Language Science Lunch Talk - Kuan-Jung Huang
Language Science Lunch Talk - Kuan-Jung Huang
Understand child sentence processing in reading: computational and experimental approaches
A domain-general learning mechanism with language exposure, when operationalized with surprisal theory, yields overall impressive results in predicting adult language data. Here we take up the developmental trajectory issue, by comparing known child language processing results to LLMs of various training scales.LMs trained with 100M child-oriented text of words can predict many linguistic effects on child reading data. Regardless of the training scales, however, LMs are not-human-like w.r.t. thematic and semantic/pragmatic processing. We discuss implications for current LLMs as theory-proxies and propose a new experimental study for further theory evaluation.
Lunch served at 12:15 PM.
About: Kuan-Jung Huang is a postdoc associate in the Department of Hearing and Speech Sciences. He obtained his Ph.D. degree in Cognitive Psychology from the University of Massachusetts Amherst. His research focuses on the role of statistical regularities and syntactic structures in sentence processing and comprehension. He particularly specializes in eye-tracking during sentence reading and cognitive modeling.