Exploring self-generated thoughts in a resting state with natural language processing

Abstract

The present study seeks to examine individuals' stream of thought in real time. Specifically, we asked participants to speak their thoughts freely out loud during a typical resting-state condition. We first examined the feasibility and reliability of the method and found that the oral reporting method did not significantly change the frequency or content characteristics of self-generated thoughts; moreover, its test-retest reliability was high. Based on methodological feasibility, we combined natural language processing (NLP) with the Bidirectional Encoder Representation from Transformers (BERT) model to directly quantify thought content. We analyzed the divergence of self-generated thought content and expressions of sadness and empirically verified the validity and behavioral significance of the metrics calculated by BERT. Furthermore, we found that reflection and brooding could be differentiated by detecting the divergence of self-generated thought content and expressions of sadness, thus deepening our understanding of rumination and depression and providing a way to distinguish adaptive from maladaptive rumination. Finally, this study provides a new framework to examine self-generated thoughts in a resting state with NLP, extending research on the continuous content of instant self-generated thoughts with applicability to resting-state functional brain imaging.

Neural Representations of Self-Generated Thought during Think-aloud fMRI

“…We briefly introduced the concept of self-generated thoughts and asked a few questions to deepen their understanding of this phenomenon. All introduction and interview information was as previously reported (Li et al, 2021). We informed participants that the reporting periods would be audio-recorded throughout although we would not listen to their reports concurrently; that the audio-recording file names would not contain any personally identifiable information; and that we would only use the text for scientific research and would not disclose any of their personal information.…”

“…To test whether there was a relationship between brain activities related to self-generated thought reporting and thought content, we calculated correlations between the indicator of thought content divergence and brain activation maps. Similar to our previous study (Li et al, 2021), we first divided each 10-minute stage into two five-minute segments. Then, we calculated the sum vector of all self-generated thought episode vectors (768-dimensional token vectors from the final layer of BERT) to represent the content of the segment.…”

“…We used MRI to synchronize fMRI data throughout the process and recorded participants’ verbal reports with an MRI compatible adaptive noise canceling microphone. Our previous research has demonstrated that the think-aloud approach has good reliability and validity (Li et al, 2021). The verbal report did not change the characteristics of resting-state self-generated thoughts, including frequency and content characteristics.…”

“…Furthermore, the metrics calculated based on natural language processing (NLP) for verbal report data had good reliability and validity. Recently, Raffaelli et al also used the think-aloud method to quantify the content of resting-state self-generated thoughts, suggesting the ecological validity of the think-aloud approach in the application of resting-state cognition (Raffaelli et al, 2021).…”

“…By comparing the free reporting and non-reporting blocks we sought to examine the impact of speaking. To further illustrate a relationship between the activation of verbally reported self-generated thoughts and thought content, we calculated correlations between brain activation patterns and divergence in thought content using an indicator that was shown to have good reliability and validity (Li et al, 2021).…”