Estimation of semantic case of Japanese dialogue by use of distance derived from statistics of dependency

In an attempt to estimate the semantic cases for nounparticle-verb triples in the ATR dialogue corpus, the authors propose a measure of distance based on statistics of dependent noun-particle-verb triples. A clustering analysis of all the triples in the corpus was conducted using the measure of distance. Competence of the proposed measure of distance is verified by examination of the distribution of the single-case clusters. By use of the score derived from the measure of distance of the training corpus, the authors conducted the estimation of the correct semantic case for a given noun-particle-verb triples in the test corpus. The result remarkably differentiates the particles with respect to the estimation accuracies. For instance, particle 'wo' has accuracies over 80 %, while 'de' has accuracies less than 40 %. The correlation analysis between the accuracy and the consistency rates indicates that the particles of higher consistency have also tendencies to higher accuracies .