Combining fMRI Data and Neural Networks to Quantify Contextual Effects in the Brain

Does word meaning change according to the context? Although this hypothesis has existed for a long time, only recently it has become possible to test it based on neuroimaging. Embodiment theories of knowledge representation suggest that word meaning consist of a collection of attributes defined in terms of various neural systems. This approach represents an unlimited number of objects through weighted attributes and the weights may change in context. This paper aims at quantifying such dynamic meanings using computational modeling. A neural network is trained with backpropagation to map attribute-based representations to fMRI images of subjects reading everyday sentences. Backpropagation is then extended to the features, demonstrating how they change in different sentence contexts for the same word. Indeed, statistically significant changes occurred across similar contexts and across different subjects, quantifying for the first time how attribute weightings for the same word are modified by context. Such dynamic representations of meaning could be used in future natural language processing systems, allowing them to mirror human performance more accurately.