Portal de Periódicos da CAPES

Relating Activity Contexts to Early Word Learning in Dense Longitudinal Data

2012; Wiley; Volume: 34; Issue: 34 Linguagem: Inglês

1551-6709

Brandon Roy, Michael C. Frank, Deb Roy,

Phonetics and Phonology Research

Relating Activity Contexts to Early Word Learning in Dense Longitudinal Data Brandon C. Roy Michael C. Frank Deb Roy The Media Laboratory Massachusetts Institute of Technology bcroy@media.mit.edu Department of Psychology Stanford University mcfrank@stanford.edu The Media Laboratory Massachusetts Institute of Technology dkroy@media.mit.edu Abstract guistic input across a narrower range of activities poses a simpler learning problem. The effect of overall linguistic input on lexical devel- opment was investigated by Huttenlocher, Haight, Bryk, Seltzer, and Lyons (1991). They were the first to docu- ment positive correlation between the quantity of child- directed speech and a child’s vocabulary size and growth rate. For individual words, increased frequency of use was also tied to earlier acquisition of those words; our own (Roy, Frank, & Roy, 2009) and other (Goodman, Dale, & Li, 2008) findings replicate this pattern. In ad- dition to frequency, words presented in single word utter- ances (Brent & Siskind, 2001) and with prosodic stress (Echols & Newport, 1992; Vosoughi, Roy, Frank, & Roy, 2010) are also acquired earlier. In addition to studying linguistic input, work in cross- situational word learning has investigated how words can be linked to referents through their consistent co- occurrence across a range of situations. In the face of referential uncertainty, a learner sensitive to the statis- tics of which words and referents co-occur can learn cor- rect word-referent pairings (Yu & Smith, 2007). But the idea of learning by gradually accumulating word-referent co-occurrences was challenged by Medina, Snedeker, Trueswell, and Gleitman (2011), on the grounds that the sheer number of possible pairings in everyday experience, coupled with memory limitations, leads to an intractable learning problem. Their data suggest a different learning strategy based on early binding between words and ref- erents, with errors corrected through natural processes of forgetting. While the natural environment is complex, it does provide structure notably absent from many laboratory- based word learning experiments. Bruner (1985) empha- sized the importance of naturally occurring, predictable formats of interaction that support communication. To study the role of formats in language acquisition, Bruner moved his research into the “clutter of life at home” via naturalistic, observational methods. One format that Bruner studied was the game of “peek-a-boo”, a recur- ring, rule-bound activity that occurs across a wide de- velopmental period. Language works in concert with the game to help reveal the meaning of words. With Bruner’s formats in mind, the goal of the present study is to investigate the activity structure of a child’s first years of life, how the child’s linguistic input links Early word learning is contingent on linguistic input, but a child’s linguistic experience is also embedded in the larger, natural structure of everyday life at home. We investigate the activity structure of life in the home of one young child, and link this structure to the child’s early word learning. Our analysis is based on the dense, naturalistic, longitudinal corpus collected for the Hu- man Speechome Project. To study activity structure, we apply probabilistic topic modeling techniques to the corpus. The emergent topics capture not only linguis- tic structure, but also spatial and temporal regularities indicative of coherent activity contexts. We consider the child’s word learning with respect to caregiver word usage frequency and word distributions across activity contexts. We find that frequency and consistency of use across context are predictive of age of acquisition. Words that are used more frequently and in more con- textually constrained settings are learned earlier, sug- gesting that activity contexts may be an important as- pect of the child’s natural learning environment and worthy of further study. Keywords: Language acquisition; word learning; non- linguistic context; topic modeling. Introduction Children’s early word learning is a remarkable achieve- ment, the result of powerful learning processes unfolding in the natural setting of a child’s first years of life. Cul- tural and individual variability in children’s early envi- ronments has led researchers to question the contribu- tions of the child’s innate faculties relative to the role of the environment. But to the extent that children are learning language, the environment must provide appro- priate conditions for learnability: There must be some consistent underlying structure for learning mechanisms to build upon. In lexical development in particular, the linguis- tic environment—what words a child hears, and how often—provides essential input for the young learner. Yet the child’s natural environment consists of other di- mensions in addition to language: spatial, physical and social dimensions, to name a few. Learners are exposed to their input in the rich, multimodal domain of every- day experience. In this work, we begin to investigate the activity structure of day-to-day life and its contri- butions to early word learning. Based on the idea that words and referents are more predictable in sufficiently constrained situations, we hypothesize that words asso- ciated with a limited range of recurrent activities will tend to be learned earlier. That is to say, consistent lin-