Two types of approach are possible for construction of object-oriented application systems : one is to identify a set of objects on the basis of a predefined problem solving framework, and the other is to design and implement a framework within given objects. Following the former methodology, we designed a framework for language processing in a speech understanding system, and then realized it as ASP (an association-based parser). Previous object-oriented language processing systems have emphasized their object-oriented representation of linguistic knowledge. But, in the case of speech understanding, control knowledge also plays a crucial role, where ambiguities are increased drastically by acoustic-phonetic errors as well as homonyms. In ASP, not only linguistic knowledge but also control knowledge is modularized into classes. This makes it easy for developers to realize various control strategies in ASP. ASP has been implemented on a Symbolics Lisp machine using Flavors, and almost 2O00 classes are involved. In particular, to enhance the extensibility of its knowledge base, ASP exploits layered software architecture, which assumes different types of programming skill for each layer. Currently, ASP is in operation as a language processing subsystem of a speech understanding system SPURT-I, and its performance has been evaluated with a 1O0O-word vocabulary. This paper gives a brief description of the framework of ASP, then explains its object-oriented implementation. The programming environment is discussed in terms of its extensibility, and the right level of object-oriented programming paradigm is clarified.

This paper proposes autoepistemic logic for two agents, which is a natural extension of Moore's autoepistemic logic. This logic formalizes beliefs which two ideally rational agents with mutual communications can infer by the extended introspection from their initial beliefs. In this paper, these beliefs are characterized by two concepts, stability and groundness. A relation between these beliefs and their initial beliefs is also investigated. Furthermore, this logic is applied to representation of hierarchical knowledge and to characterization of the reasoning based on such knowledge. This representation can be thought as a logical formalization of Minsky's frame model for knowledge representation.

Autoepistemic logic for two agents, which is a natural extension of Moore's autoepistemic logic, formalizes beliefs which two ideally rational agents with mutual communications can infer by the extended introspection from their initial beliefs. This paper investigates a decision procedure for autoepistemic logic for two agents. The procedure is based on the tableau method which is semantical and widely used for many kinds of logic such as modal logic and nonmonotonic logic. This paper also illustrates the reasoning of hierarchical knowledge by the procedure.

To improve the quality of machine translation systems, we should step toward the deeper analysis at the conceptual level. Developing the machine translation systems with deeper analysis requires the dictionaries including following information ; the set of conceptual items, the mapping relation between the surface words and the conceptual items, and the mapping relation between the conceptual items of the source language and that of the target language. There are several researches to compile such dictionaries. Japan Electronic Dictionary Research Institute (EDR) is now compiling such dictionaries on a large scale. Nirenburg, et al. at Carnegie Mellon University has proposed a systematic method to construct a conceptual dictionary. These attempts try to compile the dictionary by hand with the help of software tools. However this approach suffers from the problems such as huge amount of manual labor, the unstable result and so forth. Unlike this approach, the paper proposes a method to extract the information about the conceptual items from a pair of machine readable bilingual dictionaries in an automatic way. It is very difficult to compile the complete dictionary in a fully automatic way. The results of the method may require some refinement and modifications by human. Our goal is rather to automate the compilation process as much as possible and to decrease manual labor. In the paper, we make an approximation in that each word sense defined in the bilingual dictionary is considered as a conceptual item. Since each word sense has the proper translations in the bilingual dictionary, the above approximation is reasonable in terms of word choice in the translation, and we can easily get both the set of conceptual items and the mapping relation between the surface words and the conceptual items. The most difficult thing is to get the mapping relation between the conceptual items of the source language and that of the target language. The paper focuses on this issue. We introduce three types of translation circuits. The translation circuit is a tuple which consists of four elements, that is, a headword of both the languages and one of the word sense of both the headwords, with the condition that the word sense of one language should have the headword of the other language as a translation. We assume that the word senses in a translation circuit represent the same concept, that is, there is a mapping relation between the conceptual items (word sense) in a translation circuit. The paper describes the outline of a preliminary experiment conducted to verify this assumption. The results of the experiment are promising and some remarks are also given. We conclude the paper with pointing out the possibility by extending our method to construct the set of conceptual items which can be shared by more than two languages.

Achieving a conjunctive goal is a difficult problem in robot planning. Subgoals of a conjunctive goal can't be performed by arbitrary order because the mutual interferences exist in the processes of performing subgoals. For deriving such partial orders, we propose a new technique for planning by using temporal reasoning proposed by Allen. In our system, such a conjunctive goal composed of only two subgoals is called as minimal conjunctive goal, and given conjunctive goal can be regarded as a set of minimal conjunctive goals. All minimal conjunctive goals of a given circumstance can be obtained from production rules for that circumstance. And orders of performing subgoals of those minimal conjunctive goals can be derived by using temporal reasoning. Planning proceeds by utilizing these orders and so the backtracking of planning can be avoided, and the planning algorithm is very simple. In Allen's model, the time and space complexity in a full connected interval network is prohibitive. However, the size of the interval networks in our system is limited for minimal conjunctive goals, so that it is no necessary to worry about the complexity of time and space. In this paper, we will describe the deduction process of partial orders of performing subgoals of minimal conjunctive goals at first, and then show the algorithm of the main part of planning system. Some examples also will be shown to clarify the efficiency of our system.

Hypothetical reasoning is a basic technique for building AI system based on incomplete knowledge. Knowledge bases include hypotheses that are not always valid. Systems for hypothetical reasoning usually take a lot of time because they need to maintain the consistency of knowledge bases. This paper presents an efficient method for hypothetical reasoning on a forward-chaining system with the assumption-based truth maintenance system (ATMS). The forward-chaining system consists of a compiler of Horn clauses and normal defaults into a Rete-like network and a Rete-based inference engine. The Rete-like network as a flow graph consists of a root node, one-input nodes, two-input nodes and terminal nodes. The Rete algorithm is an efficient method for matching a large collection of objects with many conjunctive patterns. The inference engine is extended to reason in multiple contexts without conflict resolution. It gives justifications that are propositional Horn clauses to the ATMS. In the ATMS, each datum is labeled with a collection of sets of assumptions where the datum holds. The feature of a proposed reasoning method is that the inference engine gives intermediate justifications to the ATMS and stores intermediate dependent assumptions of two-input nodes, allowing faster hypothetical reasoning. By means of this method, the APRICOT/0 system for hypothetical reasoning has been implemented on the PSI-II machine (Personal Sequential Inference machine) in ESP (Extended Self-contained Prolog). An experimental result shows that APRICOT/0 is about seven times faster than a system that neither gives intermediate justifications to the ATMS nor stores the intermediate dependent assumptions under a tested knowledge base. The compared system, called SCS, is a simple combination system of a conventional inference engine and the ATMS. The cost of total ATMS label computations on APRICOT/O is generally less than the cost of the compared system in the following cases : (1) When the Rete-like network includes a two-input node shared by some clauses or defaults. (2) When a one-input node passes tokens to its successor that is a successor of another two-input node.

This paper depicts the integrated system Neuro-Prolog/II which is a synergistic cooperation of both Prolog interpreter C-Prolog and neural network simulator SunNet, each focusing on different aspects of the problem. By using Neuro-Prolog/II, users can deal with both logical information (i. e. laws, well-defined knowledge and logical rules) and non-logical information (i. e. human sensibility, ill-defined knowledge and taste) concurrently in a single system. In Neuro-Prolog/II, logical information is handled in Prolog and non-logical information is handled in neural network. This paper discusses how they may be integrated to provide a more unified and comprehensive treatment of these kinds of information. Prolog and neural network communicate with each other by execution of P-rule and N-rule. P-rule is a conventional rule used in Prolog, whereas N-rule is newly proposed for Neuro-Prolog/II to activate its corresponding neural network. The syntactic structure of N-rule is "head : ～body" which is very similar to the syntax of P-rule. We also describe a general protocol for interaction between P-rule and N-rule. If a goal can be unified with the head of N-rule, the neural network whose input layer and output layer are respectively assigned to the body and head of N-rule is activated. Since, both P-rule and N-rule can activate each other in Neuro-Prolog/II, this protocol enables Neuro-Prolog/II to deal with problems containing both logical and non-logical information in an unified manner. Furthermore, we have also developed a trial application system in Neuro-Prolog/II to demonstrate the advantages of Neuro-Prolog/II in real world problems. The trial system selects appropriate color arrangement of an outline picture among candidate colors. In the trial system, images of colors are handled in N-rules and heuristic rules of color arrangement are described in P-rules. As a result, a lot of features of human sensibility (i. e. individuality) can be handled in logic programs by using Neuro-Prolog/II. Neuro-Prolog/II is now running on SPARC station 330.

Although the rule composition can reduce the number of recognize-act cycles of production systems, it does not always improve the performance. Many works concerning this subject have recently been reported in the area of deductive learning. But they address only goal-driven and backtracking problem solvers. This paper makes some experiments on the influence of rule composition upon the performance of data-driven and non-backtracking production systems. We added a composing module to the OPS5 system and examined the impact of composites under the use of the RETE pattern matching algorithm. Several related works are also discussed.

We argue that explanation-based generalization as recently proposed in the machine learning literature is NOT essentially equivalent to partial computation, a well-known technique in the functional and logic programming literature. A training example plays an essential role in explanation-based generalization.