Àíäðóñ³â Í. Î.

×åðí³âåöüêèé íàö³îíàëüíèé óí³âåðñèòåò

 ³ì. Þ. Ôåäüêîâè÷à

Ô³ëîëîã³÷í³ íàóêè:

Åòíî-, ñîö³î- ³ ïñèõîë³íãâ³ñòèêà

Knowledge representation and knowledge processing using frames

Recent work in computational semantics has argued that the representation of meaning should not be restricted to simple word definitions [3, 134]. "Background knowledge" [4, 27] has to be included into these representations. The main aim of the knowledge representation system described here is to find a structure for this additional knowledge.

Knowledge related to words, but also to nonlexicalized items, can be stored in frames. In our system, slot names indicate different aspects under which an object can be seen, e. g. appearance, character, sociocultural norms. The values of these "super-slots" are represented in subslots with NPs as their names and VPs as their slot values. The use of verbs allows the encoding of many different relationships between the frame concept and the concepts represented by slot names.

In our frame system, single inheritance is used. However, unrestricted inheritance can result in a problematical information load at lower levels of the hierarchy. For example, if we consider that a Person is a kind of Organism, then Persons according to Konerding should inherit nearly 30 superslots by Organism, define another 11 superslots for themselves, and represented in a frame system like ours fill all these with lists of subslots as answers to the superslot- "questions".

Different frames can be chosen for a given item depending on context. The selected frame will then supply only the background knowledge needed in the particular situation. Konerding distinguishes two general frames for persons, the first one applying to persons in a certain state or with a certain property, the second one applying to persons with a certain profession. Minsky describes two frames for representing a generator from different viewpoints, the mechanical and electrical "subframes". This knowledge of the different roles people and objects can play is stored in frames called noun roles here. Every frame can be attached to a role slot of another frame as a noun role, and any role can be instantiated for itself like any other frame.

With respect to individual objects, this architecture allows the creation of an object (e. g. a Person) to which one or more roles can be attached dynamically (e.g. Politician and French Person). The object does not lose its identity (e. g. if we have named our person Jacques Chirac, this information is stored at the initial (person) level, so the name of the person remains Jacques Chirac whether we view him as a Frenchman or as a politician). The roles of the object can be viewed and compared independently. As the slot values of the role slots indicate the context in which the person or object plays the particular role, we will no longer be confronted with all the knowledge about the object at once.

At the class level, this architecture allows classes to provide one or more default roles for the neutral or non-specified context, and special roles for special contexts. Information overload caused by inheritance can thus be avoided by treating the information formerly contained in some slots as a noun role and only accessing it if necessary. For example, an Organism could have a role with the frame Self Reproducing Organism which stores all the information that would otherwise have been assigned to a reproduction superslot. Instead of inheriting that superslot without restrictions, the Person subclass will inherit the Self Reproducing Organism as a role. It can also convert it to a subtype of Self Reproducing Organism, e. g. Self Reproducing Person, filled with more human-specific information.

Research relying on a newspaper corpus showed that nation frames (Italian, Frenchman, German) focused on quite different subslots when analysed in a politico-economical context than they did in a neighbourhood or family context like the one analysed by van Dijk. For example, in the family context it is relevant what people eat and when and how they do that. In the state context, however, it is important what people think about politics, inflation, or national identity.

Consequently, the context has to be taken into account when frames are established for the frame system. As it is difficult to find a "neutral" context for a given word to occur, a concept or word is first analysed in different "special" contexts which are determined by key words and will yield different role frames of the concept. It can then be decided whether some of the slots of these role frames should be represented in the concept frame itself and/or whether one or more roles should be regarded as default roles.

Accordingly, the world knowledge stored in the frames and in the noun roles could be accessed by a text understanding system. The structure of the frames can also be exploited for other purposes like finding out resemblances between concepts, objects or some of their aspects (slots, roles), or discovering structural regularities of metonymies or metaphors.

 

References:

1. Dijk, Teun A. van. Communicating racism. Ethnic prejudice in thought and talk. Newbury Park, CA et al.: Sage, 1987. – 356 p.

2. Konerding, Klaus-Peter. Frames und lexikalisches Bedeutungswissen. Untersuchungen zur linguistischen Grundlegung einer Frametheorie und zu ihrer Anwendung in der Lexikographie. Tubingen: Niemeyer, 1993 . – 298 p.

3. Minsky, Marvin. A framework for representing knowledge. The psychology of computer vision. New York et al.: McGraw-Hill, 1975. – P. 211-277.

4. Pustejovsky, James. The Generative Lexicon. Cambridge. MA/London: MIT Press, 1995. – 257 p.

5. Yeap, Wai-Kiang. A Sketched Computational Theory of Language Comprehension. // Proceedings of the Twentieth Annual Conference of the Cognitive Science Society. – Mahwah, NJ/London: Lawrence Erlbaum. – P.1170-1175.