Building Large-Scale Japanese Pronunciation-Annotated Corpora for Reading Heteronymous Logograms

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 7113–7121 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 7113 Building Large-Scale Japanese Pronunciation-Annotated Corpora for Reading Heteronymous Logograms Fumikazu Sato,1,3 Naoki Yoshinaga,2 Masaru Kitsuregawa4,2 1 The University of Tokyo 2 Institute of Industrial Science, the University of Tokyo 3 National Diet Library 4 National Institute of Informatics {fsato0609, ynaga, kitsure}@tkl.iis.u-tokyo.ac.jp Abstract Although screen readers enable visually impaired people to read written text via speech, the ambiguities in pronunciations of heteronyms cause wrong reading, which has a serious impact on the text understanding. Especially in Japanese, there are many common heteronyms expressed by logograms (Chinese characters or kanji) that have totally different pronunciations (and meanings). In this study, to improve the accuracy of pronunciation prediction, we construct two large-scale Japanese corpora that annotate kanji characters with their pronunciations. Using existing language resources on i) book titles compiled by the National Diet Library and ii) the books in a Japanese digital library called Aozora Bunko and their Braille translations, we develop two large-scale pronunciation-annotated corpora for training pronunciation prediction models. We first extract sentence-level alignments between the Aozora Bunko text and its pronunciation converted from the Braille data. We then perform dictionary-based pattern matching based on morphological dictionaries to find word-level pronunciation alignments. We have ultimately obtained the Book Title corpus with 336M characters (16.4M book titles) and the Aozora Bunko corpus with 52M characters (1.6M sentences). We analyzed pronunciation distributions for 203 common heteronyms, and trained a BERT-based pronunciation prediction model for 93 heteronyms, which achieved an average accuracy of 0.939. Keywords: pronunciation-annotated corpus, pronunciation prediction, pre-trained model 1. Introduction The screen reader is a canonical tool not only for visu- ally impaired people to understand written text, but also for children with reading difficulties to learn from text- books. For example, in Japan, the government has en- forced in 2019 a new law on act to further the improve- ment of reading environments for visually impaired persons and begun to develop multimedia DAISY (dig- ital accessible information system) textbooks for chil- dren with reading difficulties. Although screen readers play a vital role in these kinds of activities, incorrect reading causes confusion in understanding the text, es- pecially in languages with common heteronymous lo- gograms (e.g., Japanese and Chinese). For example, in Japanese, if ‘表’ in a phrase ‘表に出る’ is pronounced as ‘hyou’ instead of ‘omote,’ the meaning changes from ‘go outside’ to ‘listed in a table’; in Chinese, if ‘好’ in a phrase ‘这个人好说话’ is pronounced as hào in- stead of h˘ao, the meaning changes from ‘this person is easy-going’ to ‘this person likes talking.’ Since pronunciations of heteronyms depend on individ- ual contexts, we want to use a machine-learning clas- sifier to predict a pronunciation of a heteronym for a given context. However, since there is no large-scale pronunciation-annotated corpus in Japanese, it is diffi- cult to train an accurate pronunciation classifier for var- ious heteronyms. Although the recent pretrain-finetune …

Content truncated to 500 words