Brain architecture and mechanisms that underlie language: an information-processing analysis

A neural network model for naming visual objects and their attributes, and understanding certain simple types of sentences has been presented. The model is based on neural processes rather than linguistic or symbolic constructs. The following are the major structural features of the model: 1. Memory stores are associative networks that perform an analysis of their inputs in real time. This analysis converts the input pattern into a "recognition pattern" that depends on the specific stored information and its location in the store. 2. The naming store is an analyzer network that converts the sensory encoding of a sentence into a "sentence pattern". The sentence pattern is obtained by averaging the recognition signals from several memory locations in the naming store. Because of the organization of words in the naming store, sentence patterns are an encoding of the structure of the sentences. 3. The sentence store is an associative memory store that recognizes sentence patterns and associates with each one an "instruction sequence" or "program" that specifies how the system should respond to the sentence. It is the enabling of access to this "program" that constitutes "understanding." Sentences are understood in real time, without the explicit grammatical analysis that is usual in sentence-understanding systems. Understanding a sentence involves two functionally distinct processes: (1) associating with its sentence pattern an instruction sequence that specifices what to do in order to generate an appropriate response to the sentence matching that pattern: for example, to look in the visual field for an object, determine its location, and so on, and (2) analysis of the tokens in the sentence that indicate specific attributes or entities to look for or determine.