The traditional approach to computational models for language comprehension is based on the assumption that language is for communication between individuals. It also assumes that there is a common grammar and a shared semantic network among members of a language community, as demonstrated by professional qualifying exams. Computational models have focused on simulating language behaviour in humans, primarily simulating questions and answers (description rather than instruction, persuasion, or entertainment). Despite over 40 years of research, this approach has stalled.
This research paradigm starts with the assumption that the brain is not merely a hard-wired mechanism for information processing. It assumes that information processing in the brain involves both data and programs that are encoded, stored, and interpreted. The paradigm hypothesizes that there is an 'inner language' that gives an evolutionary advantage. In this research, 'inner language' means the central thought (like a statement in a computer program) that is expanded into precise muscle instructions. The paradigm hypothesizes that this 'inner language' controls actions that give an evolutionary advantage, i.e. that are better in feeding, fight, flight, and reproduction. We hypothesize that the 'inner language' facilitates faster and wider-scope learning for individuals. Communication between individuals with an 'outer language' is seen as secondary, and based on actions controlled by the 'inner language', such as gestures, posturing, and mating rituals. We further hypothesize that the mechanisms supporting this information processing must have developed over millenia.
The objective is to develop a theoretical and computational framework for research on language that fits evolution, with emphasis on action and perception. Longer term objectives include investigating how language comprehension and use develop from infancy to adulthood. One should be able to build a model that is competent to answer grade 9 exam questions. If one then exposed the model to the grade 10 curriculum, the model should be able to learn the material and to answer grade 10 questions. The long term goal of this research is to create such a model.
An interesting question is where such a model might start, and what are the starting capabilities. We propose that the starting point is when the baby is born and learns to develop motor skills. We shall come back to the question of what initial capabilities are required to make this initial inductive step for language learning work.