Speaker: Madison Bunderson @mbunderson
Affiliation: Stanford University
Title: Exploring brain-behavior connections in narrative engagement using EEG inter-subject correlation
Abstract (long version below): Engagement with narrative remains elusive and hard to characterize (Galbraith & Rodriguez, 2018). This work in progress presents electroencephalography inter-subject correlation (EEG-ISC) as a relatively new tool for studying engagement with naturalistic narratives (Dmochowski et al., 2012). We collected EEG and subjective ratings from 14 adult participants who listened to social and non-social auditory narrative excerpts. Preliminary results suggest alignment with previous EEG-ISC studies (Ki et al., 2016) and novel connections between neural data, behavioral ratings, and content analysis. The use of EEG-ISC as a temporally acute neural metric has potential for further study of ecologically valid narrative content and engagement.
Despite the long-standing assertion that engagement is an essential part of the reading process, engagement remains elusive and hard to characterize1. Theories of engagement vary widely and exist relatively separate from other theories of reading, with limitations in behavioral and neuroscientific measurement of this construct. Interdisciplinary approaches to conceptualization and measurement are necessary to expand our understanding of narrative engagement.
In this work in progress, we use electroencephalography inter-subject correlation (EEG-ISC) to study engagement with naturalistic audio narratives. It has been suggested that EEG-ISC indexes brain states of narrative engagement, defined as “emotionally laden attention”2. Underlying this approach is an analysis technique that derives multiple spatial brain “components” in which ISC is maximized. Recent studies have drawn connections between EEG-ISC and narrative coherence2, attention3, memory4, and knowledge acquisition5, and spatial components and corresponding correlated activity have been found to reflect distinct sensory and integrative aspects of auditory narrative processing3. But questions remain about the narratives themselves, particularly as we seek to explain findings from the empirical literary studies literature. For example, how does varying content across narratives–such as social or non-social information–influence engagement? Our study aims to elucidate relations between narrative features and measures of engagement.
Stimuli and participants
Two categories of auditory stimuli–social and non-social literary excerpts–are the focus of the current investigation. Three excerpts per category, each approximately five minutes in length (571-618 words each), were selected based on content or usage in previous research6,7. The social category contains Gilb, Berry, and Munro texts, while the non-social category contains LeGuin, Smithsonian, and Haskell texts. Participants (18-35 years old and highly fluent in English) are assigned two excerpts from each category in a counterbalanced fashion. We are in the process of collecting 36 usable datasets and present preliminary results based on 14 datasets.
128-channel EEG is recorded while participants listen attentively to each assigned stimulus once in its entirety, performing no other task as narratives are presented. After each narrative, the participant answers questions on a 1-9 scale to report their interest and enjoyment of the passage, as well as how compelling they found it. Thus far, an average of nine usable trials have been collected for each stimulus.
Content analysis of the selected passages was performed with LIWC and its inbuilt dictionaries8. We analyzed behavioral responses by computing means, Pearson pairwise correlations for individual stimuli, and category-level t-tests to each post-narrative question. We applied an established pipeline2 for computing ISC from cleaned EEG records, first pooling the data across stimuli to compute the optimized spatial components, and then calculating ISC on a per-stimulus, per-component basis as the correlation of each data trial against all other trials2. We report ISC summed across the three maximally correlated components (RC1–RC3) for each stimulus as done in previous published work2–5.
Overall, non-social texts used more perception words (e.g., “see”, “look”) and fewer cognitive (e.g., “know”, “think”) and social words (e.g., “care”, “friend”), with the exception of the LeGuin text, which used slightly more social words than the Berry text. Preliminary t-tests indicated no differences in enjoyment, interest, or how compelling each excerpt was by category (all p > 0.62), but preliminary per-stimulus pairwise correlations suggested differences in both interest and compulsion. Texts’ average ratings across all three questions were Gilb (6.07), Smithsonian (5.97), LeGuin (4.73), Munro (4.53), Haskell (4.27), and Berry (3.46). Spatial EEG components highlight both auditory and non-auditory regions, similar to other recent auditory EEG-ISC studies3,9. Group-averaged ISC was also broadly in the range of previous reports3: LeGuin (0.0548), Gilb (0.0496), Smithsonian (0.0421), Munro (0.0375), Berry (0.0293), and Haskell (0.0240). While preliminary, these results suggest potentially meaningful brain-behavior connections, as the same top and bottom three excerpts emerge across both behavioral ratings and EEG-ISC.