Specificity of Statistically Segmented Words

• In Event: 2-041 - Specificity of Infant Statistical Learning

Integrative Statement

Infants have demonstrated remarkable abilities to rapidly track sound patterns in novel languages (Saffran et al., 1996). They can also exploit experience with statistical patterns to support word learning. That is, infants can extract words using syllable-level transitional probability patterns (i.e., high probabilities within words and low probabilities across word boundaries) and subsequently apply the segmented words to map to meanings (Graf Estes et al., 2007). Despite evidence of powerful early learning mechanisms, there have been questions about how useful statistical learning is for language acquisition and processing. For example, Endress and Mehler (2009) argued that statistical learning may allow learners to track probabilities, but it does not yield specific word-like units. In the present experiments, we examined the nature of the representations that infants form. If infants form specified word-like units, word learning should only be bolstered by units that match with statistically segmented words. If infants accept words that only overlap partially with the statistically-defined words, it may suggest statistical learning yields unspecified units, supporting arguments against the utility of statistical word segmentation in language acquisition.

We tested 17-month-olds in a task that integrated statistical word segmentation with object label learning. All infants listened to an artificial language consisting of four 3-syllable words (pabiku, tibudo, golatu, daropi). They then participated in a habituation-based label-object association task (Werker et al., 1998). Infants saw two novel objects, one at a time, and heard repetitions of each label. After habituating to two label-object pairs, infants viewed Same trials in which the original label-object pairings were maintained and Switch trials in which the label-object pairings were switched (object 1 + label 2).

In the Word label condition (n=22), the labels were two words from the artificial language (pabiku, tibudo). In the Position label condition (n=22), the labels contained syllables from the artificial language in their original positions, but novel combinations. For example, "parotu" was created from the words pabiku, daropi, golatu. If infants encode word position rather than consistent syllable units, labels maintaining syllable positions should promote learning. In the Edge label condition (n=22), the labels (paroku, tilado) maintained the first and final syllables of words. If infants only require an approximate match to segmentation experience, these labels should support learning.

We found that infants learned the Word labels, looking significantly longer on Switch trials than Same trials (p=.048). In contrast, infants displayed no evidence of learning Position labels (p=.609) or Edge labels (p=.075; marginally longer looking to Same trials). Infants only learned the labels that matched the units from the speech stream.

These findings support the idea that infants extract specified word-like representations during statistical learning. When word positions were maintained, but transitional probabilities were violated, and when labels had the same onsets and codas as the artificial language words, infant did not learn the labels. Infants required a tight match with the segmented words forms to transfer the experience to label learning.

Authors

• Katharine Graf Estes, University of California, Davis
  Presenting Author

• Dylan Antovich, UC-Davis
  Non-Presenting Author