Semantic disambiguation is a difficult problem in natural language analysis. The meaning of words is ambiguous because it cannot be uniquely determined unless information of other words in the sentence is obtained. A possible strategy for semantic disambiguation is to determine the meaning of words at the end of the sentence. This strategy might cause combinatorial explosion of the number of possible meanings, if the sentence is long. So we think it is impracticable for a system to determine the meaning of words from a number of candidates after it has finished reading the whole sentence. A desirable strategy for semantic disambiguation is to accumurate information obtained during the analysis process of a sentence and disambiguate the meaning incrementally by using the information. In this paper, we propose such a computational model of natural language analysis, called incremental disambiguation model. In our model we introduce indeterminates which represent semantic ambiguity of words. Indeterminates enable underdetermined meaning of words to be incrementally determined by information obtained later. Primitive operation of the semantic disambiguation is realized by extended unification on indeterminates. Information used for disambiguation is considered to be constraints which restrict the meaning of words. Our model has the following features : analysing a sentence from left to right without any backtracking ; disambiguating the meaning incrementally by using constraints obtained during the analysis process. So it is clear that our computational model agree with constraint programming paradigm. Our computational model is regarded as an implementation of constraint programming by extending logic programming language Prolog.

The frame problem is very important in the context of knowledge representation of both humans and computers. The present paper discusses the frame problem for humans from a viewpoint of artificial intelligence. A major claim here is that the frame problem is unsolvable for humans as well. This claim is supported by several examples. Another major claim is that from a viewpoint of partiality of information, the frame problem must be discussed generally in a wide sense of the term, instead of being subdivided into the problem of description and the problem of processing. The unsolvability of the frame problem for humans should not be regarded as a limitation of human intelligence. Contrariwise, the flexibility of human intelligence is possible thanks to the fact that they cannot solve the frame problem; i. e., the fact that they make mistakes from time to time.

The deductive reasoning (DR), mainly used in expert systems, requires complete and consistent knowledge base. This brings about the difficult problem of knowledge acquisition. For supplementing DR, it is useful to introduce case-based reasoning (CBR) that explicitly uses past cases which are the expert's experiences with success/failure results. This paper clarifies the roles and problems of the domain knowledge used in CBR, and proposes an interactive model of CBR for flexible user's interaction with CBR's process under the incomplete domain knowledge. Based on this model, an interactive CBR system was implemented. This consists of the following functions : problem indexing, case retrieval, case evaluation, case modification, and case repair. An application of this system to a problem of machine adjustment has shown that even if the user cannot find useful adjustment method by the existing approach, the system can present an appropriate method plan.

This paper presents a very simple learning algorithm (hereafter called K-algorithm), which was developed in order to automate crop management in protected cultivation. K-algorithm assumes that the rule is formulated as, for example,"if and only if X(1) ∈R(1) OR X(2) ∈R(2) then do action", where R(i) means a certain subset of the domain of a variable X(i). In the if-part not only an OR-combination but also an AND-combination or both of them are allowed (there is a limitation on the form of the combination). Learning is mainly based on the analysis of the distributions of the measured X(i) values when the action was made and those when the action was not made. Although K-algorithm has a limitation in the form of the rules to which it can be applied, it proved to be very computation saving, memory saving and showed acceptable learning ability in learning crop management rules.