Abstract
Mutual understanding during a conversation is an extremely fast and efficient process : we can process three words per second, often more. However, this observation is not consistent with laboratory experiments showing that processing a single word can take up to a second. The speed of processing is explained by our ability to predict what the speaker is going to say, in a way similar to language models. Today, there is no global model for integrating this phenomenon of prediction-based facilitation into a classical language processing architecture (from phonetics to syntax to semantics). I'll present the foundations of such a model, explaining how superficial (facilitation effects) and deep (difficulties) processes coexist. This architecture is based on a central mechanism, prediction, which I will describe from both computational and neurolinguistic perspectives. This approach is based on results obtained from recent theories in cognitive science (" prediction-by-production ") and neuroscience (" predictive coding "), which suggest that participants in a conversation use the same mechanism to produce and understand speech.
