2018 Friday Session B 1400

Friday, November 2, 2018 | Session B, Conference Auditorium | 2pm

Individual-Outcome Corpus Modeling to Constrain Parameters of Statistical Learning Models
A. Buerkin-Pontrelli, J. Coffey, D. Swingley

Infancy researchers have shown that infants command an impressive repertoire of computational abilities that they could exploit to extract phonological and lexical regularities from speech.

Making quantitative predictions based on such experiments is difficult, because experiments place insufficient quantitative and implementational constraints on learning models. It is hard to generalize from small, concentrated samples of a regularity (the lab) to huge, diffuse samples (the world). The usual remedy is to implement a computational model over a corpus, where the model exemplifies intuitions that governed studies’ stimulus creation, and report whether the model succeeds in learning language structure. More successful models are considered better.

But infants are probably not optimal learners. Here, we instead constrain a model by assessing its ability to predict individual children’s learning of specific words, based on those individuals’ experience.

We phonologized and syllabified seven moms of the Brent & Siskind (2001) IDS corpus, assigning consonants to syllables (including across word boundaries) according to probabilities influenced by maximal-onset, stress, and sonority (Prince & Smolensky, 1993; Swingley, 2005). The corpus was then passed through a statistical clustering algorithm similar to Swingley (2005), where adjacent units with high mutual information (MI) and frequency were iteratively bound together. Under various parametrizations this yielded outputs of English words, part-words (under-combinations), and non-words (over-combinations). For each infant, we ask whether our word-finder’s identifying a particular word in a mother’s speech predicted that infant’s understanding of that word (15-month CDI).

Our baseline regression (Swingley & Humphrey, 2017), used no statistical clustering. We test this regression against versions provided, as a predictor, the binary outcome of whether a parameterization of our statistical clusterer found a given word in a given corpus. Baseline predictors were: total (log) frequency; (log) frequency in isolation; concreteness; word MLU; grammatical category. We implemented our word-finder across a range of MI and Frequency threshold percentiles (5^th-95^th), producing 19 separate glms.

The models with the best predictive power had low thresholds (45^th and 50^th) for assuming wordhood (50th: ß= 0.42 ± 0.19, z =2.2; 45th: ß =0.37 ± 0.19, z = 2.0; p’s <0.05). These parameterizations were good at predicting children’s knowledge, but poor relative to the gold standard: accuracies were 0.08 (45^th) and 0.09 (50^th) and recall scores were 0.31 (45^th) and 0.30 (50^th). (More restrictive parameterizations yielded accuracies around 75%.) The statistical word-finder boosted performance by identifying low-frequency words that infants knew, while correctly rejecting (by not identifying) frequent words that infants didn’t know (Fig 1, quadrants I and IV).

Our statistical word-finder aided prediction by capturing characteristics of words infants learn in natural settings. This was true despite the relatively simple design and frugality of the model. Our results corroborate recent work showing that simple, TP-based word-finding models can offer sound predictions for actual vocabulary outcomes (Larsen et al., 2017) and expand these findings using individually matched inputs and outcomes. Future work in this domain, we believe, should strive to include more individual outcomes to fine-tune predictions about infants’ language learning.