Catálogo de publicaciones - libros

This paper discusses the fake past tense morphology used for present state in Japanese (Teramura 1984; Iatridou 2000). Unlike Korean and other languages, the past tense marker “ta” can express an unexpected finding or remembrance at the time of speech (Inoue and Ubukoshi 1997). I claim that this construction corresponds to subjunctive conditionals with a covert negative antecedent, and that counterfactuality is involved in such non-past past tense, even though the proposition expressed is factual. It is comparable to the counterfactual analysis of factive emotive predicates such as “sorry”and “glad”(Heim 1992; Giorgi and Pianesi 1997). I show that non-past past construction updates the information state, and revises the context.

In this paper, I propose a systematic account of the semantics of English present perfect (PresPerf) from both empirical and formal perspectives. I incorporate the insights of an extended now (XN) analysis ([1]) into the dynamic semantics framework, DRT ([2]). I demonstrate that PresPerf is licensed by its relation to an XN interval and that different readings are attributed to the specific way that the embedded eventuality interacts with an XN. I provide the semantics of PresPerf compositionally in terms of a general component attributed to + and specific conditions attributed to adverbials and verb semantics. In addition, I characterize PresPerf as a tense by showing that the XN interval serves as ”temporal location” ([3]), a determinant of tense. An alleged connection between PresPerf and a perfect state is only entailed in the case of events.

The workshop on Active Mining (AM-2004) was held on June 1, 2004 at Ishikawa Kousei Nenkin Kaikan in Kanazawa City, Japan, as a part of the Eighteenth Annual Conference of the Japanese Society for Artificial Intelligence (JSAI-2004). This is the third workshop that focuses on Active Mining; the first one was held on December 9, 2002, as a part of the Second IEEE International Conference on Data Mining (ICDM’02), and the second one was held on October 28, 2003 as a part of the 14th International Symposium on Methodologies for Intelligent Systems, both at Maebashi TERRASA, Maebashi City, Japan.

In this paper, we summarize our endeavor for spiral discovery of a separate prediction model from chronic hepatitis data. We have initially proposed various learning/discovery methods including time-series decision tree, PrototypeLines, and peculiarity-oriented mining method for mining the data. This experience has motivated us to model physicians as considering typical cases with the specific disease and ruling out clearly exceptional cases. We have developed a spiral discovery system which learns a prediction model for each type of cases, and obtained promising results from experiments.

Chance discovery is the process of human interaction with the environment for discovering events significant for making a decision. We executed the process of chance discovery, on the blood-test data for hepatitis B, for obtaining scenarios telling when and how symptoms essential for treatment appear. In the double-helical process of chance discovery, the presented scenario maps are evaluated and fed back to the following cycles, to obtain novel and potentially useful knowledge for treatment. Due to the combination of the objective facts in the data and the subjective focus of the hepatologists’ concerns in this process, the relation between the changes of iron quantities due to iron-carrying proteins and the cure of hepatitis B with interferon, has got clarified visually.

Data mining is the efficient discovery of patterns in large databases, and classification rules are perhaps the most important type of patterns in data mining applications. However, the number of such classification rules is generally very big that selection of interesting ones among all discovered rules becomes an important task. In this paper, factors related to the interestingness of a rule are investigated and some new factors are proposed. Following this, an interactive rule interestingness-learning algorithm (IRIL) is developed to automatically label the classification rules either as “interesting” or “uninteresting” with limited user participation. In our study, VFP (Voting Feature Projections), a feature projection based incremental classification learning algorithm, is also developed in the framework of IRIL. The concept description learned by the VFP algorithm constitutes a novel hybrid approach for interestingness analysis of classification rules.

Data mining systems semi-automatically discover knowledge by examining large volumes of data, but the knowledge so discovered is not always novel to users. We introduce a discovered-rule-filtering approach that uses information retrieval results from the Internet to assess rules discovered by data mining and find those that are novel to the user. To implement this approach, we create 2 methods: the micro view method and the macro view method. In the micro view method, we extract keywords from a discovered rule and rank the rule referring to the number of hits returned when the keywords are submitted to an appropriate database. In the macro view method, we first retrieve documents by submitting every pair of extracted keywords and then form keyword clusters according to the results. We evaluated the methods by sending out a questionnaire to medical students and using the MEDLINE database as our Internet source. The evaluation indicates that the macro view method is promising.

The relevance feedback based on a keyword map is proposed so that a Web interface can be more interactive. There exists vast amount of information in the Web, from which users usually gather information without definite information needs. Therefore, it is difficult for users to organize and understand what they have gathered from the Web. From this viewpoint, we have proposed the concept of RBA-based interaction, in which analysis operation aims to assist users in understanding the context of their web interaction. However, the currently developed interface focuses on the information flow from the interface to users. As the first step for realizing relevance feedback (RF) based on interactive keyword map, this paper proposes the algorithm for extracting the pair of keywords that reflects a user’s interest from the keyword map. Experimental results are given for showing how the algorithm works on the keyword map that is modified by the user, and for discussing the difference between the RF based on keyword map and conventional RF methods.

The huge number of descriptive features is often a problem in data mining. We analyzed structure activity data for dopamine antagonists, which involves selecting useful features from numerous fragments extracted from their chemical structures. Correlation coefficients among categorical variables were used to select attributes. Chemists evaluated the rules obtained by the cascade model, and the importance of attribute selection was confirmed.

In this paper, we propose an approach which can improve Inductive Logic Programming in multiclass problems. This approach is based on the idea that if a whole rule cannot be applied to an example, some partial matches of the rule can be useful. The most suitable class should be the class whose important partial matches cover the example more than those from other classes. Hence, the partial matches of the rule, called , are first extracted from the original rules. Then, we utilize the idea of Winnow algorithm to weigh each partial rule. Finally, the partial rules and the weights are combined and used to classify new examples. The weights of partial rules show another aspect of the knowledge which can be discovered from the data set. In the experiments, we apply our approach to a multiclass real-world problem, classification of dopamine antagonist molecules. The experimental results show that the proposed method gives the improvement over the original rules and yields 88.58% accuracy by running 10-fold cross validation.