臺灣博碩士論文加值系統

本論文之研究主題為語音文件自動摘要，包括使用結構性支撐向量機、領域調適並接上抽象式摘要的新架構，並以中文課程錄音作為實驗語料，包含語音訊號以及語音辨識轉寫~(Transcription)~。我們首先使用語意、韻律、相似度三大類特徵以不同的機器學習模型做督導式摘要抽取，包括支撐向量機~(Support Vector Machine)~、排序支撐向量機~(Rank Support Vector Machine)~、以及結構支撐向量機~(Structured Support Vector Machine)~，並和常用的非督導式最大邊際關聯法~(Maximum Margin Relevance)~比較。實驗顯示結構式支撐向量機可整體學習~(Joint Learn)~出原本需人為定義的參數，且優於其他督導式及非督導式的方法。此外，對各類特徵進行分析時，也發現韻律特徵對文件摘要有極大幫助，且各類特徵有加成性，串接後對文件摘要皆有幫助。

督導式摘要的最大問題在於人工標註摘要的取得不易，且不同語料的人工標註答案都是有差異~(Bias)~的。因此，本論文進一步研究領域調適之方法，利用其他領域的人工標註摘要資訊，迭代訓練結構式支撐向量機並利用特徵轉換減少領域間資料分佈的差異。在抽取課程錄音摘要的實驗中，透過利用新聞語料的人工標註答案和領域調適，發現可以得到相當大的進步。

最後，本論文對抽象式文件摘要提出了一套新架構並作了初步研究，有效利用詞與詞前後文的關係建圖~(Graph)~，圖中節點代表詞並在上面做搜尋產生候選語句，最後利用詞N連語言模型、POS N連語言模型以及主題模型對候選語句做重排序，對將來抽象式摘要的研究提供了良好的起點。

誌謝.......................................... i
中文摘要....................................... ii
一、導論....................................... 1
1.1 研究背景.................................. 1
1.2 本論文研究方向.............................. 2
1.3 本論文研究貢獻.............................. 3
1.4 章節安排.................................. 4
二、背景知識介紹.................................. 5
2.1 語音文件摘要背景介紹與相關研究 ................... 5
2.1.1 文件摘要之分類及應用...................... 5
2.1.2 文件摘要之方法.......................... 7
2.1.3 語音文件摘要與文字摘要之不同 ................ 10
2.1.4 傳統非督導式語音文件摘要方式-最大邊際關聯法(MMR) . . . 11
2.2 結構式支撐向量機............................. 12
2.2.1 機器學習.............................. 13
2.2.2 支撐向量機 ............................ 14
2.2.3 鑒別函數(DiscriminativeFunction) ............... 22
2.2.4 減損函數(LossFunction) ..................... 23
2.3 文件摘要評估方式............................. 24
2.3.1 準確率 (Precision) 、召回率 (Recall) 與F評估 (F-measure) . . 25
2.3.2 ROUGE-N ............................. 25
2.4 本章總結.................................. 26
三、利用結構支撐向量機及韻律特徵之抽取式語音文件摘要 . . . . . .27
3.1 簡介..................................... 27
3.2 督導式模型融入最大邊際關聯法..................... 27
3.3 結構式支撐向量機及目標函數之定義 .................. 29
3.4 抽取特徵.................................. 31
3.4.1 語意特徵.............................. 31
3.4.2 相似度特徵 ............................ 33
3.4.3 韻律特徵.............................. 34
3.4.4 其他特徵.............................. 38
3.5 實驗基礎設置 ............................... 39
3.5.1 實驗語料與辨識.......................... 39
3.5.2 參考摘要之形成.......................... 39
3.5.3 實驗配置.............................. 40
3.5.4 評估方式.............................. 40
3.6 實驗結果與分析 .............................. 40
3.6.1 基準實驗.............................. 40
3.6.2 初步特徵效力分析與串接 .................... 43
3.6.3 結構式支撐向量機法 ....................... 44
3.6.4 韻律特徵效力分析 ........................ 45
3.7 本章總結.................................. 46
四、使用結構支撐向量機作領域調適於跨語料之抽取式語音文件摘要 . . .47
4.1 簡介..................................... 47
4.2 架構與流程(Framework).......................... 48
4.3 領域調適.................................. 49
4.3.1 自我標註.............................. 49
4.3.2 特徵轉換.............................. 51
4.4 特徵抽取.................................. 53
4.4.1 領域獨有特徵 ........................... 53
4.4.2 共有特徵.............................. 54
4.5 實驗基礎設置 ............................... 56
4.5.1 實驗語料與辨識系統 ....................... 56
4.5.2 參考摘要之形成.......................... 56
4.5.3 實驗配置.............................. 57
4.5.4 評估方法.............................. 57
4.6 實驗結果與分析 .............................. 57
4.6.1 跨領域特徵效能分析 ....................... 57
4.6.2 自我標記以及領域獨有特徵之效能分析 ............ 60
4.6.3 特徵轉換.............................. 60
4.7 本章總結.................................. 63
五、以圖建構抽象式文件摘要之新架構...................... 64
5.1 簡介..................................... 64
5.2 架構與流程(Framework) ......................... 65
5.3 候選語句生成 ............................... 66
5.3.1 前處理 ............................... 66
5.3.2 建圖 ................................ 67
5.3.3 光束搜尋.............................. 68
5.4 抽象摘要生成 ............................... 69
5.4.1 訓練模型.............................. 69
5.4.2 候選語句評估 ........................... 71
5.5 實驗基礎設置 ............................... 72
5.5.1 使用語料.............................. 72
5.5.2 參數配置.............................. 72
5.5.3 實驗評估.............................. 73
5.6 實驗結果.................................. 73
5.7 小結..................................... 75
六、結論與展望 ................................... 76
6.1 結論..................................... 76
6.2 未來研究方向 ............................... 77
參考文獻....................................... 79

[1] http://www.ebay.com/.
[2] http://www.agoda.com.tw/.
[3] https://www.coursera.org/.
[4] T. Hazen J. Glas and R. Barzilay, “Recent progress in the mit spoken lecture pro- cessing project,” in Proceedings of 8th An- nual Conference of the International Speech Communication Association, 2007.
[5] Ani Nenkova, “A survey of text summarization techniques,” in Multimedia, 2001.
[6] Yang Liu and Dikel Hakkani-Tur, “Spoken language understanding.(ch13 ),” 2005,
vol. 40, pp. 1–60.
[7] http://www.bing.com/.
[8] Asli Celikyilmaz and Dilek Hakkani-Tur, “Concept-based classification for multi- document summarization,” in ICASSP, 2011.
[9] Asli Celikyilmaz and Dilek Hakkani-Tur, “A hybrid hierarchical model for multi- document summarization,” in ACL, 2010.
[10] Ying-Lang Chang and Jen-Tzung Chien, “Latent dirichlet learning for document summarization,” in ICASSP, 2009.
[11] R. Motwani L. Page, S. Brin and T. Winograd, “The pagerank citation ranking: bringing order to the web,” in Stanford Digital Library Technologies Project, 1998.
79
[12] Ching-Feng Yeh Yun-Nung Chen, Yu Huang and Lin-Shan Lee, “Spoken lecture summarization by random walk over a graph constructed with automatically ex- tracted key terms,meeting utterences by importance,” in INTERSPEECH, 2011.
[13] Yun-Nung Chen Hung-Yi Lee and Lin-Shan Lee, “Improved speech summariza- tion and spoken term detection with graphical analysis of utterance similarities,” in APSIPA ASC, 2011.
[14] Michel Galley, “Intra-speaker topic modeling for improved multi-party meeting summarization with integrated random walk,” in NAACL-HLT, 2012.
[15] Yun-Nung Chen and Florian Metze, “Integrating intra-speaker topic modeling and temporal-based inter-speaker topic modeling in random walk for imporved multi- party meeting summarization,” in INTERSPEECH, 2012.
[16] Yun-Nung Chen and Florian Metze, “Two-layer mutually reinforced random walk for improved multi-party meeting summarization,” in SLT, 2012.
[17] G. Erkan and D. R. Radev, “Lexrank: Graph-based centrality as salience in text summarization,” in JAIR, 2004.
[18] Jeff Bilmes Hui Lin and Shasha Xie, “Graph-based submodular selection for extrac- tive summarization,” in ASRU, 2009.
[19] Korbinian Reidhammer Nikhil Garg, Benoit Favre and Dilek Hakkani-Tur, “Clus- terrank: A graph based method for meeting summarization,” in Interspeech, 2009.
[20] Michel Galley, “A skip-chain conditional random field for ranking meeting utter- ences by importance,” in EMNLP, 2006.
80
[21] Trevor Cohn Ahmet Aker and Robert Gaizauska, “Multi-document summarization using a* search and discriminative training,” in EMNLP, 2010.
[22] Berlin Chen and Shih-Hsiang Lin, “A risk-aware modeling framework for speech summarization,” in IEEE Transactions on Audio, Speech Language Processing, 2012.
[23] Shih-Hsiang Lin Yueng-Tien Lo and Berlin Chen, “Constructing effective ranking models for speech summarization,” in ICASSP, 2012.
[24] Shih-Hsiang Lin and Berlin Chen, “A risk minimization framework for extractive speech summarization,” in ACL, 2010.
[25] Shih-Hsiang Lin Berlin Chen and Jia-Wen Liu, “Extractive speech summarization using evaluation metric-related training criteria,” in Inf. Process. Manage, 2013.
[26] Shasha Xie and Yang Liu, “Improving supervised learning for meeting summariza- tion using sampling and regression,” 2010.
[27] Shasha Xie and Dilek Hakkani-Tur, “Leveraging sentence weights in a concept- based optimization framework for extractive meeting summarization,” in Inter- speech, 2009.
[28] Dan Gillick and Dilek Hakkani-Tur, “A global optimization framewok for meeting summarization,” in ICASSP, 2009.
[29] Shih-Hsiang Lin and Berlin Chan, “A risk minimization framework for extractive speech summarization,” in Association of Computation and Linguistic, 2010.
81
[30] Shih-Hsiang Lin and Berlin Chan, “A risk-aware modeling framework for speech summarization,” in IEEE Transaction on Audio Speech and Language Processing, 2012.
[31] Berlin Chan, “Extractive speech summarization – from the view of decision theory,” in Association of Computation and Linguistic, 2010.
[32] Klaus Zechner and Alex Waibel, “Minimizing word error rate in textual summaries of spoken language,” in ACL, 2000.
[33] Justin Zhang and Pascale Fung, “Extractive speech summarization by active learn- ing,” in ASRU, 2009.
[34] Yang Liu and Shasha Xie, “Impact of automatic sentence segmentation on meeting summarization,” in ACL, 2008.
[35] Bin Li Yang Liu, Feifan Liu and Shasha Xie, “Do disfluencies affect meeting sum- marization: A pilot study on the impact of disfluencies,” in IEEE SLT, 2007.
[36] Shasha Xie Yang Liu and Fei Liu., “Using n-best recognition output for extractive summarization and keyword extraction in meeting speech,” in ICASSP, 2010.
[37] Shasha Xie and Yang Liu., “Using confusion networks for speech summarization,” in ACL, 2010.
[38] Sameer Maskey, “Comparing lexical, acoustic/prosodic structural and discourse features for speech summarization,” in Interspeech, 2005.
[39] Jian Zhang, “Speech summarization without lexical features for mandarin broadcast news.,” in HLT-NAACL, 2007.
82
[40] ShaShaXieandYangLiu,“Acomparativestudyonspeechsummarizationofbroad- cast news and lecture speech,” in ASRU, 2009.
[41] ShaSha Xie and Yang Liu, “Integrating prosodic features in extractive meeting sum- marization,” in Interspeech, 2010.
[42] J. Hirschberg, “Communication and prosody: functional aspects of prosody,” in Speech Communication, 2002, pp. 31–43.
[43] H. Chan J. Zhang and P. Fung, “Improving lecture speech summarization using rhetorical information,” in Proceedings of Automatic Speech Recognition and Un- derstanding, 2007.
[44] Gerald Penn Xiaodan Zhu and F. Rudzicz, “Summarizing multiple spoken docu- ments: Finding evidence from untranscribed audio,” in ACL, 2009.
[45] Jaime Carbonell and Jade Goldstein, “The use of mmr, diversity-based reranking for reordering documents and producing summaries,” in ACM SIGIR, 1998.
[46] Shasha Xie and Yang Liu, “Using corpus and knowledge-based similarity measure in maximum marginal relevance for meeting summarization,” in ICASSP, 2008.
[47] Ruben Sipos Pannaga Shivaswamy and Thorsten Joachims, “Large-margin learning of submodular summarization methods,” 2011.
[48] Thorsten Joachims Thomas Hofmann and Yasemin Altun, “Support vector machine learning for interdependent and structured output spaces,” in International Confer- ence on Machine Learning, 2004.
83
[49] CHRISTOPHER J.C. BURGES, “A tutorial on support vector machines for pattern recognition,” in Data Mining and Knowledge Discovery, 1998, pp. 121–167.
[50] Chin-Yew Lin, “Rouge: A package for automatic evaluation of summaries,” in SLT, 2004.
[51] Berlin Chen Shih-Hsiang Lin and H. Min Wang., “A comparative study of prob- abilistic ranking models for chinese spoken document summarization.,” in ACM Transactions on Asian Language Information Processing., 2009.
[52] KubotaAndoandTongZhang,“Aframeworkforlearningpredictivestructuresfrom multiple tasks and unlabeled data,” in JMLR, 2005.
[53] Ryan McDonald John Biltzer and Fernando Perir, “Domain adaptation with struc- ture correspondence learning,” in Association of Computation and Linguistic, 2006.
[54] Mark Dredze John Biltzer and Fernando Pereira, “Domain adaptation for sentiment classification,” in Association of Computation and Linguistic, 2007.
[55] John Biltzer, “Domain adaptation of natural language processing systems,” in As- sociation of Computation and Linguistic, 2008.
[56] Oana Sandu and Giuseppe Carenini, “Domain adaptation to summarize human con- versations,” in DANLP, 2010.
[57] Mingli Wu Kam-Fai Wong and Wenjie Li., “Extractive summarization using super- vised and semi-supervised learning,” in ACL, 2008.
[58] Mingli Wu Kam-Fai Wong and Wenjie Li, “Extractive summarization using super- vised and semi-supervised learning,” in INTERSPEECH, 2008.
84
[59] Hui Lin Shasha Xie and Yang Liu, “Semi-supervised extractive speech summariza- tion via co-training algorithm,” in INTERSPEECH, 2010.
[60] S. Kong Y. Chen, Y. Huang and L. Lee, “Automatic key term extraction from spo- ken course lectures using branching entropy and prosodic/semantic features,” in Proceedings of Spoken Language Technology, 2010.
[61] Hung-Yi Lee Yun-Nung Chen, Yu Huang and Lin-Shan Lee, “Unsupervised two- stage keyword extraction from spoken documents by topic coherence and support vector machine,” in ICASSP, 2012.
[62] A. Nenkova and R. Passonneau, “Evaluating content selection in summarization: The pyramid method,” in HLT-NAACL, 2004.
[63] R. Passonneau A. Nenkova and K. McKeown, “The pyramid method: Incorporating human content selection variation in summarization evaluation,” in ACM, 2007.
[64] Alexander Rong Jin and G. Hauptmann, “Automatic title generation for spoken broadcast news.,” 2001.
[65] Katja Filippova, “Multi-sentence compression: Finding shortest paths in word graphs,” in International Conference on Computational Linguistics, 2010.
[66] Fei Liu and Yang Liu, “From extractive to abstractive meeting summaries: Can it be done by sentence compression?,” in ACL, 2009.
[67] Fei Liu and Yang Liu, “Using spoken utterance compression for meeting summa- rization: A pilot study,” in IEEE SLT, 2010.
85
[68] Dan Gillick Taylor Berg-Kirkpatrick and Dan Klein, “Jointly learning to extract and compress,” in ACL HLT, 2011.
[69] ChengXiang Zhai Kavita Ganesan and Jiawei Han., “Opinosis: A graph-based ap- proach to abstractive summarization of highly redundant opinions,” in International Conference on Computational Linguistics, 2010.
[70] Kavita Ganesan and ChengXiang Zhai, “Micropinion generation: An unsupervised approach to generating ultra-concise summaries of opinions,” in WWW, 2012.