Conversing with computers naturally.

Query a personal computer in plain English and the machine more often than not replies with a curt error message. A microcomputer's low speed and small memory strictly limit its capacity for handling grammar and English sentences. Such a computer can't figure out what users want unless they express their needs precisely in a language it can understand.

But as personal computer users struggle to retrieve and collate information from increasingly complex and unwieldy computer files and data bases, something more than memorizing lists of commands or following sequences of instructions is needed. To meet this need, computer programmers are now trying to introduce a more natural, conversational tone.

Software packages like INTELLECT (developed by Artificial Intelligence Corp. of Waltham, Mass.), which allow users at a terminal of a large computer to retrieve information simply by taping in English-language questions, are already popular. These programs can even deal with incomplete or ungrammatical requests.

One of the more ambitious schemes to shoehorn a large natural-language system into a personal computer is the Natural Access System, developed by Bozena H. Thompson and Frederick B. Thompson of the California Institute of Technology in Pasadena. Last week, the researchers described their approach at the National Computer Conference in Chicago.

The key is to use the input sentence, together with a dictionary and grammar table, to identify which procedures and pieces of data must be brought into the computer's main memory. This is done with an elaborate "paging" scheme that is closely linked to the processing or parsing of the sentece. As the sentence is parsed, small blocks or "pages" of relevant digital information are delivered from peripheral storage to the main memory.

"The user's input sentence is all that is needed to orchestrate the loading of the data and procedures necessary for the processing of that sentece," the researchers report.

Although the main memory may be small, the system can call upon practically an unlimited number of pages stored elsewhere. As a result, in answering a single question, it can, as needed, integrate text, pictures, data, display formats, statistical processing and information scattered throughout several files. A user can, for example, make a request like "Display a bar graph of pineapple production over the last five years in Hawaii" and, if all of the needed information is accessible to the computer, expect an answer without doing anything more.

The Natural Access System is now implemented and running on microcomputers such as the IBM Personal Computer AT with 640 kilobytes of internal memory. Tests show that this type of computer, for instance, given a list of 1,500 cities in one data base, takes about 20 seconds to answer: "What is the average population of cities?" Answers to some queries are available within 4 seconds. The Thompsons expect to demonstrate their system publicly in September.

The first and so far most successful company to come up with a natural-language interface for microcomputers is Microrim Inc. of Bellevue, Wash. Although not as comprehensive as the Thompson program, its "Clout" software processes database queries typed in free-flowing English. To understand a sentence, the program checks a set of "dictionaries" that describes what the data base contains and lists key words and "glue" words used to put sentences together.

A related product, Savvy (Excalibur Technologies Corp. of Albuquerque, N.M.), at least in its earlier versions, ignores grammar and vocabulary and instead relies on recognizing patterns -- matching strings of characters without analyzing their meaning -- to identify key words and phrases typical of the user. Although such a scheme is more vulnerable to error, it takes up so little memory that it fits on computers like the Apple II.