<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1819">
  <Title>SS</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Prosodic phrase prediction or prosodic phrasing plays an important role in improving the naturalness and intelligence of TTS systems. Linguistic research shows that the utterance produced by human is structured in a hierarchy of prosodic units, including phonological phrase, intonation phrase and utterance. (Abney, 1995) But the output of text analysis of TTS systems is often a structure of syntactic units, such as words or phrases, which are not equivalent to the prosodic ones.</Paragraph>
    <Paragraph position="1"> Therefore the object of prosodic phrasing is to map the syntactic structure into its prosodic counterpart.</Paragraph>
    <Paragraph position="2"> A lot of methods have been introduced to predict prosodic phrase in English text such as Classification and Regression Tree (Wang and Hirschberg, 1992), Hidden Markov Model (Paul and Alan, 1998). For Chinese prosodic phrasing, the traditional method is based on handcrafted rules. Recurrent Neural Network (Ying and Shi, 2001) as well as POS bigram and CART based methods (Yao and Min, 2001) is also experimented recently. Due to the difference in training corpus and evaluation methods between researchers, these results are generally less comparable.</Paragraph>
    <Paragraph position="3"> In this paper, a rule-learning approach is proposed to predict prosodic phrase in unrestricted Chinese text. Rule-based systems are simple and easy to understand. But handcrafted rules are usually difficult to construct, maintain and evaluate. Thus two typical rule-learning algorithms (C4.5 induction and transformation-based learning) are employed to automatically induce prediction rules from examples instead of human. Generally speaking, automatic rule-learning has two obvious advantages over the previous methods: 1) Statistical methods like bigram or HMM usually need large training corpus to avoid sparse data problem while rule-learning doesn't have the restriction.</Paragraph>
    <Paragraph position="4"> In the case of prosodic phrase prediction, the corpus with prosodic labelling is often relatively small. Rule-learning is just suitable for this task.</Paragraph>
    <Paragraph position="5"> 2) CART, RNN or other neural network methods have good learning ability but the learned knowledge is represented as trees or network weights, which are not so much understandable as rules.</Paragraph>
    <Paragraph position="6"> Once rules are learned from examples, they can be analyzed by human to check if they agree with the common linguistic knowledge.</Paragraph>
    <Paragraph position="7"> We can add prediction rules converted from our linguistic knowledge to the rule set, which is especially useful when the training corpus doesn't cover wide enough phenomena of prosodic phrasing. Furthermore, we can try to interpret and understand rules learned by machine so as to enrich our linguistic knowledge. Hence rule-learning also helps us mine knowledge from examples.</Paragraph>
    <Paragraph position="8"> Since features related to prosodic phrasing come from various linguistic sources, several comparative experiments are conducted to select the most effective features from the candidates. The paper also suggests general evaluation parameters for prosodic phrase prediction. With these parameters, our methods are compared with RNN and bigram based statistical methods on the same corpus. The experiments show that the automatic rule-learning approach can achieve better prediction accuracy than the non-rule based methods and yet retain the advantage of the simplicity and understandability of rule systems. The paper proceeds as follows.</Paragraph>
    <Paragraph position="9"> Section 2 introduces the rule-learning algorithms we used. Section 3 describes prosodic phrase prediction and its evaluation parameters. Section 4 discusses the feature selection and rule-learning experiments in detail. Section 5 reports the evaluation results of rule based and none-rule based methods.</Paragraph>
    <Paragraph position="10"> Section 6 presents the conclusion and the view of future work.</Paragraph>
  </Section>
class="xml-element"></Paper>