Dataset Construction Method for Word Reading Disambiguation

Dataset Construction Method for Word Reading Disambiguation Koki Nishiyama Kazuhide Yamamoto Department of Electrical Engineering Nagaoka University of Technology 1603-1 Kami-Tomioka, Nagaoka, Niigata, 940-2188, JAPAN nishiyama@jnlp.org Hideharu Nakajima NTT Communication Science Labs. NTT Corporation 2-4 Hikari-dai Seika-cho, Soraku-gun, Kyoto, 619-0237, JAPAN nakajima.hideharu@lab.ntt.co.jp Abstract The scarcity of large corpora in reading dis- ambiguated words is a major limitation in lin- guistic analysis and the initiation of a statisti- cal approach to word reading disambiguation. As readings of words are usually not written in documents like meanings of words, there- fore, human annotation is necessary but ex- pensive. In this study, a method is proposed to construct a reading disambiguated dataset for word reading disambiguation. The method constructs a dataset of sentences wherein words with ambiguity in reading (pronuncia- tion), called heteronyms, are tagged for cor- rect reading. In this method, a word with unique reading is labeled to a heteronym, and this unique word is used as a query word to collect sentences that include the word. The word in the collected sentences is replaced by the original ambiguous word and the read- ing corresponding to that of the query word is tagged as the pronunciation of the het- eronym. It was confirmed through experi- ments that the method was able to collect data effectively, and the collected data was numer- ically balanced among all the readings of the heteronym. 1 Introduction Text-to-speech (TTS) synthesis requires prosodic phrase boundaries and pronunciation (reading and stress/accents) (Hall and Sproat, 2013; Sproat and Hall, 2014). These are determined on the basis of linguistic analysis with natural language processing technologies. Although an adjacent word informa- tion such as a statistical language model is applied for the processing, a few words in narrow contexts are used that sometimes lead to errors. Among the errors, those in pronunciation are noticeable in the TTS situations. Furthermore, it has been reported that the subjective evaluation of TTS by listeners is particularly sensitive to homograph disambigua- tion errors (Braga et al., 2007; Gorman et al., 2018). Thus, homograph disambiguation has been one of the important tasks in synthesis and has been tackled by many researchers in the past (Yarowsky, 1997; Braga et al., 2007). Homographs are words that share the same writ- ten form as other words but have a different mean- ing, and can be classified into either homonyms or heteronyms, depending on their pronunciation, e.g., homonyms have the same pronunciation, such as lie (untruth) [l´Ai] and lie (to recline) [l´Ai], whereas het- eronyms have different pronunciation, such as desert (region) [d´ez@rt] and desert (to leave) [diz’@:rt]. Thus, disambiguation of heteronyms is a more im- portant task in TTS synthesis, as the correct pro- nunciation of the word in its context must be deter- mined. Statistical methods have been considered to be better techniques for disambiguation, such as word sense/homograph disambiguation (Yarowsky, 1992; Yarowsky, 1997; Mihalcea and Moldovan, 1999; Gorman et al., 2018). These methods require large corpora that are manually tagged with correct mean- ings or pronunciations. The …

Content truncated to 500 words