[PS-1.29] The Picture Guessing Game: Robust Active Statistical Learning of Syntactic Regularities

Trecca, F. 1 , Frinsel, F. 2 & Christiansen, M. H. 2, 1

1 School of Communication and Culture, Aarhus University, Aarhus, Denmark
2 Department of Psychology, Cornell University, Ithaca, New York

Language learning requires successfully integrating syntactic, semantic, and other cues to the meaning of an utterance. We present a new experimental paradigm that combines artificial language learning with sentence-picture matching, to investigate active statistical learning via multiple-cue integration. English-speaking participants were presented auditorily with sequences of nonsense words that followed non-English dative patterns: prepositional-dative (S-O1-prep-O2-V, e.g., kav-jux-ma-rus-sook) or double-object (S-O2-O1-V, e.g., kav-rus-jux-sook). For each sequence, participants were shown four images of three characters in different who-gave-what-to-whom scenes and instructed to click the picture matching the sentence. The language contained 9 nouns (depicted with images of human or animal characters) and 3 verbs (depicted as arrows of different colors). Feedback was provided for correct guesses by repeating the auditory sequence along with the correct picture. Performance was high (M = .76, SD = 1.8) and increased across trials (B = .16, p < .001), but less so for the syntactically more complex double-object sentences (B = -.17, p = .024). The robustness of learning was tested by introducing implausible referents (animal character as agents) and/or Brownian noise in the last 40 trials. Neither manipulation decreased learning. Thus, active statistical learning can support robust learning of distributional structure.