Natural language understanding systems had been developed during 20th century after computers were invented. This paper surveyed the computer systems in the view point of language cognition. Though many of the computer systems for natural language processing were not so successful nor cognitive, some of them such as Winograd's and Schank's were able to understand texts but the text topics were bounded to small worlds.
The paper explains that the difficluty was caused by a crucial feature of natural language, dependency to the context. It means that meanings of words in a text are determined when its context is given. This difficulty cannot be conquered by the development of computers such as their speed or memory amount.
In the last half fo this paper, a cognitive system research to natural language understanding is introduced. First part of the research is an electronic concept dictionary constructed with a large scale associative experiment which provides more than one hundred thousand words from human memory. The other part is a pulse neural network system on which the concept dictionary is implemented. The neural network system is applied to analyze metaphorical expressions. The results show a promising perspective of the cognitive appproach to natural language processing systems.