Quilian semantic network(ID:1913/cof001)Original frames systemRoss Quillian's formalism of semantic memory into the triplet of Frame/Slot/Facet Formalised and generalised by Minsky, it became the basis for frame languages Structures: Related languages
References: Introduction Behind the development of every new computer language there lies a set of problems and a set of programming structures with whose aid the problems can be managed. With Fortran the problem was to solve algebraic equations without the need for a great deal of I/O bookkeeping and without concern for detailed computer-word-packing statements. Behind JovrAL lay the command-control problem, which customarily dealt with complex data structures and the need to use every bit of computer memory efficiently. IPL grew in response to a need to use associative list structures of nonnumeric symbolic data in a computer. Lisp answered the need for a high-level functional language to handle recursive algebraic and symbolic structures. Comit was the machine translator's approach to handling natural language strings. In developing a special concept dictionary for storing and retrieving the meanings of words and phrases of English, the authors have found it desirable to use a complex network whose nodes are interrelated in several ways. Basically, the idea is that of a dictionary of English words in which each word has associated with it word-class information and lists of attribute-value pairs that define various aspects of its meaning and usage in terms of pointers to other words in the dictionary. In a data structure language such as Jovial, in addition to ordinary table structures several additional levels of associative coding are required to give easy access to the data without excessive costs in either space or processing time. Because of the many levels of associative linking required the authors decided to use Lisp, at least for early experimental work with the system. Advantages of Lisp extended beyond the ease of producing complex data structures; they also included the simplicity of producing complex, often recursive functions for maintaining and querying the dictionaiy. An additional advantage is gained in the fact that although Lisp is primarily an interpretive system it does allow for the compiling of a fast-running completed program. The most serious disadvantage of Lisp for our system is that in present versions1 it is limited to core memory for most uses. This limitation means that a dictionary of the type we are studying could not exceed two or three hundred words. Since we are aiming for an eventual vocabulary of from five to fifty thousand words, the limitation to core memory is intolerable. Either an expansion of Lisp will be required or the writing of a special language using auxiliary memory for handling cyclical structures in large complex networks will grow out of our experiments with the conceptual dictionary. Extract: The Problem The Problem The major shortcoming of all existing retrieval systems is their inability to handle anything vaguely resembling the meaning of words taken singly, let alone the meaning of the language strings that they comprise. Synonym dictionaries and thesauri have often been added but have proved but feeble makeshifts offering little improvement over the use of root forms of words alone. To the extent that automatic syntactic analysis has been available it has only emphasized the need for word and phrase meanings. Five years of Synthex research toward the development of question-answering systems based on natural language text have confirmed this inability to deal with meanings. In these five years many approaches have been attempted toward representing some aspects related to the meaning of words. Most have been unsuccessful. It was learned early that the use of a synonym dictionary did not greatly improve our understanding of text in response to questions. More recently it was realized that even a well-coded thesaurus was not an answer. At various times attempts were made to save syntactic contexts associated with words as possible representations of meanings; these too, although promising, did not appear to be a reasonable answer. With more recent research, particularly that of Bobrow [1964], Raphael [1964], Quillian [1965], and Thompson [1964], it has become apparent that, in addition to dictionary-type meanings, there is a need for something that can best be characterized as a knowledge of the world (e.g., Cows eat grass. Walls are vertical. Grass doesn't eat., etc.). Without something representing knowledge of the world it can hardly be hoped that a word or sentence can be understood. The consequence of this line of thought is the realization that the problem requires the development of a conceptual dictionary that would contain definitional material, associative material, and some representation of knowledge of the world. These three aspects of a word's meaning seem to be the minimum that will allow for enough understanding of English to make question answering even a reasonable probability. Extract: The Conceptual Dictionary The Conceptual Dictionary In a conceptual dictionary each word would be characterized by (a) a set of class memberships, (b) a set of attributes, (c) a set of associations, and (d) a set of active and passive actions. The set of class memberships includes statements of the form the/an X(noun) is a/an F(noun). Thus, "an aard-vark is an animal" or "an aardvark is a mammal" are both examples of statements giving rise to class membership characteristics. For many nouns the class membership set is one of the basic definitional aspects of the word. Attributes characterizing a word are such that if y is an attribute of x, then "x is y" is a true statement and the string "the yx" is grammatical. Thus if "scaly" is an attribute of "aardvark," then "an aardvark is scaly" is true and "the scaly aardvark" is grammatical. Associates of a word are in a loose part-whole relationship. If x has a y, then y is an associate of x\ thus "John has a wallet" and "John has a nose" provide the two associates, "nose" and "wallet," for John. The set of actions characterizing a word are derived from the verbs and their complements or objects that appear in context with the word. The two sentences "Natives eat aardvarks" and "Aardvarks eat ants" provide "eaten by natives" and "eat ants" as passive and active actions related to aardvarks. The idea underlying this schema is that the meaning of a word can be conceptualized in bands of closeness of relation. The class membership of an object is so closely related as to be an integral part of it or its perception. Attributes and things closely associated with a word are seen as important but less essential, while the actions associating one word and another may range from important to irrelevant. Extract: Toward an Operational System Toward an Operational System The conceptual dictionary briefly described above exists now as an experimental Lisp program in. 47k of core memory in the ARPA-SDC Q-32 Time-Sharing system. It is currently limited to a dictionary of 200-300 words and a relatively small set of program functions. What is required of even an early operational system is the ability to handle from 20 to 50 thousand words and a rather large set of functions for dealing with questions of increasing difficulty. Our expectations are that Lisp will be expanded into a system that uses up to four million words of disk to augment core and that we will be able to pay an increased cost of response time in favor of continuing to use Lisp for a large operational system. If that cost is prohibitive it will be necessary to produce a system tailored to the needs of the conceptual dictionary, and that will be able to use auxiliary memory efficiently to deal with a very large network of complexly linked words. Although it is our belief that new problems create the need for new languages, it is apparent that existing languages are largely sufficient for our language processing problems, but in many cases, especially among the list-oriented languages, they simply have not geared themselves to the large amounts of data and data processing required in this special field. Extract: Acknowledgments Acknowledgments. I wish to acknowledge my debt to the twenty or so people who have studied question-answering systems over the past decade. Their work is reviewed elsewhere [Simmons 1965]. Headers knowing the Quillian system will recognize that the result of the author's three years of acquaintance with Quillian was the appropriation wholeheartedly of his ideas insofar as the author was able to understand them. A special debt is also expressed to Fred Thompson for leading the author to an understanding of parsing directly into a data structure and for acquainting him with TEMPO'S forthcoming language system of associative cycles. Programming and detail design of the conceptual dictionary described in this paper were accomplished by John Burger. Extract: Discussion Discussion Salton opened the discussion with the comment that a system such as this is inherently not extendable. The system will operate nicely with various kinds of "fish," but will run into trouble when "whales" appear, since the system will not know how to deal with aquatic mammals. He compared the system to the "Baseball" system, in which one cannot go beyond a limited range of questions, e.g., one has trouble if a new team appears. Young objected to this view, saying he believed the two systems were quite different, and that the present system was in principle infinitely extendable and very general. He said he could conceive of a system one level above this one which could deal with generalizations and which would make handling propositions easier than with the present special programming. Burger said that work with higher level relationships was planned, but that the immediate extensions would be more trivial, in the way of inserting a great deal of knowledge about the world, of the kind any child has (e.g., "Walls are vertical."). Gora observed that the present system is based upon two forms of the verb "is" and the verb "has," and that more relationships were needed. Burger said the system was not in principle so limited. Gorn then asked how they would deal with the statement " 'Word' is a word." Burger had no immediate answer, though he thought they would eventually be able to deal with such cases. Responding to a question from Cheydleur, Burger said they had about 50 different functions. Mooers asked if there were any inherent features of Lisp which limited their work. Burger said that a fundamental limitation was the limitation to use of core storage alone in the SDC version of Lisp. They could not use disks. The second limitation was the inability to break out individual characters from the "atoms" of Lisp. This prevents au easy and direct way of treating the similarity between "dog" and "dogs." At present this relationship has to be put in as a separate piece of information. Mitchell then mentioned that for a very much larger data base a limitation will be the time required to pass the dictionaries through the machine. He said one of the big improvements in speed due to syntax-directed compilers was that they had less need to refer to a very large dictionary. Burger admitted that this was a very important problem, and one which is being considered. What can be done outside of Lisp is being studied, since a problem in using an auxiliary memory with a list-structure system like Lisp is that interrelationships between lists are broken up if just one section is brought from the auxiliary memory. So far, a good solution to the problem has not been found. Abstract: An experimental system that uses LISP to make a conceptual dictionary is described. The dictionary associates with each English word the syntactic information, definitional material, and references to the contexts in which it has been used to define other words. Such relations as class inclusion, possession, and active or passive actions are used as definitional material. The resulting structure serves as a powerful vehicle for research on the logic of question answering. Examples of methods of inputting information and answering simple English questions are given. An important conclusion is that, although LISP and other list processing languages are ideally suited for producing complex associative structures, they are inadequate vehicles for language processing on any large scale—at least until they can use auxiliary memory as a continuous extension of core memory. in [ACM] CACM 9(03) March 1966 includes proceedings of the ACM Programming Languages and Pragmatics Conference, San Dimas, California, August 1965 view details in Minsky, Marvin (ed.), "Semantic Information Processing" Cambridge, MA: MIT Press 1968 view details |