Item

Abstract

Memory is short-lived. To avoid information loss, humans have to chunk continuous speech into discrete units, each containing multiple words. Earlier work has shown that auditory short-term memory is limited to 2-3 seconds (Baddeley et al., 1975). For language, a proposed window of 6 words (Frazier and Fodor, 1978) ​​translates to 2.4 seconds when assuming a rate of 150 words per minute (Tauroza and Allison, 1990). This time constraint may come from the limited duration of the underlying electrophysiological windows. It has been suggested that cycles of low-frequency neural activity serve the formation of multi-word chunks. Previous studies have shown that phase angles of oscillatory activity in the delta band (<4 Hz) predict the offsets of multi-word chunks (Meyer et al., 2016), in particular when these occur after 2.7 seconds (Henke and Meyer, 2021). In the current project, we pursue the possibility that this constraint does not only reflect time, but also relates to the amount of information per chunk. Prior studies have shown similar information rates (~39 bits/s) for syllables across different languages, implying that the brain can handle a constant amount of information per incoming syllable (Coupé et al., 2019; Pellegrino et al. 2011). For higher levels, the uniform information density (UID) hypothesis in psycholinguistics also suggests that speakers prefer utterances that convey/distribute information uniformly across speech signals (Aylett and Turk 2004; Jaeger 2010). Accordingly, there might be an upper limit to the amount of information that can be processed within a given time window. Here, we investigate if chunk boundaries are defined by the (accumulated) amount of information. In particular, we aim to examine if chunking-related neural activity correlates with chunk boundaries as defined by the cumulative sum of word surprisals. We will analyze chunking-related neural activity (i.e., delta-band oscillations) in the electroencephalography while participants listen to a naturalistic story. We define a chunk boundary as a time point when the accumulated sum of surprisal values–which are extracted from GPT-2 –exceeds a certain threshold. We compare different cut-off thresholds to determine if multi-word chunking indeed relates to the amount of processed information: We hypothesize to observe delta-band phase clustering at the end of a chunk for the surprisal threshold that defines an optimal amount of accumulated information. Additionally, we may also observe a closure positive shift (CPS) that reflects the closure of a chunk. Overall, our results will help to understand cognitive limitations on language processing. Specifically, we will uncover whether limitations of information processing determine multi-word chunks and are reflected by neural processing windows.