Presentation

Victoria Poulton¹, Máté Aller¹, Lucy MacGregor¹, Matt Davis¹; ¹University of Cambridge

How do listeners settle on the appropriate meaning of ambiguous words? Previous behavioural research shows that multiple alternative meanings of ambiguous words (e.g., “bank”) are transiently activated before the contextually-appropriate meaning is selected (Onifer & Swinney, Mem. Cognit., 1981; Swinney et al., Mem. Cognit., 1979). Ambiguous words therefore allow us to study how and when listeners incorporate contextual information during word recognition and meaning selection. Here, we contrast two neural computations proposed to drive meaning selection: (1) “sharpening” computations, in which representations of predictable contextually-appropriate meanings are enhanced or sharpened; and (2) “prediction error” computations, in which predictions are subtracted from the meaning of heard words.The goal of the present study is to determine the neural time course and source localisation of the computation of meaning during spoken sentence comprehension. Often, univariate analyses are employed to investigate the impact of different factors (e.g., lexical information, prediction, semantic constraint) on the amplitude of ERP/ERF components like the N400. The change in N400 response magnitude has been interpreted to reflect the magnitude of prediction error (Rabovsky et al., Nat. Hum. Behav., 2018). However, changes in response magnitude are consistent with multiple computations, including sharpening or prediction error (Aitchison & Lengyel, Curr. Opin. Neurobiol., 2017). Therefore, we will instead use a multivariate approach with representational similarity analysis (RSA) in order to assess neural responses to sentence-final ambiguous words and matched single-meaning controls. Multivariate analyses of selected and suppressed meanings allow us to distinguish sharpening and prediction error computations during the comprehension of ambiguous words, as these neurocomputational proposals make opposing predictions for the similarity of neural responses within critical pairs of trials. To date, simultaneous MEG/EEG data have been collected from three participants; and we intend to pre-register an analysis with N>30 datasets. Participants listened to sets of sentences, with examples like “While sailing down the river, she noticed the trees along the…”, which are biased towards one meaning of a target ambiguous word (here, “bank”). A sentence in the set can end in either the ambiguous word, a contextually-consistent synonym (e.g., “shore”), or an unpredictable word related to the non-selected, alternative meaning (e.g., “account”). For each critical comparison (i.e., ambiguous “bank” vs. synonym “shore” and ambiguous “bank” vs. alternative “account”), we will calculate the multivariate dissimilarity between the neural response patterns across samples (i.e., spatial dissimilarity over time) and across space (i.e., temporal dissimilarity across sensors/sources). Sharpening proposals predict low dissimilarity (i.e., high similarity) between neural responses to ambiguous words (“bank”) and contextually-consistent synonyms (“shore”) as the predictable meaning (<riverbank>) is shared. In contrast, prediction error proposals predict low dissimilarity for the other comparison – ambiguous words (“bank”) and alternative meaning words (“account”) – as both of these words will elicit a qualitatively similar prediction error for the unexpected meaning of the ambiguous word (<financial institution>). With the high temporal and spatial resolution of combined MEG/EEG, the results of this study will allow us to determine the time course and spatial dynamics of ongoing neural computations underlying meaning comprehension from spoken sentences.

Topic Areas: Meaning: Lexical Semantics, Speech Perception