Share on Facebook Tweet on Twitter Share on LinkedIn Share by email
Predicting Word Pronunciation in Japanese

Jun Hatori and Hisami Suzuki

Abstract

This paper addresses the problem of predicting the pronunciation of Japanese words, especially those that are newly created and therefore not in the dictionary. This is an important task for many applications including text-to-speech and text input method, and is also challenging, because Japanese kanji (ideographic) characters typically have multiple possible pronunciations. We approach this problem by considering it as a simplified machine translation/transliteration task, and propose a solution that takes advantage of the recent technologies developed for machine translation and transliteration research. More specifically, we divide the problem into two subtasks: (1) Discovering the pronunciation of new words or those words that are difficult to pronounce by mining unannotated text, much like the creation of a bilingual dictionary using the web; (2) Building a decoder for the task of pronunciation prediction, for which we apply the state-of-the-art discriminative substring-based approach. Our experimental results show that our classifier for validating the word-pronunciation pairs harvested from unannotated text achieves over 98% precision and recall. On the pronunciation prediction task of unseen words, our decoder achieves over 70% accuracy, which significantly improves over the previously proposed models.

Details

Publication typeProceedings
Published inCICLing 2011, Lecture Notes in Computer Science (6609)
PublisherSpringer Verlag
> Publications > Predicting Word Pronunciation in Japanese