetieputafo teemaytohsighteemaytukeetafotieputaylatohsightayla tukeeteemaytieputaylateemaytukeeteemaytohsi tee maytohsighteemaytukeetafotieputaylatohsightaylatu keeteemaytieputaylateemaytukeeteemaytohsightukee tafoteemaytukeetohsighteemaytohsighteemaytohsight afoteemaytafotohsightukeetieputukeetieputafoteemay tohsighteemaytukeetafotieputaylatohsightaylatukee teemaytieputaylateemaytukeeteemaytohsi teemaytoh sighteemaytukeetafotieputaylatohsightaylatukeetee maytieputaylateemaytukeeteemaytohsightukeetafotee maytukeetohsighteemaytohsighteemaytohsightafotee maytafotohsightukeetieputukeetieputafoteemaytoh sighteemaytukeetafotieputaylatohsightaylatukeetee maytieputaylateemaytukeeteemaytohsi teemaytoh sighteemaytukeetafotieputaylatohsightaylatukeetee maytieputaylateemaytukeeteemaytohsightukeetafotee maytukeetohsighteemaytohsighteemaytohsightafotee maytafotohsightukeetieputukeetieputafoteemaytoh sighteemaytukeetafotieputaylatohsightaylatukeetee maytieputaylateemaytukeeteemaytohsi teemaytoh sighteemaytukeetafotieputaylatohsightaylatukeetee maytieputaylateemaytukeeteemaytohsightukeetafotee maytukeetohsi.
By Sarah D. Sahni; Mark S. Seidenberg and Jenny R. Saffran
Reported by Author; Author; Author