Interaction with computers using natural language has been a major goal of artificial intelligence. Though many natural language processing systems have been developed, most of them assume that input sentences are grammatically correct. However, when users communicate with the system, they often ues grammatically ill-formed sentences, especially in spoken dialogues. For example, the users omit some words, change the word order, or make some careless errors such as agreement errors, misspellings or adding of extra words. To use NLP systems in real applications, we need to construct an NLP system that can handle not only grammatically well-formed inputs but also grammatically ill-formed inputs. This paper describes an integrated method for processing grammatically ill-formed inputs. We use partial parses of the input sentence for recovering from parsing failure. In order to select partial parses appropriate for error recovery, cost and reward are assigned to them. The notion of cost and reward is introduced in order to select a partial parse appropriate for error recovery. Cost and reward represent the badness and goodness of a partial parse, respectively. The most appropriate partial parse is selected on the basis of cost and reward trade-off. Cost is calculated by cost-based unification proposed here. The system contains three modules. Module A handles local ill-formedness such as constraint violations. Module B handles non-local ill-formedness such as word order violations, and Module C handles non-local ill-formedness such as contextual ellipses. These three modules work in a uniform framework based on the notions of cost and reward.

If a computer is to approach the same capability of understanding that humans possess, it will need many types of knowledge. One such type of knowledge pertains to the concepts of objects. There are many ways to define these concepts. One way is to define them using sets of attributes. The knowledge concerning objects and their attributes is called object-attribute knowledge. When a task domain is not restricted, describing by hand the attributes for all objects that should be described is too much work. This paper presents a method for acquiring and modifying object-attribute knowledge based on several examples indicating whether or not an object has an attribute. In this method, an existing thesaurus showing a classification hierarchy of objects is used to control object-attribute knowledge inheritance. This thesaurus does not need to be built specifically for object-attribute knowledge inheritance and is handled as a set of hypotheses which can be modified to yield an accurate inheritance. When an example is given, thesaurus hypotheses and new hypotheses concerning object-attribute knowledge are modified to make both sets of hypotheses consistent with all examples. This modification is done using three levels of hypotheses which are assigned automatically. If a contradiction occurs with a given example, the hypotheses of a lower level are modified to resolve the contradiction. If this fails to resolve the contradiction, upper-level hypotheses are modified. Whether or not an object has an attribute is inferred by using the set of hypotheses and inference rules even if no example indicating that the object has the attribute has been given. Thus, it is not necessary to describe the attributes of all objects by hand. This method is applied to acquire and modify the knowledge pertaining to the measure-attribute of objects. Measure-attribute is an attribute concerning measurable objects properties. Two different thesauruses are used for knowledge inheritance. Whichever of them is used, this method properly infers the measure-attribute of many objects from a few examples.

This paper studies the problem of learning decision trees when the attributes of the domain are tree-structured. Quinlan suggests a pre-processing approach to this problem. When the size of the hierarchies used is huge, Quinlan's approach is not efficient and effective. We introduce our own approach which handles tree-structured attributes directly without the need for pre-processing. We present experiments on natural and artificial data that suggest that our direct approach leads to better generalization performance than the Quinlan-encoding approach and runs roughly two to four times faster.

When we carry out planning to make arrangements, we may reach some point in dispute in planning because of physical obstacles and/or logical contradictions. When we concerned with mechanizations for adjustments of planning proposals at meeting, we are motivated to formulate the point in dispute and its dissolution with reference to space and time descriptions in proposals. In this sense, we take the situation theory presented by J. Barwise as the basis. Making use of the situation theory, we formally define the point in dispute from logical conflict, timing and spatial problem viewpoints. To do so, we present the idea of situation model which represents a set of situations supporting an infon [Barwise 89]. We take the notion of dual infons, which represent a logical inconsistency, time or space conflict. The point in dispute reflects the existence of dual infons. By means of dual infons, we firstly define the point in dispute of the situation model, and next present a method to dissolve the point by removing the set of situations, expressed in terms of the situation models, which support the point in dispute. Three kinds of points in dispute are introduced, being discussed in a unified way so that they may be dealt with in Lisp and be dissolved.

We propose a multi-agent system for hypothetical reasoning based on a large-scale computational theory on essential characteristics of neocortical processing. For a problem-solving on a real world environment, we require both a large-scale computational theory and a robust local computational theory. As a large-scale computational theory, we develop a hypothetical reasoning system by introducing a knowledge-based control on agents and a local commu-nication among agents. These agents communicate each other to reach a globally consistent solution while they locally perform hypothesis generation, representation and evaluation based on a memory-based reasoning as a robust local computational theory. This memory-based reasoning is defined by a principal component analysis, and applies both a deductive reasoning and an inductive reasoning with a least amount of memory that are requisites for hypothetical reasoning. By its multiple representation of same-type knowledge, and its intrinsic local control for decision-state-dependent recall of that knowledge, the proposed agents also serve as symbolic representations of the signal description of a respective feature. Since vision is a typical case for problem-solving by hypothetical reasoning, the proposed general architecture has been used to implement a model on face recognition to verify its performance.

Real-time systems consist of many concurrent processes and behave on strict timing conditions. It is important to verify the timing conditions of real-time systems. In this paper, we propose extended real-time temporal logic and effective real-time model checking as follows. (1) The timing description of extended real-time temporal logic consists of both freeze quantification and branching time structure in dense time model. For this extension, extended real-time temporal logic admits timing constraints between distant contexts. (2) System specification is described by parallel composition of timed automata. Timed Kripke structure is automatically generated from timed automaton. (3) Real time model checking consists of both labelling algorithm and geometric region. This method is shown effective by some example.

As the opportunity of using computer systems spreads in everyday life, the need for friendly and intelligent interfaces are increasing. It is desirable that one can interact with computer systems as if one interacts with other person. Our interface agent called Visual Software Agent (VSA) has a realistic human face which moves in real-time on a CRT responding to a user, and can interact with users using natural languages. These two communication modalities help users interact naturally with this system; thus it is suitable for systems used by various people with little knowledge about computers, such as guidance systems at department stores, shopping arcades, companies, universities, etc. In this paper, we propose a dialog management system (DMS) with an automatic learning capability of cooperative answering strategy. This DMS consists of six modules, i.e., a state recognizer, an interpreter, an answer generator, a user-model manager, a learning module and a domain database. The system's answer is determined depending on semantic representation of user utterance and the current state of a user-model. Furthermore, the learning module modifies answering rules' strength using a reinforcement learning algorithm so that the system can learn cooperative answering strategy automatically from its experiences. With this learning module, VSA can modify its behavior of interactions with users; consequently, we can develop friendly guidance systems easily. Although the scale of experiments is still small, its results indicate the applicability of the learning module to real systems.

In this paper, we apply our multi-autoepistemic logic (MAEL), a natural extension of Moore's AEL with multi-theories, to formalization of hierarchical knowledge and temporal knowledge. In MAEL, inheritance, persistence and causality are represented both by meta-beliefs about relations between agents' beliefs and by communications of meta-beliefs between agents. Solving some typical examples such as multiple inheritance, Yale shooting problem and its extension, we show that MAEL provides a unified framework for representation of inheritance, persistence and causality. Especially the representation contains in itself a mechanism to control directions of reasoning about persistence and causality.