Modeling Spontaneous Thought: A Network- and Langauge-based Computational Method

Modeling Spontaneous Thought: A Network- and Langauge-based Computational Method

This repository includes the data and codes to generate analysis results and figures. 

Data anonymization

To ensure participant confidentiality, we replaced all reported concepts and details—including names mentioned in their personal narratives—with alphanumeric codes. This anonymization process did not compromise the integrity of our analysis, and the results align with our published findings.

Dependencies

CanlabCore

cocoanCORE

spm12

Matlab 2022b

Conda 24.9.2

Docker version 24.0.7, build afdd53b

nvidia-docker2

xz (XZ Utils) 5.2.5, 

liblzma 5.2.5

Python 3.6

Tensorflow 1.15.0

Python library versions are provided by the requirements.txt file.

Docker environment

We trained the Transformer model in a docker container. We provided "fast_nlp_tf1.tar.xz" compressed by xz utils.

To build the environment, follow the command lines below.

$ xz -d -p {# of processes} fast_nlp_tf1.tar.xz

$ docker run --gpus all --name fast_nlp_env -v {path of this repo}:/Projects -itd fast_nlp:tf1.15.0-py3

Abstract

Spontaneous thought plays a crucial role in shaping affective traits and mental health. However, its dynamic and unconstrained nature makes it challenging to quantify and model effectively. To address this, we employed the Free Association Semantic Task (FAST) to obtain self-generated spontaneous thoughts, which were then analyzed using network modeling and natural language processing (NLP) to decode key affective content dimensions of self-generated thought, including valence, self-relevance, and time. In two studies (n = 213 and n = 137), we found that degree centrality and semantic distance between consecutive concepts were associated with the overall self-relevance level of self-generated thought. To capture the trial-by-trial dynamics of content dimensions, we developed a Transformer-based model, with which we extracted dynamic features and were able to predict individual differences in general negative affectivity. These findings highlight the potential of computational linguistic and network models to quantify spontaneous thought and predict affective traits, offering a scalable approach for real-time, automated mental health assessments while reducing reliance on retrospective self-reports.