Statistical Aggregation and Hypothesis Testing Mechanisms Interact during Word Learning

Referential utterances are by their nature ambiguous to novice language learners. Each utterance consists of multiple layers of information that must be decoded: 1) the linguistic structure (how the sounds and words should be packaged into meaningful units), 2) the world structure (how people, objects and actions in the world relate to one another and which is the current focus of attention) and 3) the relation between the language and the world. Tracking structure over time, frequently referred to as “statistical learning,” is one way to resolve these ambiguities and gain access to additional patterns within and across domains. For example, in statistical learning of word boundaries, human learners, including infants, track the relative predictability of sequences of syllables in the language stream. Such tracking facilitates word segmentation because individual words can be used in many difference linguistic contexts. Thus, syllable sequences within words will tend to travel together, or be more predictable, than the sequences that bridge word boundaries (e.g., the word sequence was preceded by the words of, syllable and the in the sentences above, making the syllable bigram se-quence more predictable than of-se, the-se or ble-se). Temporal cues in the auditory domain, such as the relative predictability of syllable sequences discussed above, may be augmented by additional cues from other domains or modalities. For example, there is evidence that word segmentation and word learning (mapping segmented words to visual referents) are more efficiently achieved when the two processes are done simultaneously rather than sequentially (Yurovsky, Yu & Smith, 2012). Such learning, of tracking language and world structures over time, has been termed “crosssituational” learning, with the different world-situations, or contexts, serving the disambiguating role that the different syllable contexts serve in statistical word segmentation.