Temporal dynamics of words in daylong audio recordings

• In Event: Reorienting the study of language learning using naturalistic data

Abstract

There is increasing consensus that aspects of the language environment predict language outcomes (Lieven, 2016). However, there is less agreement regarding which measures of the environment predict which outcomes. There is uncertainty regarding which constructs serve as theoretically relevant predictors of language outcomes, and how these constructs should be operationalized (Montag et al., 2018). Thus, the challenges associated with identifying and operationalizing appropriate dimensions of the environment are both theoretical and methodological.

Daylong audio recordings of children’s auditory environments (Gilkerson & Richards, 2008; VanDam et al., 2016) can suggest new dimensions of language input that may be relevant for understanding language learning. The present work analyzed three fully transcribed daylong audio recordings of children’s language environments. One recording was collected and transcribed by VanDam (2018) and two were collected by Fausey and Mendoza (2018) and transcribed by my lab. This work addresses ways in which daylong audio recordings, specifically daylong dynamics of words, can help identify and operationalize dimensions of language input that may contribute to language learning.

The amount of language that children encounter is not consistent across the day (Figure 1). There are two consequences of this variability. First, the profile of speech throughout the day may vary across or within children. Hypothetically, two children could hear the same number of words, but with wildly different daylong distributions. Second, words are bursty: different contexts yield different language. The words uttered at mealtime, playtime or bathtime are not the same, so encountering speech in more different contexts will itself yield more lexically diverse input. A child’s profile of language density over time and the lexical diversity of a child’s language input may not be dissociable constructs. Relevant input dimensions related to the dynamics of language over time may only come from daylong recordings.

Next, individual words may be consistent or variable in the contexts in which they appear. Figure 2 shows the occurrence of different words in a day. Some words appear consistently all day (you, look) or in reliable contexts (nap, diaper) while other words appear in isolated contexts like book reading (price, robot). Some words may have different temporal distributions across days or families. For example, school is more frequent in Figure 2a but occurs in limited contexts. In Figure 2b, school is less frequent but appears more consistently throughout the day. Given well-established tendencies for variable versus consistent or blocked versus interleaved training items to affect learning (Estes, 1955; Carvalho & Goldstone, 2015; Vlach & Sandhofer, 2011), the temporal dynamics of words may be important for understanding why some words, categories, or sentences are learned before others. Ongoing experimental work investigates variability and consistency of sentence contexts and its effect on novel word learning, to tie observations from corpus analyses to human behavior.

The daylong dynamics of speech structure the sequences and timing with which children encounter the language that ultimately becomes the data for learning. Understanding these daylong dynamics may be important for understanding the input and the learning processes that underlie language learning.

Author

• Jessica L Montag, University of Illinois at Urbana-Champaign
  Presenting Author