In this paper, we propose a standard utterane-unit tagging scheme, which has been developed by the discourse tagging working group under SIG-SLUD, JSAI. Utterance-unit tagging mainly addresses the type of illocutionary force and the role of The interaction unit. We have made a first version of the tagging scheme by surveying existing tagging schemes developed by several research groups. We have evaluated it on an experimental basis and thereby revised it to the new version that we propose as a standard scheme. The reliability of this scheme is demonstrated by another tagging experiment.

The Japanese Map Task Corpus was created between 1994 and 1998 and contains a collection of 23 hour digital recordings, digitized maps and orthographic transcriptions of 128 dialogues by 64 native Japanese speakers. Map task dialogues are dialogues participated in by two speakers, the instruction giver who has a map with a route and the instruction follower who has a map without a route. The giver verbally instructs the follower to draw a route on his map. The two maps are slightly different so that there may emerge a natural interaction in spite of the fact that the flow of information internal to the task is basically one way. The principle and design of the recordings are described with special reference to the augmentations and improvements to the original HCRC Map Task corpus. Annotations to the orthographic transcriptions are viewed as "tags" that provide the start and end times of utterances, the duration of pauses, non-verbal events and synchronization of overlapping utterances, in a format which provides a view to giving a basis for further tagging in terms of linguistic and discourse phenomena in a interchangeable and sharable manner. Discourse and linguistic phenomena peculiar to spontaneous spoken dialogues, such as overlapping, are analyzed and the method of recording such phenomena in the transcription is discussed and proposed, with an implication for the requirement that one dialogue be represented in one digitized sound file for the preservation and reference of the information on timing. The tags emp1oyed in the corpus also provide an easy way of characterizing it in terms of the number and the duration of utteraI1ces and pauses. The statistical figures thus ob-tained are relatively independent of design factors like kinds of maps, but familiarity does significantly correlate with the duration and number of utterances.

Recently, large scale dialogue corpus with annotation of semantics and/or discourse is required in higher level natural language processing. The production of such a corpus demands a great amount of human effort because such a corpus involves higher level knowledge. A program that infers utterance-unit tags using other level tags would help reduce tagging work. Thus, we have proposed a method to infer the utterance-unit tag from both the text corpus and its morpheme analysis result. Also we have developed a GUI tool to support discourse level tagging, and have implemented the inferring method as one of the functions of this tool. The experimental result to evaluate this method shows a correct answer of 86% in c1osed tests and 73% in open tests.

This report describes tagging of discourse seglnents for spoken dialogues. We carried out tagging experiments for dialogues of two tasks, meeting space reservation and route direction, using two versions of written instructions. The agreement of tags between the coders is quantitatively evaluated with reliability in terms of the kappa coefficient (K). The first version of written instructions uses a tagging scheme of discourse segments based on Grosz and Sidner's discourse structure model. The discourse segment is marked with the beginning and the ending points, and the structllre of discourse segments can be nesting. The coders reached K=0.63 on the existence of discourse segment tags, and K=0.45 on the tags with the nesting level. Major disagreements of the discourse segment tag were ending points of the segment and nesting structure. In the second version, the tagging scheme was renewed based on characteristics of dialogue, which are exchange structure and ambiguous ending of topic. The written instruction requires coders to mark boundaries of the discourse segments with topic break index (TBI). The coders reached K=0.65 ignoring the TBI's, K=0.56 on the tags with TBI, and K=0.63 on the boundaries with TBI=2.

We investigate a relation between agreement/disagreement expressions and discourse tags in Japanese Discourse Tagging Manual (JDTAG) and DAMSL. An agreement/disagreement expression is defined as the latter utterance of consecutive utterances, which shows agreement/disagreement with the former. We analyzed these expressions in dialogues recorded in a cooperative task experiment. Subjects used two types of agreement expressions: Agreement words and Agreement Kyowa. Subjects used two types of disagreement expressions: to the partner's utterance, and to their own. We found that the tags in JDTAG and DAMSL do not treat the intention of agreement or disagreement on Agreement Kyowa and Disagreement expressions to one's own utterances. We discuss the discourse tags for these expressions.

An improved way to select a feature subset for decision tree induction is presented. The method is based on a heuristic that is defined by means of explanation paths of the decision tree. Although the method emp1oys a control structure that is similar to the wrapper model, it is much more eflicient because the heuristic that controls the search requires to call an induction algorithm much less times than the wrapper model. Experiments prove that the algorithm is as effective as the wrapper model, not only to simplify decision trees but also to improve generalization ability. L-usefulness, a newly defined property of a feature and theoretical background of the method, is also discussed.

Two approaches to inducing a concept represented in first order logic are inductive logic programming (ILP) and genetic programming (GP). In ILP, concept learning can be considered as a search in the space specified by the background knowledge, and in which the goal concept is represented by Horn clauses. On the other hand, in GP, the search space is specified by terminal and nonterminal symbols, and the goal is represented generally by S-expressions. These two approaches are very similar in terms of their methods and goals, yet their combination in previous work is rare. In this paper, we propose a method that synthesizes the inductive logic programming and genetic programming approaches. The concept behind this approach is to combine the search method of GP, that is, Genetic Algorithm, with the type and mode methods of ILP. We have implemented a system called SYNGIP (SYNthesized system with Genetic programming and Inductive logic Programming) based on the method. Experimental results show that the proposed method can be used to treat, in the same way, learning from training examples that do not have discrete classes, and learning from both positive and negative training examples. Moreover, the proposed method constitutes a novel solution to the c1osure problem and provides a new bias for concept learning.

Externalizing the contents which one learns about instructional material is a fruitful way to facilitate or reinforce his/her learning. However, such knowledge externalization does not always occur spontaneously. Alternatively, it is not so easy for learners because they may make excessive mental efforts which will cause cognitive overload. In this paper, we propose an intelligent assistance in externalizing knowledge structure to be learned with instructional material, which in advance provides learners with part of the externalized structure. This intends to show them what to and how to externalize, and to reduce the externalization load. Considering the context of learning a text describing computer vocabulary, this paper also illustrates the intelligent assistance which is called Diagram Tailoring, and describes an experiment on it. As a result, we have ascertained that Diagram Tailoring contributes to enhancing an understanding of a text. We also demonstrate an intelligent learning system which accomplishes Diagram Tailoring. In this system, learners are encouraged to compose a diagram to visualize the processes of externalizing their knowledge. In order to control their diagramming load, in addition, the system properly presents part of the diagram. The load control is done with the learner models representing their diagramming processes and diagramming capability. Such tailoring technique enables the learners to facilitate or reinforce the retention of their externalized knowledge more effectively.

This paper proposes a computation method of relative least general generalization using liteal association and MDL criteria in inductive inferene with a logical framework. The method does not require additional conditions, which are difncult to be described in KDD applications, to select relevant ground atoms from background knowledge. The notion literal association is a relationship among literals through common variables in the literals. The MDL criteria selects literals in the generalized clause by the description length of literals as a post-processing. The search using these properties is processed with constraints that the generalization covers no negative examples in top-down manner. The validity of the proposed method is shown by an experiment to learn list processing predicates from examples in logic programming traditions.

Many problems in AI can be formulated as Constraint Satisfaction Problems (CSPs). In actual problems, Constraint Optimization Problem (COP) is often more important. In this paber, we present a new efficient method for COP, more exactly, for VCOP (value-based COP), in which every value of CSP is assigned an integer cost and the CSP solution with a minimal cost is to be found. Since the computational cost of COP is very high, i.e., NP-hard, simple search methods working in binary (0-1) space are not good enough in terms of efficiency. Thus we try to utilize a linear programming method, namely simplex method, to find a near-optimal solution efliciently. In our method, COP is transformed at first into a 0-1 integer linear programming problem. Then we solve it's relaxation problem which allows a real-number solution rather than a O-1 solution. With using the real-number optimal solution, we execute a local search method called GLS (guided local search) to find a 0-l near-optimal solution of COP. GLS is a general and compact optimization technique with a capability of escaping from local optimal points. In our method, the initial search point for GLS is determined referring to the real-number optimal solution of the simplex method. Also, terms regarding the distance between a temporal search point and the real-number optimal solution are added into the augmented cost function of GLS to control the search to intensively investigate the neighborhood of the real-number optimal solution. This method shows a good performance empirically in computational efEiciency and in solution quality on COP, especially on loosely constrained COP. In a wide sense, our research here has shown that the combination of the linear programming technique with search method is beneficial to achieve search efficiency in constraint optimization.

We discuss the effect of inconsistent heuristics on the completeness property of the Moving-Target Search (MTS), a real-time search algorithm for changing goals. Although MTS can be considered as a natural extension of the stationary-goal real-time search algorithm LRTA, it is augmented with an extra operation to force the consistency of heuristic estimates. We argue that this extra operation, which ensures the updating of the heuristic estimates is carried out only when it increases the current value, can be safely eliminated from MTS as well. We also show that the completeness is preserved even if the heuristic function is inadmissible (and hence inconsistent). These results elucidate the nature of the MTS algorithm, and broaden the class of heuristic functions that can be used with MTS without violating the completeness.

The real world includes various unknown elements, objects, phenomena, and so on. Causes of phenomena to be observed is difficult to specify, because they might be uncoverable for the observer. That is, it takes too much time to consider all the possible causes, or sometimes they are even unable to be considered because they are unknown. Now then, we aim at constructing a learning method by which to learn unknown causal-links and unknown basic causes, i.e., deeper causes than those assumable as hypotheses. A major contribution of this paper is that the input data are not combinations of simultaneous occurence of events, but the average occurence of each event given as probabilities. By determining the position of new nodes in the probabilitic network, the user reasons for him/herself, what the new node means in the real world. The system is applied to analyzing data obtained by a few questionaires.

Production systems are established methods for encoding knowledge in expert systems. It has been pointed out the lack of supporting too1s for developing knowledge-base to use production systems more widely. Especially, there has been a demand for a verification system for production systems. In this report, we present our verification system for OPS5-based production systems. It consists of three steps to verify the knowledge-base. (1) Detect errors automatically. (2) Suggest a correction of errors to user. (3) Correct the errors if user permitted. Our system introduces new concepts called "variable expand" and "expected WME" for detecting more errors than other verification systems. By using "variable expand", we can restrict variable range. And using "expected WME", we can detect some errors that required dynamic information, only using static information from knowledge-base.