A major challenge in natural language understanding is to build a comprehensive common-sense knowledge base in the efficient way possible This thesis discusses automatic knowledge acqursition from text, especrally the acquisition of causal relations We consider four types of causal relations, namely, cause, effect, precondition and means They are based on agents' volitionality, as proposed in discourse understanding The idea behind knowledge acquisition is to use resultative connective markers such as "because", "but" and "if" as linguistic cues However, there is no guarantee that such a connective marker always signals the same type of causal relation Therefore, we need to create a computational model that is able to classify samples according to the causal relation. In this work, we focus on Japanese complex sentences including the word ため(because) The following questions are asked (1) What kinds and how much causal knowledge is present in the document collection, (2) How accurately can relation instances be identified, and (3) How can acquired causal knowledge be made available to applications First, we investigated the distribution of causal relation instances in Japanese newspaper articles The main part of this investigation was conducted based on human judgments using lingnstic tests. We confirmed that it is possible to acquire causal relation instances from approximately 90 % of samples Second, we assessed how accurately we can automatically acquire causal relation instances by experiments Using a machine learning technique, we achieved 80 % recall with over 95 % precision for the cause, precondition and means relations, and 30 % recall with 90 % precision for the effect relation Furthermore, the classification results suggest that one can expect to acquire over 27,000 instances of causal relations from one year of Japanese newspaper articles Third, we applied the acquired causal knowledge to annotate words with its desirability From this investigation, it became clear that causal relation instances, at least instances of cause relations and means relations, are useful for assigning desirability of words

The author has proposed a method for knowledge creation called "Knowledge Liquidization and Crystallization", and has developed a system called "Knowledge Nebula Crystallizer" to support the process Though several theories about knowledge creation have been proposed and companies have realized their importance, most of the companies face to difficulty of applying them to their real practices The aim of this study is to establish a method for supporting concept articulation in knowledge management processes in a real practice and a system to support it Knowledge has been regarded as something formalizable and universally sharable However, Knowledge is not something "solid" but "liquid". Rather, it is something like "liquid" It takes any shape to fit to a certain context, and a portion of it is extracted and fused into new knowledge that fits to a new context Exhibition design was selected as an application of our method In this paper the author describes the application of the proposed method and system The system is named "Knowledge Nebula Crystallizer for Exhibition Design (KNC4ED) " Evaluation on the method and the system based on user studies and discussions with professional exhibition designers is also described The contributions of this study are - Though researches on knowledge creation processes have been conducted for their theoretical aspects, the author has proposed and developed a method and a system to enable the processes based on the research results including knowledge management, creativity support, Human-Computer Interaction, and wearable computers - The author has applied the method and the system to the realworld design problem, i e, exhibition design Through the user studies, the author has observed that the method and system have supported the process of knowledge creation - The validity of the method and the system for practical applications have been proved through the discussions with the senior manager and the senior designer at the advertising company The result of this study is applicable not only to the exilibition design field but also to the other business practices

The technology capable of capturing and analyzing the changes on the Web is no doubt vital, in providing the needed information m time for one to stay competent in this fast changing information age. This dissertation presents an approach toward the automatic journalism of new information (changes) on the Web. These information changes on the Web can be classified into two types: "flow" and "stock". "Flow" type information (i.e news) come to the Web constantly and regularly, at a rather fast pace "Stock" type information, mainly the static web pages, change unpredictably doesn't know at when and in what form. Our system aims to innovate the technology and use a new TF * PDF (Term Frequency * Proportional Document Frequency) algorithm to detect the prominent topics in the changes. In the framework and domain of problem addressed, this algorithm is more superior than the conventional TF * IDF algorithm in a way that it doesn't need retrospective corpus, besides posing minimal risk of losing the tracks of detection and tracking of popular topics. Also, our system requires less computational complexity while offering more flexibility. It crawl the Web, collects the changes and journalizes a summary of popular topics to the user. It does more than the conventional web tracking systems that just acknowledges the URLs of changed pages. It can become our personalized e-journalist on the Web and periodically provide us with the collection and e-publication of currently popular events.

By rapid progress of network and storage technologies, a huge amount of weakly structured data such as Web pages and XML data, called semistructured data, have been available on Internet and intranets Therefore, there have been increasing demands for efficient methods that extract rules and patterns from semi-structured data, namely semi-structured data mining However, semi-structured data are a huge amount of complex and hetero-geneous data modeled by trees or graphs Thus, we can not directly apply to semi-structured data traditional data mining methods for relational databases Hence, it is an important challenge to develop efficient and scalable methods for semi-structured data mining In this thesis, we model semi-structured data as labeled trees, and study efficient semi-structured data mining algorithms for various classes of tree patterns In Chap 3, we consider the problem of discovering frequent ordered tree patterns from semi-structured data and developed an efficient algorithm FREQT for the problem The key technique is an efficlent enumeration technique called rightmost expansion, which enumerates all the ordered trees in 0 (1) time per pattern Consequently, FREQT computes all the frequent patterns m 0 (kb^2m) time per pattern without duplication, where k is the size of the pattern, b is the maximum branching of an input data tree, and m is the number of occurrences of the pattern In Chap 4, we then extend the algorithm FREQT to the optimized pattern discovery problem and give an efficient algorithin OPTT for mining optimized ordered tree patterns In Chap 5, we consider the frequent unordered tree pattern discovery problem for semi-structured data and developed an efficient algorithm UNOT for the problem In Chap 6, we study a variant of the frequent tree discovery problem, called frequent pattern mining from semi-structured data streams, where the input to the mining algorithm is not a static dataset, but a rapidly changing and unbounded data stream We developed an online algorithm StreamT, which discovers all the frequent ordered trees appearing in a given data stream by scanning it once Finally, we conclude this thesis and discuss future research directions in Chap 7