Proceedings of the Annual Meeting of the Cognitive Science Society

The pervasive use of distributional semantic models or wordembeddings is due to their remarkable ability to represent themeanings of words for both practical application and cognitivemodeling. However, little has been known about what kind ofinformation is encoded in text-based word vectors. This lack ofunderstanding is particularly problematic when distributionalsemantics is regarded as a model of semantic representationfor abstract concepts. This paper attempts to reveal the internalknowledge encoded in distributional word vectors by the anal-ysis using Binder et al.’s (2016) brain-based vectors, explicitlystructured conceptual representations based on neurobiologi-cally motivated attributes. In the analysis, the mapping fromtext-based vectors to brain-based vectors is trained and predic-tion performance is evaluated by comparing the estimated andoriginal brain-based vectors. The analysis demonstrates thatsocial and cognitive information is predicted with the highestaccuracy by text-based vectors, but emotional information isnot predicted so accurately. This result is discussed in terms ofembodied theories for abstract concepts.