The Other Clutter: Verbal Clutter and Statistical Word-Referent Learning

• In Event: Lexical Development: environment and mechanisms

Abstract

A wealth of research reveals the power of cross-situational statistical learning in discovering word-referent mappings (see Smith & Yu, 2008; Yu & Smith, 2007). Many of these studies focus on how infants figure out to which of many objects do words refer (i.e., the visual clutter problem). Through corpus analyses, we illustrate descriptively that the verbal clutter to which infants are exposed may have been underestimated. We then experimentally test the tolerance of statistical word-referent learning by constructing a learning task that better resembles the verbal clutter infants face.
STUDY 1: CORPUS ANALYSIS
We analyzed audio-visual recordings of 37 parent-infant interactions (the 9-month-old Rollins Corpus; Rollins, 2003). We transcribed parent speech and identified referential utterances (utterances that were about at least one of the toy objects). The result was a corpus of 5,686 referential utterances (e.g., Fig. 1A). For each subject, we analyzed 50 randomly selected utterances and asked: (1) how many objects were referred to in each utterance, (2) how many different word types infants heard, and (3) how often was an object’s name heard. All analyses were repeated at the object level of analysis as well.
Results revealed that first, most referential utterances were directed to only a single object (M = 0.98, SD = 0.02; Fig. 1B). Second, infants encountered upwards of 90 different word types per 50 utterances (M = 91.08, SD = 20.03). Third, object names were rare, occurring in only a subset of referential utterances (M = 0.29, SD = 0.10; Fig.1D) and accounting for a small fraction of the total words infants heard (M = 0.08, SD = 0.03; Fig. 1E). Object-level analyses mirrored these subject-level patterns (Fig. 1B-E). Together, these results suggest that parents’ speech focuses mostly on one object at a time. However, to find the object’s name, infants must sift through a lot of verbal clutter.
STUDY 2: CROSS SITUATIONAL WORD LEARNING EXPERIMENT
To test whether human statistical learning is sufficiently powerful to overcome this verbal clutter, we created an artificial word learning experiment based on the statistics of the Corpus analysis. Specifically, we translated a random selection of six referential utterances for each of 12 objects into a string of novel words (Fig. 2A). On each trial, adults (N = 20) saw a single novel object and heard the string of words. After the learning phase (72 trials), participants were tested on whether they learned the links between object names and their referents (Fig. 2B). Surprisingly, participants learned at a rate that exceeded chance performance, M = 15.4%, SD = 10%, t(19)= 3.22, p = .005 (Fig. 2C).
DISCUSSION
Statistical word-referent learning has come into prominence as an account of early word learning. Previous studies have revealed that human statistical learning can overcome the challenges posed by the uncertain and probabilistic nature of reference (see Vouloumanos, 2008; Yu & Smith, 2007). Here, we demonstrate that statistical learning in adults is sufficiently robust and tolerant to also overcome the verbal clutter characteristic of infants’ learning environments.

Authors

• Miranda N Long, University of Connecticut
  Presenting Author

• Sumarga H. Suanda, University of Connecticut
  Non-Presenting Author