David L. Waltz
Research in computer understanding of natural language has led to the construction of programs which can handle a number of different types of language, including questions about the contents of data bases, stories and news articles, dialogues, and scene descriptions. This research draws on and has in turn had an effect on many other research areas, including software engineering, linguistics, psychology, philosophy, and knowledge representation. This paper provides a brief history and overview of the field, along with examples and explanations of the operation of several natural language understanding programs. The limitations of our current technology are discussed, and assessments are given of the most promising current research directions.
I would like to call programs that deal with large numbers of long natural language texts, produced not by programmers or linguists, but by the outside world, Realistic Language Comprehension (RLC) systems. These systems need to be very robust and flexible, capable of dealing not only with malformed input, but with their own inadequacies. Unfortunately, the standard methods used in our current programs are unlikely to help because they fail worst with the texts that are the easiest for people to understand.
Four radical techniques used in recent natural language programs that may lead to RLC systems are: (1) skimming texts to fill in high-level knowledge structures; (2) using content rather than function words to guide analysis; (3)