An Investigation Of Semantic Representation In Human Brain And Computer Science

We can acquire semantic representations by experiencing or reading texts.These learning experiences shape and change our semantic system in our brain.In computer science(CS),the semantics of a word can be learned from large text corpus and represented as a vector in the high dimensional vector space models.Can computer models predict neural activity of language comprehension? Which vector space CS model of semantic representation is closer to the ‘ground truth'--the semantic representation in the human brain?To investigate these issue,three CS semantic models were selected.CS semantic models computed the cosine similarity between vectors of two words as the relative semantic distance of a word pair,which was used as the index of the model's semantic representation.EEG(electroencephalography)semantic priming experiment was carried out,and the N400 ERP component was obtained as an index of semantic distance representations in the brain.We focused on the N400 component as it can provide a time window to observe the relationship between human semantic representation and CS models.We examined correlations between CS measure of cosine similarity and the neural measure of N400.Our results showed that the CS semantic models can predict the N400 component,which suggests that the CS model semantics representation capture the fundamental features of semantics in human brain--the semantic representations are similar in the CS models and in the brain.Further analyses found that the GloVe model best approximated the semantic representations in the human brain.Our research built a new framework to evaluate representation similarity between complex systems.