2018 Sat Poster 6588

Saturday, November 3, 2018 | Poster Session II, Metcalf Small | 3:15pm

Distributional Learning is Constrained by Prior Language Experience
S. Simackova, K. Chladkova

Distributional learning can be understood as inattentive tracking of stimulus probability distributions in the input. In the domain of speech, the literature suggests that listeners exposed to a bimodal distribution of stimuli along an auditory dimension subsequently discriminate a novel contrast along that dimension better than those exposed to a unimodal distribution [1,2]. However, a number of recent studies failed to find the expected distributional-training effects [3,4].

While adults are sensitive to distributional statistics across modalities [5,6], whether or not this sensitivity results in the formation of new categories or contrasts may be subject to factors other than the input statistics, for instance, prior experience. Specifically for speech, adults, unlike infants, already have a speech sound system in place. This brings up the question of whether the native phonological system can affect how adults learn from the sounds’ statistics. To this end, we compared distributional speech sound learning in participants with different language backgrounds and tested whether learning outcomes are affected by prior language experience.

Czech and Greek adults, whose native phonologies contain and lack, respectively, the phonological vowel length feature, were exposed to tokens of a novel vowel quality [ɑ] falling into a unimodal or bimodal distribution along the durational dimension. Sensitivity to duration differences was assessed before and after training to reveal whether and what the listeners learned from the distributional statistics. We  hypothesized that if the ability to track probability distributions is affected by the native phonology in that learners can easily reuse their native-language contrasts or features,  we would find the effects of distributional training to be larger for an already colonized dimension, i.e. in Czechs, as opposed to an unused dimension, i.e. in Greeks. If, on the contrary, the ability to track speech sounds’ distributions is not modulated by native phonology, the effects of training would be comparable between Czechs and Greeks.

A linear mixed model detected an interaction of Language, Test (pre, post), Training (unimodal, bimodal), and Duration difference (identical pair, 15%-, 30%-, 45%-difference), t=-2.205, p=0.028. For the most difficult difference overall (15%- difference), two groups improved their discrimination accuracy from pre- to post-test: bimodally-trained Czechs and unimodally-trained Greeks. Language background thus affected how listeners learn from the distributions. Czechs, who employ the trained auditory dimension to distinguish native vowels and possess higher-order representations that can guide the bottom-up learning, benefited from bimodal exposure and learned to more accurately differentiate short vs. long [ɑ]s. Reversely, Greeks, unfamiliar with the trained dimension and lacking mental representations for length, did not profit from the distributional statistics. Instead, they showed an unexpected training effect after unimodal exposure: they sensitized to within-category variation of the frequently presented stimuli.

We conclude that in adults, exposure to sounds’ probability distributions may not be sufficient for novel categories or contrasts to be formed unless listeners have feedback from some established mental representations. The requirement of this form of implicit feedback indicates that, at least in the domain of speech, distributional learning may not be available as a purely unsupervised mechanism throughout lifetime.

References

  1. Maye, & Gerken, L.A. (2001). Learning phonemes: How far can the input take us? BUCLD 25 Proceedings, 480-490.
  2. Clayards, M., Tanenhaus, M. K., Aslin, R. N., & Jacobs, R. A. (2008). Perception of speech reflects optimal use of probabilistic speech Cognition, 108, 804–809.
  3. Wanrooij, , Boersma, P., & van Zuijen, T.L. (2014). Distributional vowel training is less effective for adults than for infants. A study using the mismatch response. PLoS One, 9(10), 1-13.
  4. Wanrooij, K., De Vos, J.F., & Boersma, P. (2015). Distributional vowel training may not be effective for Dutch adults. In Proceedings of the 18th International Congress of Phonetic Sciences, paper
  5. Love, C. (2003). The multifaceted nature of unsupervised category learning. Psychonomic Bulleting & Review, 10, 190–197.
  6. Garrido, I., Teng, C.L.J., Taylor, J.A., Rowe, E.G., & Mattingley, J.B. (2016). Surprise responses in the human brain demonstrate statistical learning under high concurrent cognitive demand. npj Science of Learning 1, 16006.