Locating Boundaries for Prosodic Constituents in Unrestricted Mandarin Texts

61 Computational Linguistics and Chinese Language Processing Vol. 6, No. 1, February 2001, pp. 61-82  Computational Linguistics Society of R.O.C. Locating Boundaries for Prosodic Constituents in Unrestricted Mandarin Texts Min Chu*, Yao Qian+ * Microsoft Research Asia, 5F, Beijing Sigma Center No. 49, Zhichun Road, Haidian District, Beijing 100080, P.R.C., E-mail: minchu@microsoft.com + Shanghai Normal University, 100 Guilin Road,Shanghai 200234, P.R.C., E-mail: yqian@shtu.edu.cn Abstract This paper proposes a three-tier prosodic hierarchy, including prosodic word, intermediate phrase and intonational phrase tiers, for Mandarin that emphasizes the use of the prosodic word instead of the lexical word as the basic prosodic unit. Both the surface difference and perceptual difference show that this is helpful for achieving high naturalness in text-to-speech conversion. Three approaches, the basic CART approach, the bottom-up hierarchical approach and the modified hierarchical approach, are presented for locating the boundaries of three prosodic constituents in unrestricted Mandarin texts. Two sets of features are used in the basic CART method: one contains syntactic phrasal information and the other does not. The one with syntactic phrasal information results in about a 1% increase in accuracy and an 11% decrease in error-cost. The performance of the modified hierarchical method produces the highest accuracy, 83%, and lowest error cost when no syntactic phrasal information is provided. It shows advantages in detecting the boundaries of intonational phrases at locations without breaking punctuation. 71.1% precision and 52.4% recall are achieved. Experiments on acceptability reveal that only 26% of the mis-assigned break indices are real infelicitous errors, and that the perceptual difference between the automatically assigned break indices and the manually annotated break indices are small. 1. Introduction The state-of-the-art text-to-speech (TTS) systems are able to produce very natural synthesized 62 Min Chu, Yao Qian speech if they are provided with a correct phonetic string and with prosodic features extracted from human pronunciation of the string [Chu and Lu, 1996; Dutoit et al., 1996]. Automatic prosody generators, however, cannot yet deliver high quality prosody. One of the main obstacles to automatic generation of prosody is the difficulty of identifying the hierarchical prosodic constituents from texts automatically. It has been proven through many experiments [Liberman and Prince, 1977; Gee and Grosjean, 1983; Selkirk, 1984; Ladd and Campbell, 1991] that prosody constituents are not always identical to those of the surface syntax. The relationship between prosody and syntax is not well understood. While, representing prosodic constituents by means of syntactic constituents directly cannot produce very natural prosody, the boundaries of prosodic constituents can be derived from syntactic information. Some early studies used rules to parse prosodic structures. Stochastic models have been used more frequently in recent studies. In some works [Wang and Hirschberg, 1991; Hirschberg and Prieto, 1996; Lee and Oh, 1999], break indices have been predicted using the automatic classification and regression tree(CART) from information such as four-word part-of-speech(POS) windows, pitch accent types, the sentence length, the distance from the beginning of the sentence and the end of the sentence, etc. A Markov model is used in …

Content truncated to 500 words