Natural Language Processing by Reasoning and Learning (draft for comment)

This paper proposes a new approach toward natural language processing in a general-purpose intelligent system. The major features that differs this approach from the traditional ones are: natural language understanding and production are not carried out by a separate module in the system, but by the reasoning-learning mechanism responsible for all kinds of cognitive functions; furthermore, when a sentence in a natural language is processed, its syntactic, semantic, and pragmatic aspects are taken into consideration altogether. This approach has been formalized in the form of a logic, and partially implemented as a reasoning system. Some preliminary results are introduced and discussed. 1 Two Chomskyan Assumptions 1.1 The modularity of language processing Since the very beginning of the study, natural language processing (NLP) has been recognized as a major aspect of artificial intelligence (AI). For example, by proposing his test for machine intelligence [Turing, 1950], Alan Turing implicitly assumes that the existence of other cognitive functions needed by intelligence can be recognized in a system’s language performance. Though the ability of using natural languages is widely acknowledged as necessary for (human-like) intelligence, its relation with other cognitive functions (such as perception, reasoning, learning, and action) is a controversial topic. Two well-known positions are represented by Noam Chomsky and Jean Piaget, respectively, and their opinions are contrasted in a debate in 1975 participated by them and some other researchers [Piattelli-Palmarini, 1980]. One key topic of the debate is the relation between language and intelligence. According to Chomsky, human language competence is basically innate and modular, as a