FOIA on Martin Luther King

Read from the scan by GLM-OCR; expect the odd misread word.

# The Understanding of Natural Language by Computers

Proposition: Computers are making long strides towards understanding natural language as used by human beings, and will eventually understand such language as well as many human beings do.

What do we mean by: computers? understanding language? natural language? And what is the evidence for this proposition?

The word "computers" here refers to powerful computers that have appropriate programs written by human beings plus the programmed capacity to improve their programs by using experience. An example of such improvement is Dr. A. L. Samuels' famous checker-playing program, which can learn from experience, and which plays far better checkers than Dr. Samuels himself can play.

The experience may consist of the following at least:

Answers from human beings to questions posed by the computer;

Differences between computed results and a priori specified results;

Signals from the environment, such as instrument readings;

information obtained by "looks" at the environment, as for example recognition of the character A, as in optical character recognition;

The interpretation of words, as with FORTRAN expressions.

Probably there are even more categories of experience which a powerful computer program can use to modify itself to become even a better program.

In regard to "understanding language", there are over 500 languages which computers have been programmed to understand, when written precisely according to stated rules: among them, BASIC, FORTRAN, COBOL, many kinds of machine language, etc.

"Natural language" is the ordinary language used by human beings, subject to the requirement that for a computer implicit understandings must be stated. For example, when a speaker exclaims "Ouch!", a great deal of the meaning comes from the listener's observations of the speaker. But a computer, like a blind man, cannot "see" the situation, and must be told more than the ordinary listener.

With some definitions taken care of, let us consider evidence for the main proposition stated above. This evidence includes the information published in several recent articles in Computers and Automation.

The article "Do What I Mean' - The Programmer's Assistant", by Warren Teitelman, in the April 1972 issue, described a "programmer's assistant", called "DWIM"; this was a "front end" or preprocessor to an interactive program for operating with LISP expressions. The front end was tolerant of the programmer's mistakes in typing, in low-

mistake occurred, it would inquire of the human programmer what he meant, suggesting the correctly spelled alternative, and so on. In this way the human programmer was relieved of much of the burden of expressing himself exactly and correctly the first time, and his efficiency in using the LISP interactive program greatly increased.

The three articles "Computer Programming Using Natural Language" by Edmund C. Berkeley, Andy Langer, and Casper Otten, in the June, July, and August 1973 issues, demonstrated the understanding by a computer program called GENIE of at least some sets of instructions in ordinary natural language. The vocabulary though small was free; and there was unlimited freedom in putting the words together, with about a 90% chance of complete understanding. The computer program that did this could be called 15% GENIE, because, as the authors emphasized, it was still in an early stage of development.

The article "Latest Computers See, Hear, Speak, and Sing - and May Outthink Man" by David Brand, in the October 1973 issue, enumerated many instances of computers (and robots equipped with computers) which could deal with concepts expressed in natural language. One of the programs mentioned could for example decipher natural language commands dealing with the stacking of blocks of various shapes, sizes, and colors.

The main tasks for understanding natural language appear to be the following:

Recognition of the framework of a sentence;

Recognition of the common meaning of groups of synonyms, the collection of words that "say the same thing";

Knowledge of context: the context is regularly specified to the computer, so that it does not have to deduce the context from "what is being said";

Knowledge of a limited vocabulary consisting of perhaps 300 to 500 words - knowledge in the sense that the computer program can attach meaning to the words either by themselves or in phrases, as for example it may attach the meaning of doing something three times to the numeral 3;

Capacity to accept variation in the way something is stated by a human programmer.

The three articles referred to above show instances of the achievement of all of these tasks separately. Combining all the achievements should not be too hard, especially since children as young as four years old show complete capacities to listen to many kinds of natural language, understand, and respond.