Portal de Periódicos da CAPES

Local statistical learning under cross-situational uncertainty.

2011; Wiley; Volume: 33; Issue: 33 Linguagem: Inglês

1551-6709

Luca Onnis, Shimon Edelmann, Heidi Waterfall,

Language and cultural evolution

Local statistical learning under cross-situational uncertainty Luca Onnis (lucao@hawaii.edu) Department of Second Language Studies, University of Hawaii, Honolulu, HI 96822 USA Center for Second Language Research, University of Hawaii, Honolulu, HI 96822 USA Shimon Edelman (edelman@cornell.edu) Department of Psychology, Cornell University, Ithaca, NY, 14853, USA Heidi Waterfall (heidi.waterfall@gmail.com) Department of Psychology, Cornell University, Ithaca, NY, 14853, USA Department of Psychology, University of Chicago, Chicago, IL 60637, USA Abstract Statistical learning research often assumes that learners collect global statistics across the entire set of stimuli they are exposed to. In naturalistic settings, this assumption of global access to training data is problematic because it implies that the cognitive system must keep track of an exponentially growing number of relations while determining which of those relations is significant. We investigated a more plausible assumption, namely that learning proceeds incrementally, using small windows of opportunity in which the relevant relations are assumed to hold over temporally contiguous objects or events. This local statistical learning hypothesis was tested on the learning of novel word-to-world mappings under conditions of uncertainty. Results suggest that temporal contiguity and contrast are effective in multimodal learning, and that the order of presentation of data can therefore make a significant difference. Keywords: statistical learning; cross-situational learning; variation sets; language acquisition. What principles guide learning from multiple parallel streams of sensory information? How can humans find structure in sustained exposure to auditory and visual stimuli? Much of learning can be characterized as finding patterns in space and time under conditions of high uncertainty – from deriving categories from experience (e.g., Tenenbaum & Griffiths, 2001), through learning word meanings from their co-occurrence with perceived events in the world (e.g., Frank, Goodman, & Tenenbaum, 2009), to acquiring the different levels of linguistic structure (e.g., Solan, Horn, Ruppin, & Edelman, 2005). Behavioral studies in statistical learning (e.g., Gomez & Gerken, 2000) indicate that infants, children, and adults can extract regularities across a set of exemplars distributed in time and/or space. A core assumption of such studies is that to extract such regularities, learners collect global statistics across the entire set of stimuli they are exposed to, often over multiple trials or training sessions. When applied to naturalistic learning, this ‘global’ assumption is problematic: it requires that the cognitive system keep track of an exponentially growing number of relations among various pieces of data while identifying the significant relations. A more plausible assumption, investigated here, is that learning proceeds incrementally, using small windows of opportunity in which the relevant relations are those that hold over spatially and/or temporally neighboring objects or events. For example, in a study by Onnis, Waterfall, & Edelman (2008), adult learners were asked to individuate the novel words of an “alien” language from unsegmented sentences such as kedmalburafuloropesai. In the absence of acoustic and prosodic cues (sentences were generated by speech synthesis software), each sentence could in principle be composed of a range of possible words ranging from a single long word (as is not uncommon in polysynthetic languages such as West Greenlandic), to as many words as there were syllables. Onnis et al. (2008) found that learners were significantly better at the word segmentation task when some consecutive sentences in the training set overlapped in some of their syllables (e.g., kedmalburafuloropesai followed by rafuloro), compared to a control condition in which the order of the same set of sentences was scrambled so that no parts of adjacent sentences matched. When aligned, the partially matching sentences suggest candidate