Japanese Particle Error Correction employing Classification Model

要旨

We present a grammatical error correction system for Japanese particles based on the classification method. We define a confusion set of the particles for detection of particle errors and prediction of the correct word. Our method can handle not only substitutions but also insertions and deletions. For building the training data, we used two datasets: a large amount of native language data and corrected learners' sentences. That is, we did not require a parallel corpus of learners. We show the results for Japanese particle error correction on the NAIST Goyo corpus, evaluated by the MaxMatch (M2) score. In addition, we analyze the effect of percentage changes in deletion labels while building the training data and analyze the prediction probability threshold at correction. Our best model achieved 46.4 F0.5.

主題