Investigating Effect of Rich Syntactic Features on Mandarin Prosodic Boundaries Prediction

Abstract

Previous researches indicated that syntactic information could improve the performance of automatic prosodic boundaries prediction. However, few researches focused on the usefulness of various syntactic features on different phrase boundaries prediction in detail, especially the verification on large-scale corpus. This paper investigates the effect of different syntactic feature combinations for phrase boundaries prediction based on a large-scale Mandarin corpus. Syntactic phrase structure and dependency relationship are both introduced and compared in the experiments. The evaluations of all the features in different time-span as well as feature importance are also carried out. The experimental results show that prediction models of prosodic word and prosodic phrase achieve the best performance with syntactic phrase and dependency features, while the models with dependency features outperform other models when predicting intonational phrase. Furthermore, the performance of intonational phrase is improved obviously by concerning syntactic global features. Meanwhile, the experiments also find that the relationship between syntactic local features and lower level prosodic boundaries: prosodic phrase and prosodic word, is closer.

BLSTM-CRF Based End-to-End Prosodic Boundary Prediction with Context Sensitive Embeddings in a Text-to-Speech Front-End

“…These studies investigate a great number of features and their relevance to prosodic boundaries prediction [3][4][5][6][7]. Recently, some syntactic features [8,9] and embedding features [10][11][12] have been employed to augment or replace the traditional linguistic features (like POS, word-terminal syllables etc.). The other category focuses on the aspect of modeling methods.…”

“…With the previous researches on prosodic boundary prediction, this task has been accomplished by dividing it into t-wo stages. The first stage is to extract rich features from raw texts [3][4][5][6][7][8][9][10][11][12], while the second stage takes as input the rich features and predicts corresponding boundary for each word. Even the newly developed methods [10][11][12] had to transform the raw texts to word embeddings first before feeding them into the prediction model.…”

“…Even the newly developed methods [10][11][12] had to transform the raw texts to word embeddings first before feeding them into the prediction model. This two-stage procedure is often timeconsuming and language-specific, since it requires much expert linguistic knowledge to define linguistic features [7][8][9]. In this paper, we integrate these two stages and learn the whole process end-to-end, which removes a major bottleneck in modeling prosodic boundary for new languages.…”

Improving Fluency of Spoken Mandarin for Nonnative Speakers by Prosodic Boundary Prediction Based on Deep Learning

“…Lu et al [20] model the relationships between prosodic boundaries and lexicon words for prosodic boundary prediction by combining self-attention with MTL and setting word segmentation as an auxiliary task. Because the prosodic structure of Mandarin is related to the syntactic structure [21], the traditional shallow linguistic feature has been augmented or replaced by embedding features and some syntactic features [22][23][24][25][26]. There is still room to improve the prosodic prediction of Chinese and use the predicted result to perfect international students' speaking Mandarin.…”