Research On The Prosody Boundary Prediction For Foreign Students Speaking Mandarin

The development of the China-proposed Belt and Road Initiative has expanded the scale of foreign students in China,making the status of Chinese in the world languages gradually improved.However,when foreign students learn Chinese,they are always affected by the pronunciation habits of their native language.Predicting the prosodic structure from the Chinese text can help foreign students improve their Chinese proficiency and make them speak Mandarin with cadence.This thesis analyzes the relationship between the Mandarin syntactic structure and prosodic structure with foreign students as the research object,focusing on the prediction of prosodic words and prosodic phrases and finally evaluates the fluency scores of foreign students speaking Mandarin.The research results of the thesis have important theoretical significance and application value for revealing the relationship between text and speech and improving the oral fluency of foreign students.The main works and originality of this thesis are as follows:Firstly,a large-scale text corpus and a 1300-sentence recording corpus of foreign students have been established.The text corpus contains 100,000 Chinese sentences labeled with part-of-speech and prosodic boundary,which manually marked under the guidance of linguistic experts.In the recording corpus,a total of 6 foreign students and1 native Chinese speaker recorded voices.The recording corpus is statistically evaluated from the coverage and comprehensiveness of syllables and phonemes.The final result shows that these two corpora can be used for the research of Mandarin prosody structure prediction and the evaluation of foreign students speaking Mandarin.Secondly,the prediction of Mandarin prosody boundary based on deep learning is realized.Three methods of Mandarin prosody boundary prediction based on deep neural networks are Bi-directional Long Short-term Memory(Bi LSTM)model,Sequence to sequence(Seq2seq)model and Sequence to sequence with Attention(Seq2seq＿Attention)model.At the same time,a feature for prediction of Mandarin prosodic phrases,named Syntactic Hierarchical Number(SHN),is proposed to describe the relationship between the syntactic structure and prosodic structure of Mandarin sentences.Combining different prosodic features such as part of speech and word length,the boundary prediction experiments of prosodic words and prosodic phrases are carried out respectively.The experimental results show that the Seq2seq＿Attention model performs best in the prediction of prosodic words with F1-score of 98.14%.The Seq2seq＿Attention model with SHN(Seq2seq＿Attention＿SHN)is more effective than other methods in prediction of prosodic phrases with F1-score of 83.12%.Finally,the fluency of foreign students speaking Mandarin is evaluated.In order to verify that improvement of the Mandarin prosody boundary prediction methods on the oral level of foreign students,100 sentences with prosodic boundary annotations are selected and recorded by foreign students in the experimental results of prosody boundary prediction.The pronunciation level was automatically evaluated by the voice evaluation software system for the fluency score(value range 0～100).The speech evaluation results show that the oral fluency score of foreign students has increased by7.31-15.30 and average score increase of 12.11,which proves that the research work in this thesis can help foreign students master the prosodic structure and express better oral level.