This paper describes a method of extracting a large set of hyponymy relations with a high precision from hierarchical layouts in Wikipedia articles. Hyponymy relation has been studied as one of the principal knowledge for information retrieval and web directory, which helps users to access the growing web. Various methods have been proposed to automatically acquire hyponymy relations. In this article, we first extract hyponymy relation candidates from sections and itemizations in hierarchical layouts of Wikipedia articles, and then filter out irrelevant candidates by using a machine learning technique. In experiments, we successfully extracted more than 1.35 million relations from the hierarchical layouts in the Japanese version of Wikipedia, with a precision of 90%.

View full abstract

Recently the development of computers and networks makes amount of information huge. It is very difficult to find necessary information in the huge information. The existing retrieval system uses not the meaning of input words but the notation of them. Therefore, different words bring a defferent result of retreieval even if they have the same meaning. A user of the system has to consider the input words to search the necessary information. This paper proposes the quantification technique of the semantic distance between documents based on relevance of the word to realize the search that captured the meaning of the document. Concretely the related degree between words is calculated by concept-base and the resemblance degree between documents is calculated by Earth Mover’s Distance. Besides this paper proposes method that no existence word on concept-base is defined as a concept based on Web information to expand concept-base automatically. Retrieval experiments using the NTCIR3-WEB in comparison with the other method have shown that our method is effective than other method.

View full abstract

Japanese numeral classifier (NC) expresses quantificational expressions by following number nouns. The parser based on phrase structure grammar like Lexical-Functional Grammar (LFG) or Head-driven Phrase Structure Grammar (HPSG) has to define the relationships between noun and NC or number and NC by the grammatical rules. NC has various relationships syntactically because of its’ various characteristics. The treatment of NCs in LFG formalism should take account of their lexical and syntactical characteristics. This report proposed LFG rules for NCs and examined their validity and accuracy. The comparison between Japanese f-structure inducted by the rules proposed in this report and English f-structure and Chinese f-structure which correspond to the Japanese f-structures show that the rules make f-structures parallel between same phrases in Japanese and in other languages. The experiment of evaluations shows the rules improve F-score of NCs for currency and unit from 52.10% to 77.90% and F-score of NCs for quantity from 81.92% to 87.32% and F-score of NCs for Order from 59.62% to 81.26%.

View full abstract

Adjacent symbol connection constraints (ASCCs) are very useful for not only morphological analysis of non-segmenting language such as Japanese language, but also for continuous speech recognition of any language. By incorporating ASCCs into an LR parsing table, it is possible to reduce the size of the table, as well as reject any locally implausible parsing results. Although several algorithms have been proposed, they cannot remove all of the unnecessary actions because they consider only local context. This paper proposes a new algorithm and show some evaluation results. The proposed algorithm incorporates ASCCs by searching for global action chains from the initial state to the final state. According to the results, the proposed algorithm can remove about 1.2% more actions than a conventional algorithm, and the parsing time can be reduced by about 2.4%. Lastly, we show the completeness of our algorithm.

View full abstract