Prediction and Memory Retrieval in Dependency Resolution

Abstract

Successful language comprehension requires the rapid deployment of working memory resources alongside the capacity to predict upcoming linguistic input. While previous research views these as competing factors, this dissertation explores a unified theory of processing complexity and evaluates the interaction between memory and prediction. The evaluation focuses on how language-users deploy these factors to form long-distance dependencies in Mandarin. Specifically, I investigate how memory retrieval of a target word is affected by: (i) the time elapsed since the word first appears, (ii) interference from a neighboring distractor word that shares some linguistic features, and (iii) linguistic expectations of the target word. Neuroelectric signals of the human brain during naturalistic language comprehension were acquired by two electroencephalography (EEG) experiments. The experiments examine the resolution of noun-phrase ellipsis and subject-verb agreement using, respectively, carefully designed experimental stimuli and a naturally occurring audiobook story. The data are analyzed in terms of their fit to the quantitative predictions from computational models of these expectation and memory retrieval processes. This approach allows for a comparison between predictions of a symbolic cognitive model and a non-symbolic large language model. I report the first ever empirical evidence of the modulation of memory retrieval by linguistic expectations with a controlled experiment. I then report the first cortical electrophysiological evidence of the memory effects during naturalistic story listening, and suggest that interference modeled with cue-based working-memory retrieval framework may generalize to more everyday comprehension situation. The primary contributions of this work are, first, to unveil the biological underpinning of cognitive operations essential for how people understand complex sentences in a way that generalizes across languages and second, to contribute to methodological advancement in combining computational modeling and cognitive neuroscience to study naturalistic language comprehension in real time which can be generalize to real-world situations.