Modeling Statistical Insensitivity: Sources of Suboptimal Behavior

Abstract

Children acquiring languages with noun classes (grammatical gender) have ample statistical information available that characterizes the distribution of nouns into these classes, but their use of this information to classify novel nouns differs from the predictions made by an optimal Bayesian classifier. We use rational analysis to investigate the hypothesis that children are classifying nouns optimally with respect to a distribution that does not match the surface distribution of statistical features in their input. We propose three ways in which children's apparent statistical insensitivity might arise, and find that all three provide ways to account for the difference between children's behavior and the optimal classifier. A fourth model combines two of these proposals and finds that children's insensitivity is best modeled as a bias to ignore certain features during classification, rather than an inability to encode those features during learning. These results provide insight into children's developing knowledge of noun classes and highlight the complex ways in which statistical information from the input interacts with children's learning processes.

Do children privilege phonological cues in noun class learning?

“…First, there is an effect of early availability of cues on learning which appears to hold across development; both adult and child learners tend to rely more on early available cues, regardless of whether the cues come from noun-internal phonological features or word meanings. Our findings are also consistent with the possibility that children have an active bias against semantic cues; when both a noun-internal and external cue are available, children prefer to use the noun-internal cue (Gagliardi, 2012;Culbertson & Wilson, 2013;Gagliardi & Lidz, 2014;Gagliardi et al, 2017).…”

“…A number of possible explanations for this surprising cross-linguistic tendency have been proposed, including an active bias against using external cues like semantics, when noun-internal phonological cues are available (Gagliardi, 2012;Culbertson & Wilson, 2013;Gagliardi & Lidz, 2014;Gagliardi, Feldman, & Lidz, 2017). In Culbertson et al (2017), we presented evidence from artificial language learning experiments with adults suggesting that the over-reliance on phonology may be due instead to the fact that phonological cues are generally available earlier than semantic cues (Carroll, 1999;Polinsky & Jackson, 1999;Demuth, 2000;Culbertson & Wilson, 2013;Gagliardi et al, 2017); learners acquire early representations of phonological dependencies (e.g., between a gendered determiner and a noun) before acquiring the semantic referents of nouns. Because the system is initially built on the basis of these phonological cues, semantic cues acquired later take time to be integrated into the system.…”

Brown, Smith, Samara, & Wonnacott (pre-print). Semantic cues in language learning: An artificial language study with adult and child learners.

“…Mulford attributes this to the unreliable and often multifunctional nature of phonological cues in Icelandic, which suggests that the consistency and robustness of cues in the input may be critical. Alternatively, young children may rely more on phonological cues than semantic cues simply because of their earlier availability (i.e., infants are exposed to word forms before they successfully acquire formmeaning mappings; Gagliardi, Feldman, & Lidz, 2017). A final possibility is that children disprefer semantic cues for reasons unrelated to their relative consistency or availability.…”

Parameter Inference for Computational Cognitive Models with Approximate Bayesian Computation

“…Logačev and Vasishth () optimized the parameters of a sentence parsing model using Nelder–Mead. Gagliardi, Feldman, and Lidz () optimized the parameters of a language learning model using the fminsearch. For situations where the optimized function has multiple optima, a combination of Nelder–Mead optimization and grid search has been used.…”