Question classification, an important phase in question answering systems, is the taskof identifying the type of a given question among a set of predefined types.This studyuses unlabeled questions in combination with labeled questions for semi-supervisedlearning, to improve the precision of question classification task.For semi-supervisedalgorithm, we selected Tri-training because it is a simple but efficient co-training stylealgorithm.However, Tri-training is not well suitable for question data, so we give twoproposals to modify Tri-training, to make it more suitable.In order to enable itsthree classifiers to have different initial hypotheses, Tri-training bootstrap-samplesthe originally labeled set to get different sets for training the three classifiers.Theprecisions of three classifiers are decreased because of the bootstrap-sampling.Withthe purpose to avoid this drawback by allowing each classifier to be initially trainedon the originally labeled set while still ensuring the diversity of three classifiers, ourfirst proposal is to use multiple algorithms for classifiers in Tri-training;the secondproposal is to use multiple algorithms for classifiers in combination with multipleviews, and our experiments show promising results.

View full abstract

In this paper we present an approach to machine translation from Japanese to JSL (Japanese Sign Language).There is no standard way of writing JSL, and thatbrings complexity and difficulty on natural language processing for JSL.We defineda Japanese gloss-based notation system for JSL.This notation system enables usto divide the MT process into two stages, namely, the text-to-text translation stage (Japanese to JSL text) and text-to-motion synthesis stage (JSL text to JSL motion).Our current focus is on the former stage.This notation also allows us to apply theexisting MT techniques to JSL.We implemented a pilot MT system using jaw, apattern transfer-based MT engine that we have been developing to construct MT sys-tems from Japanese to various (oral) languages.To test the validity of our approachand to find the problems of it, we conducted a translation experiment with sentencesextracted from videos for JSL learners.

View full abstract

Some systems of moods of Japanese sentences are presented.However, it is notdefinitely shown what kind of method those known mood systems are constitutedwith.In addition, it does not seem that they are constituted through analyzing sentences included in various Japanese web pages.Therefore, in linguistic informationprocessing for Japanese web pages, it is very likely that those known mood systemsare insufficient at a point of exhaustiveness of moods.In this article, we describes a ystematic method, in detail, with which we analyze sentences included in 11, 034, 409Japanese web pages collected by NTCIR project and we collect new moods withknown moods.A basic procedure of the method is as follows: (1) Divide a Japanese sentence into words with ChaSen; (2) Focus on a sentence-end-word that probably hasvarious kinds of moods; (3) Assign a proper mood to the sentence-end-word manually. We show the collected new moods, and compare the collected moods and the other known moods in order to clarify what moods can not be collected and what moodshave been already presented among the collected new moods. Based on findings from the comparison, we propose an expanded mood system such that more exhaustivenessof moods is provided.

View full abstract

For natural computer conversation, if a computer has typical responses, and thechangeable parts of sentences can be changed by association, more flexible and morevarious conversations can be done.However, there is a risk that the generation ofresponse sentences by a computer results in a combinations of feeling of wrongnesscaused by the mechanical combination of words.This paper focused on a relationof nouns and adjective phrases.Then the knowledge structure of how to use nounsand adjective phrases is modeled by arranging the relation in a point of feeling ofwrongness.Also, this paper proposes a technique for detection relation of nounsand adjective phrases by creating a knowledge model from generation of responsesentences.Using the method described in this report, we showed that this techniquewas able to very accurately judge usages of nouns and adjective phrases with 87%accuracy, thus demonstrating the effectiveness of the technique.

View full abstract