Near, Far, and Frequent: Both Adjacent and Non-Adjacent Word Co-Occurrence Features Predict Age-of-Acquisition

• In Event: Poster Session 11
  In Poster Session: PS 11 Topic - Language, Communication

Abstract

Why do children learn some words before others? Age-of-acquisition (AoA) is highly non-random, and multiple factors predict variability of this index of learnability. Some proposed factors include usage frequency, salience, complexity, etc. Moreover, the contexts where words occur are semi-predictable. However "context" is not simple but multifaceted: It includes not only words adjacent to a target word - which often specify the word's suitable syntactic frames - but also words within the utterance but non-adjacent to it - which often are semantically or thematically associated. Do regularities in both of these contextual "scopes" predict when infants acquire words? Further, do both scopes predict AoA over-and-above frequency-based word properties known to partly predict learnability (e.g., Braginsky et al., 2017)?

To investigate this, from the CHILDES database we selecting naturalistic American-English corpora of caregiver speech to children under 48 months, regularized meta-tags (e.g., utterance boundaries), and reduced words to lemmas, for a meta-dataset of over 1.0 million utterances (4.5 million words) from 3,794 transcriptions. Word types were cross-referenced with AoA norms from Wordbank (Words & Gestures and Words & Sentences; Frank et al., 2016). Each CDI type was fit to a logistic curve, with the 0.5 crossing age set as AoA, yielding production AoAs for 656 words and comprehension AoAs for 383 words.

For adjacent ("syntax-loaded") context estimates, each pair of successive within-utterance words that occurred at least 1000x in the meta-corpus was represented in a normalized frequency value matrix (656x439). Each row then represents a word's adjacent-co-occurrence "profile." The matrix was dimensionally-reduced via PCA. The resulting PCs capture word categories that tend to co-occur in certain sequential pairs. For non-adjacent ("thematically-loaded") context estimates we used a modified COALS (Rohde et al., 2004) model: normalized frequencies of words pairs that co-occurred within-utterance, 2 to 5 words apart, at least 1000x in the meta-corpus were represented in a 656x1422 matrix. Again, PCA reduced this to discrete 'thematic feature' PCs. The procedure is schematized in Figure 1. Extensive analysis, to be presented, demonstrates that the PCs correspond to meaningful syntactic or thematic categories, and associated sentence-frames.

Regression models tested whether the contextual 'syntactic' and 'thematic' feature components (10 of each; validated by scree plots) predict AoAs for 656 words, above frequency-based predictors. For both comprehension and production norms, adding syntactic features, thematic features, or both improved prediction above frequency ('Baseline') information (Table 1). The full model (frequency+syntactic+thematic features) predicted adjusted R^2=.634 and .460 for production and comprehension AoAs, respectively. Syntactic features predicted production better than comprehension AoA, but thematic features predicted comprehension better. Earliest words were predicted by syntactic features related to simple noun or verb frames, or thematic features associated with eating or face-to-face play. The results suggest that beyond frequency, predictable contexts - both adjacent and non-adjacent - in which words occur additively contribute to word learnability.

Authors

• Lucas Chang, University of California, San Diego
  Non-Presenting Author

• Gedeon O Deak, University of California - San Diego
  Presenting Author