skip to content

Prediction and neural computations during ambiguous word comprehension: a preregistered MEG/EEG study

Victoria R. Poulton, Máté Aller, Lucy MacGregor, and Matthew H. Davis

MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, UK

How does the brain determine the relevant meaning of ambiguous words from unfolding spoken language? Here, we contrast two theoretical accounts of meaning computation: sharpening, in which the brain combines priors and input to sharpen predictable information in the stimulus; and prediction error, in which the brain computes the difference between priors and input, thereby representing the unpredictable information in the stimulus. Both computations are integral to various proposals which employ predictive coding and Bayesian inference in their explanations of spoken language comprehension. In this preregistered study, we capitalized on the multiple meanings of ambiguous words, using spoken sentences ending in either a predictable ambiguous word (e.g., bank), a selected-meaning synonym (e.g., shore) or an unpredictable alternative-meaning word (e.g., cashpoint). Ongoing brain activity was recorded using simultaneous MEG/EEG (N=34). We will present the preliminary results of representational similarity analysis (RSA), which will include timeseries of neural similarities, measured as the correlation (R) between activity patterns in the following contrasts: ambiguous-vs-synonym, ambiguous-vs-alternative, and ambiguous-vs-other (control with expected low similarity). Greater neural similarity (increased R) for the ambiguous-vs-synonym contrast indicates evidence for sharpening computations, as the ambiguous word and synonym share similar predictable semantic information. Conversely, greater neural similarity for the ambiguous-vs-alternative contrast indicates evidence for prediction error, as the unpredictable information activated upon encountering the alternative-meaning word (cashpoint) should be qualitatively similar to unpredictable information activated by the ambiguous word (bank). For full details, please see our  preregistration on OSF [view-only link]. Figures and example stimuli are attached below. 

Table 1: Example stimuli, presented to participants as continuous spoken sentences. All sentences we biased toward the ambiguous word, were pretested in an online visual cloze test.
Figure 1: A schematic of hypothesized neural activity patterns resulting from different computations. Sharpening computations in left panel will be indicated by activity patterns across sensors that are more similar between “bank” and “shore” (i.e., increased R for the ambiguous-vs-synonym contrast). Conversely, computations of prediction error will be indicated by activity patterns that are more similar between “bank” and “cashpoint” (i.e., increased R for the ambiguous-vs-alternative contrast).
Table 2: Hypothesized neural similarity based on the Account and depending on the time window under investigation (before word onset or after word onset).