2018 Friday Poster 6516

Friday, November 2, 2018 | Poster Session I, Metcalf Small | 3pm

Frequency or Predictability? The Effect of Entropy on Statistical Learning in Children and Adults
O. Lavi-Rotbain, I. Arnon

Frequency effects are prevalent across many aspects of language learning and processing1–3. Frequency is, however, only one measure of the distributional structure of the linguistic environment. More complex measures, like entropy, also impact language: for instance, the length of words in the lexicon is impacted by both frequency and entropy4–6. However, little work to date has examined the impact of such measures on learning. Here, we examine the effect of entropy on statistical learning (SL) in children and adults, and find that reducing entropy is beneficial for learning in both.

Most SL studies present learners with a uniform distribution where all elements appear an equal number of times. This uniform distribution differs from natural language7, and results in a less predictable language compared to one with a non-uniform distribution (where words differ in frequency). This difference can be captured using entropy, with higher entropy indicating a less predictable language. Lower entropy could lead to better learning of the language as a whole and/or to better learning of the lower frequency items (compared to a uniform distribution with the same frequency). In previous work, adults learned equally well from a uniform and Zipfian distributions, despite the reduced entropy of the latter8-10. However, this may have been driven by increased accuracy for the more frequent items: reduced entropy may still improve learning for the lower frequency items. We examine this prediction by looking at both children and adults, across several entropy levels and exposure durations.

We used a word segmentation task with four ‘words’ modelled on Saffran et al. (1996). We manipulated entropy by making one word more frequent: it appeared 80% of the time (low entropy), 55% (medium) or 25% (uniform distribution with high entropy). The word chosen as the frequent one was counterbalanced across subjects. As predicted, adults benefitted from the reduction in entropy (Table 1). They showed better segmentation of the language as a whole in the low entropy condition (low vs. medium: t(66.55)=4.0, p<0.001; low vs. high: t(57.29)=4.6, p<0.001), and better learning of the low frequency words in this condition, despite hearing them less (infrequent word accuracy: low vs. medium, t(65.65)=3.5, p<0.001; low vs. high, t(56.87)=3.59, p<0.001). Children also benefitted from the reduction in entropy, but only when the infrequent words appeared often enough (Table 2). When each word appeared only 19 times, children did not learn the language (they were at chance, t(17)=0.13, p=0.9). However, when exposure was increased, the low entropy condition facilitated children’s learning of both the language as a whole compared to uniform distribution with the same length (t(27.94)=1.99, p=0.056), and of the infrequent words compared to a uniform distribution with the same frequency (t(25.91)=2.56, p<0.05). These results suggest that reduction in entropy can facilitate SL learning both overall and for infrequent items, with two restrictions: a large enough reduction in entropy for adults, and a minimal frequency of infrequent items for children. We discuss broader implications for the role of entropy in language learning.

References

  1. Goodman, J. C., Dale, P. S. & Li, P. Does frequency count? Parental input and the acquisition of vocabulary. J. Child Lang. 35, 515–531 (2008).
  2. Bannard, C. & Matthews, D. Stored Word Sequences in Language Learning. Psychol. Sci. 19, 241–248 (2008).
  1. Arnon, I. & Snider, N. More than words: Frequency effects for multi-word phrases. J. Lang. 62, 67–82 (2010).
  2. Piantadosi, S. T., Tily, H. & Gibson, E. Word lengths are optimized for efficient communication. Proc. Natl. Acad. Sci. U. S. A. 108, 3526–3529 (2011).
  3. Cohen Priva, U. Not so fast: Fast speech correlates with lower lexical and structural information. Cognition 160, 27–34 (2017).
  4. Matthews, D. & Bannard, C. Children’s production of unfamiliar word sequences is predicted by positional variability and latent classes in a large sample of child-directed speech. Cogn. Sci. 34, 465–488 (2010).
  5. Piantadosi, S. T. Zipf’s word frequency law in natural language: A critical review and future directions. Psychon. Bull. Rev. 21, 1112–1130 (2014).
  6. Kurumada, C., Meylan, S. C. & Frank, M. C. Zipfian frequency distributions facilitate word segmentation in context. Cognition 127, 439–453 (2013).
  7. Blythe, R. A., Smith, K. & Smith, A. D. M. Learning times for large lexicons through cross- situational learning. Cogn. Sci. 34, 620–642 (2010).
  8. Vogt, P. Exploring the Robustness of Cross-Situational Learning Under Zipfian Cogn. Sci. 36, 726–739 (2012).