Vector representation of words and concepts

How could a neural vector represent the words of language, both in terms of their form and their meaning ? At the perceptual level, the lecture will examine the hypothesis that phonemes and syllables are represented by neural codes " factorized " into vowels and consonants, and even phonetic features. This decomposition seems to exist even in very young children, and gives neural reality to the phonetic features postulated by linguists. At the level of meaning, latent semantic analysis and, more recently, the Word2Vec and GloVe models assign each word a concept vector in a high-dimensional space, and manage to capture some of the human judgments of similarity between concepts. We will examine several recent experiments that assess how well these models match the neural encoding of concepts in human cortex.