| Language has been considered as one of the hallmarks of intelligence for centuries. Therefore language use has become the most important test criteria for artificial intelligence. But, in spite of impressive advantages in computer technology over the past 50 years, natural language still poses a challenge for AI workers. Neither Classical AI systems nor connectionist nets or hybrid systems have produced language understanding in machines.; I suggest that this failure is, at least in part, caused by focusing on some features of the language used by a competent adult speaker: abstract reasoning, problem solving, and information processing. Such an approach does not appreciate how complex human language is.; Terrance Deacon's account of language highlights the importance of symbolic competence. Only a creature that understands how arbitrary symbols (words) refer to objects in the world and to other symbols can understand the meaning of these symbols. This calls into question one of the tenets of AI: that symbols are the atomic building blocks of language. Psycholinguistic research over the past 50 years reveals that human children acquire the ability to comprehend and to produce language in a multistage process that includes sense perception (mapping from speech sounds onto visually perceived objects), statistical learning, demonstration and encouragement by other language users, and continuing practice. AI models need to account for all of these processes and find a way to replicate the social aspect of language.; Finally, it has been suggested that language, like other higher cognitive abilities, needs to be embedded into a physical system that interacts successfully with its environment. Promising research has shown that it is possible to model some aspects of simple interactions of natural organisms and create "artificial life". Whether or not these systems can eventually account for all aspects of human cognition remains to be seen. |
