Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Characterizing Semantic Verbal Fluency Tasks using Behavior, Computational Models and fMRI
Poster Session B, Friday, September 12, 4:30 - 6:00 pm, Field House
This poster is part of the Sandbox Series.
Martin Wegrzyn1,2, Bedia Vidua2, Valentina Alimenti2, Katalin Havas2, Patryk Markiewicz2, Johanna Kissler1,2,3; 1Department of Psychology, Bielefeld University, Bielefeld, Germany, 2CRC Linguistic Creativity in Communication, Bielefeld University, Bielefeld, Germany, 3Center for Cognitive Interaction Technology, Bielefeld, Germany
In a semantic verbal fluency task individuals are asked to name as many exemplars of a given category (e.g. “animals”) as possible. Beyond the number of produced words, the task can be analyzed regarding the quality or creativity of responses - for example if a speaker produces more original words or is able to switch between distant words. Semantic distance can be quantified using computational models based on word embeddings. Furthermore, behavioral and computational data can be related to brain activity, with the goal of inferring which cognitive processes are engaged during the task. Combining these different data types can allow a deeper understanding of how verbal fluency performance depends on the semantic category used and what role creative thinking plays. One hundred participants performed a behavioral version of the verbal fluency task with twenty-four semantic categories, along with the Alternate Uses Task (AUT) as a measure of creativity. We analyzed the number of words produced for each category and examined their correlation with the AUT. Word embeddings (fasttext) were used to represent the semantic categories and their distance to each other. A sample of ten participants performed the verbal fluency tasks while undergoing functional MRI. Voxel-wise analyses were used to relate activity strength to the number of words produced. In addition, Representational Similarity Analysis (RSA) was used to compare the similarity of brain activity patterns to the word embeddings. The behavioral data allowed to order the twenty-four categories regarding the average number of words produced. Speakers produced the highest number of words for the categories “first names” and “animals” and the lowest numbers for “amphibians” and “mountains”. Correlations with the AUT revealed highest correlations for the categories “fabrics” and “vessels”, while correlations with “precious stones” and “amphibians” were the lowest. Word embeddings allowed to group the categories into intuitive patterns, e.g. the category “body parts” being close to “clothes” which in turn is close to the category “fabrics”. The fMRI data showed that the fluency tasks consistently activate a left-lateralized perisylvian language network, including Broca's area. Correlations with the behavioral data showed that the activity of the fronto-parietal multiple demand network increased with the number of words produced. The similarity of whole-brain activity patterns revealed similar groupings to the ones found in the word embeddings, with the RSA showing a correlation of r=0.28 between model and neural data. The results highlight how behavioral and neural performance in verbal fluency tasks varies depending on the semantic category used. The fMRI data showed that activity in the multiple demand network is closely linked to the number of words produced. A next step will be to analyze how brain activity differs for categories depending on their relation to creativity measures, such as the AUT. Regarding the RSA, we plan a searchlight analysis to identify the brain regions where word embeddings and brain activity are most closely related. This will provide a more precise characterization of how verbal fluency tasks differ in their cognitive and neural underpinnings depending on the semantic category used.
Topic Areas: Control, Selection, and Executive Processes, Language Production