臺灣博碩士論文加值系統

目前已有許多閩南語處理相關研究，但對於台語古詩朗讀還未臻完善。本論文希望能提出一套設計以台語朗讀古詩詞的系統，以目前能取得的資料，會先以唐詩這類字數句數較規律的文體開始研究。
本論文研究分成兩部分，第一部分在於決定詩句當中每個字的讀音，尤其當有兩種以上發音的時候。第二部分在於決定朗讀詩句時，每個位置要唸本調或需要變調。
在決定讀音部分，本論文使用整理自楊青矗之「國台雙語辭典」和沈富進之「彙音寶鑑」的單字音典做為選取讀音的依據。單字音典中，每個字的各種發音都標記有文白讀、漳泉腔、常用讀音、來源字典等等資訊。當一個字有兩種以上之發音時，本論文實驗了各種不同的讀音選擇策略。除此之外，我們也測試改以實驗資料集做為選音範例的實驗，並且觀察範例來自相同或相異資料集對正確率的影響。實驗結果顯示，來源是相同語料集的統計模型系統效果最好，範例沒有時則選擇單字音典第一種發音效果最好。最佳的正確率為96.35%。
在決定是否變調的部分，本論文先觀察了詩句中位置與本、變調之間的關係，評估了選用大宗情形的效能。本論文另外提出數種特徵來預測各字是否讀變調或本調，包括出現位置、是否常為本調位置、斷詞情形、對句相對位置資訊、輕聲調用字等等，再用機器學習的方法訓練出分類器。實驗結果顯示，各系統在五言詩及七言詩判斷變調的效能互有好壞，機器學習所得之分類器效果略佳，對句相對位置資訊是最有幫助的特徵。

There are many studies dealing with Taiwanese, but not yet a text-to-speech for ancient poems. This thesis wants to propose a system which can read Chinese ancient poems in Taiwanese. We will start from the poems in Tang Dynasty, for these data are what we have found in the Internet, and the number of words and sentences are more regular.
This thesis deals with two major issue in reading Chinese ancient poems. The first issue is pronunciation choosing, especially when a Chinese character has more than one reading provided in a dictionary. The second issue is tone sandhi problem. We have to decide which tones of characters need to be changed and which will remain as original tones.
Regarding pronunciation choosing, this thesis proposed two approaches. The first approach is to choose pronunciation provided in a Taiwanese Dictionary for Chinese Characters, based on options such as ranking or literary/colloquial readings. The second approach is to choose the most frequent pronunciations in the training set. The source of the test set may be the same or different from the training set. We will see how it affects the evaluation results. The best system chooses the most frequent pronunciations in the training set from the same source, whose accuracy is 96.53%. When a character does not appear in the training set, the best system will choose its first pronunciation in the dictionary.
Regarding the tone sandhi problem, we first observed the tone changing rate at different position in a sentence of a poem and estimated how well would it be by simply choosing the majority. We also proposed many features to predict the position of tone changing by machine learning, including position, majority, word segmentation information, features from the paired sentence, and neutral tone characters. The experiment results show that the rule-based system and the best learning-based classifier outperformed each other in different datasets. The learning-based classifier achieved a little better performance, while features from the paired sentence were very useful.

摘要 i
Abstract ii
致謝 iii
目錄 iv
圖目錄 vi
表目錄 vii
第一章 緒論 1
1.1研究動機 1
1.2相關研究 1
1.3問題定義及論文架構 3
第二章 實驗資料來源 4
2.1鯤島園地本土文化網站 4
2.2台語詩詞朗誦YouTube頻道 7
2.2.1 spwang1000頻道 7
2.2.2 Janice Wu 頻道 8
2.3瑞峰國小母語教材 9
2.3.1台語吟唐詩 9
2.3.2台語唸經典 11
2.4單字音典 13
第三章 拼音系統轉換及實驗資料前處理 14
3.1擴充單字音典 14
3.2拼音系統轉換 15
3.3實驗資料聽打及確認本調 15
3.3.1 鯤島園地千家詩之本調讀音確認 16
3.3.2 YouTube頻道影片聽打 17
3.3.3 瑞峰國小台語資料之變調讀音確認 17
第四章 古詩朗讀台語選音實驗 19
4.1依單字音典附加資訊選音 19
4.2相同來源範例為本的選音策略 20
4.3不同來源範例為本的選音策略 21
第五章 古詩朗讀本變調預測實驗 24
5.1變調規則 24
5.1.1基本變調規則 24
5.1.2輕聲調變調規則 25
5.2僅考慮位置資訊的變調決定規則 26
5.3機器學習訓練變調分類器 27
5.3.1判斷本變調相關特徵 27
5.3.2本變調判斷實驗 30
第六章 結論及未來展望 32
參考文獻 33
附錄一、實驗資料集人工修正部份 34
附錄二 35
附錄三、不同來源範例選音實驗結果 37

Chuan-Jie Lin, Hsin-Hsi Chen (1999) "A Mandarin to Taiwanese Min Nan Machine Translation System with Speech Synthesis of Taiwanese Min Nan." IJCLCLP，4(1) (1999)
Yih-Jeng Lin, Ming-Shing Yu, Wei-Lun Li(2012) "Applying Association Rules in Solving the Polysemy Problem in a Chinese to Taiwanese TTS System"
[In Chinese]. ROCLING2012
Neng-Huang Pan, Ming-Shing Yu, Pei-Chun Tsai(2012) "A Prediction Module for Taiwanese Tone Sandhi Based on the Decision Tree Algorithm" [In Chinese]. ROCLING 2012
Ming-Shing Yu, Cheng-Rong Tsai (2008) "An Implementation of Toneless Input for Mandarin and Taiwanese" [In Chinese]. ROCLING 2008
Yih-Jeng Lin, Ming-Shing Yu, Chin-Yu Lin (2008) "Using Chi-Square Automatic Interaction Detector to Solve the Polysemy Problems in a Chinese to Taiwanese TTS System." ISDA (1) 2008: 362-367
Wei-jay Huang, Jhih-rou Lin, Ren-Yuan Lyu, Yuang-Chin Chiang, Jyh-Shing Roger Jang, Ming-Tat Ko (2012) "Automatic Time Alignment for a Taiwanese Read Speech Corpus and its Application to Constructing Audiobooks with Text-Speech Synchronization" [In Chinese]. ROCLING 2012
Ren-Yuan Lyu, Chi-yu Chen, Yuang-Chin Chiang, Min-shung Liang (2000) "A bi-lingual Mandarin/taiwanese (min-nan), large vocabulary, continuous speech recognition system based on the tong-yong phonetic alphabet(TYPA)." INTERSPEECH 2000: 226-229
Yu-Jhe Li, Chung-Che Wang, Liang-Yu Chen, Jyh-Shing Roger Jang, Ren-Yuan Lyu (2013) "Using Speech Assessment Technique for the Validation of Taiwanese Speech Corpus" [In Chinese]. ROCLING 2013
Dau-Cheng Lyu, Hong-Wen Hsien, Yung-Xian Lee, Zhong-Ing Liou, Chun-Nan Hsu, Yung-Jien Chiang, Ren-Yuan Lyu (2004) "The study of pronunciation variations in Mandarin and Taiwanese and its application in PDA" [In Chinese]. ROCLING 2004
黃志超 (2015) 範例為本的國語─台語翻譯之研究，碩士論文，國立臺灣海洋大學。