Predicting Conceptual Change during Naturalistic Reading with fMRI
Benjamin Schloss1, Ping Li1; 1The Pennsylvania State University
We report a model that can predict online changes in neural representation of science concepts resulting from reading expository texts about those concepts. We modeled the expected change using Google’s Word2Vec model, an error driven model of conceptual learning, and the Bound Encoding of the Aggregate Language Environment model (BEAGLE; Jones & Mewhort, 2007), a Hebbian model of conceptual learning. Then, we regressed the difference in the BOLD responses between the first and last time a participant looks at the same target word onto the corresponding changes in the computational models’ representational units before and after reading the text. BOLD responses were estimated using simultaneously acquired eye-tracking and multiband fMRI data. This paradigm allowed self-paced reading, while maintaining a high degree of precision in aligning the onset of an initial fixation, an index of the onset of word processing, to the BOLD response of a target word. Multiband fMRI was used to increase the temporal resolution of our data from the standard 2 seconds per volume to 400 ms per volume. Our best models achieve significant accuracy in approximately 75% of the 50 participants. Among participants whose data could be successfully predicted, average accuracy for pairwise left out words discrimination is 64%, but ranging up to 78% for some individuals. The results suggest that a modest amount of the online change in neural representations during reading can be accounted for by functions of distributional co-occurrence statistics.
Topic Area: LANGUAGE: Semantic