Poster #113 - Decoding Infants' Pre-activation of Lexical Representations via fNIRS

• In Event: 1-061 - Poster Session 02
  In Poster Session: PS 02 Section - Language, Communication

Integrative Statement

Verbal prediction is claimed to be an important mechanism supporting language development (Elman, 1990; Dell & Chang, 2014). In line with this view, prior findings indicate that infants can predict upcoming words during language processing (Reuter et al., 2018). Prior studies also observe a positive correlation between prediction and vocabulary size, further suggesting that prediction supports language development (Borovsky et al., 2012; Reuter et al., 2018). However, previous research investigating verbal prediction has relied on behavioral methods. Thus, a current empirical goal is to determine whether infants use top-down connections to pre-activate lexical representations during language processing. Here, we extend prior fNIRS decoding methods (Emberson, Zinszer, et al., 2017) to decode infants’ lexical representations and to investigate whether infants are able to pre-activate lexical representations.
In Experiment 1, we tested infants 15 to 18 months old (n=14). Infants viewed lexical trial movies (4.5 seconds duration). Each lexical trial included a looming object (e.g., a cookie) and four repetitions of the target noun. Between each lexical trial, infants viewed an inter-stimulus movie, each of which included neutral visual and auditory stimuli (e.g., falling snowflakes and soft music). We used a LABNIRS Shimadzu system (74 channels, distributed across occipital, temporal, frontal and parietal ROIs, infant fibers) to obtain neural signatures for 4 early-learned words (ball, cookie, cup, and shoe). Using channel stability, defined at the group-level, the top 50% of channels were included for analysis. Oxygenated hemoglobin was averaged in each channel from 4 to 8.5 seconds after stimulus onset (Figure 1A). A group model was created for all infants but one by averaging neural signatures of each lexical trial (Figure 1B). This group model was then used to evaluate the lexical trials for the left-out infant using Pearson correlations. This process was iterated for all infants. Lexical items were decoded at a rate of 68% accuracy. This finding establishes that lexical decoding is possible for infants using fNIRS.
In Experiment 2, we extended this decoding methodology to investigate verbal prediction. A new group of infants (n=5) viewed the same lexical trials and inter-stimulus movies as in Experiment 1. Additionally, infants viewed predictive trials, which included a neutral visual stimulus (i.e., a sock puppet video) and sentences that are semantically-related to the lexical trials (e.g., Let’s have a drink! Have some water! Have some juice!). We replicated the decoding findings from Experiment 1, using identical parameters, with a decoding accuracy of 80%. We then used the lexical group model to decode the corresponding predictive sentences (Figure 1C) with an accuracy of 60%.
In sum, these experiments make both methodological and theoretical contributions: First, it is possible to decode infants’ lexical representations using fNIRS decoding methods. Second, these methods provide a promising means of evaluating verbal prediction in infancy and determining to what extent infants pre-activate lexical representations. Future work will use fNIRS decoding to investigate verbal prediction among infants at-risk for language impairment in order to uncover the cognitive and neural origins of language delays.

Authors

• Claire E Robertson, Princeton University
  Presenting Author

• Tracy Reuter, Princeton University
  Non-Presenting Author

• Ben Zinszer, University of Texas at Austin
  Non-Presenting Author

• Lauren Emberson, Princeton University
  Non-Presenting Author