臺灣博碩士論文加值系統

面試是一種很常被使用的入學管道，在真正的面試過程中，面試官通常會根據他本身的背景知識，針對面試者先前的回答中特定的主題或特定的文字內容進行追問。而當面試者先前回答有多於一個句子時，面試官也會選擇一項特定句子做追問。因此對於一個面試練習系統，如何使系統擁有完整的背景知識且藉此產生追問句是相當重要的一環。本論文提出一個能夠使用知識本體產生追問句的面試練習系統，以幫助學生能在入學面試前多加練習如何回答問題，增加技巧及信心。
針對不同模型，本論文使用了三種不同的語料，第一種為MHMC面試對話語料庫，此語料透過每次由兩位學生來扮演面試官與面試者的方式收集完成，總共收集了260場面試，其中會專注於有被面試官追問的面試者回答上。第二項語料為ConceptNet所有中文資源描述框架三元體(RDF triple)，扣除不需要的關係(Predicate)後共362,414個Triple。第三項語料為由八位學生針對MHMC面試對話語料庫中適合被追問的回答句，並根據知識本體而建立的追問句，共有6,943個ADQ pairs。
本論文之研究主題為資訊工程領域知識本體的擴充及面試練習系統的追問句生成。在生成追問句之前，需先決定要針對使用者回答中的哪一個句子做追問，本論文採用卷積神經張量網路(Convolutional Neural Tensor Network, CNTN)，以預測哪個句子最適合作為追問對象句。而在知識本體擴充方面，將以ConceptNet中的RDF Triple作為訓練語料，並應用神經張量網路(Neural Tensor Network, NTN)針對ConceptNet中的每種關係(Predicate)分別訓練，藉此學習輸入的兩個詞是否具有關係，以作為知識本體擴充的模型，並且建立資訊工程領域的知識本體。最後在追問句生成之部分，首先針對ConceptNet中的每種關係建立不同的模板，接著利用追問對象句中的詞彙查詢資訊工程領域知識本體，以得到相關的Triple，並且將這些Triple填入對應的模板中，以得到對應數量的追問句，最後利用CNTN句子相似度比較模型以找出一個與追問對象句最相關的句子作為系統最終的追問句。
在評量上採用五次交叉驗證來做實驗評估。本論文使用的架構表現皆比傳統方法要來的好，在選擇追問對象句的準確度有81.94%，而句子相似度比較的準確度高達92.28%。在NTN模型方面則是能夠有效的對知識本體進行擴充，雖然在部分Predicate為傳統方法的表現較佳，但可看出NTN模型是平均表現最良好也最穩定的。另外本論文也施測主觀分析，也從結果得知，在擴充知識本體的實驗上能有效建立出可靠的Triple。最後生成出與使用者回答句相關的追問句，該句不但包含Triple的資訊，而且具有良好的語意。

Admissions interviews are an essential part of the application process for students pursuing a quality education. In a real-world interview process, when the interviewee’s response consists of more than one sentence, the interviewers will often choose the sentence that most directly relates to their personal background knowledge to inquire more about. In most question generation systems, the aim is to generate follow-up questions which are related to the meaning beyond the literal content of any given sentence. Thus, an ontology for the interview domain is necessary to provide more information in order to generate appropriate follow-up questions. This thesis proposed a method to generate follow-up questions using a specific ontology to help the students majoring in computer science (CS) practice how to answer questions during an interview.
In this thesis three different corpuses are applied. The first is MHMC Interview Corpus (MHMC-IV Corpus). In this particular corpus, two students take turns role playing as an interviewer and an interviewee to complete a set of 260 interview dialogs. The second is Chinese RDF Triples collected from ConceptNet. There are 362,414 Triples after removing those with unsuitable predicate. The last is Answer sentence and follow-up question pairs (ADQ pairs) collected by 8 students using the Answer turn in MHMC-IV Corpus and relevant Triples to make follow-up questions. Finally, 6,943 ADQ pairs are collected.
The main purpose of this thesis is to use a domain ontology to allow the interview coaching system to generate follow-up questions which are more related to the implied or intended meaning, beyond the literal content, of the user’s previous answer. A Convolutional Neural Tensor Network (CNTN) is used to select a key sentence to ask a follow-up question about; then a Neural Tensor Network (NTN) is used to model the relationship between Subjects and Objects for each predicate from the ConceptNet. The newly generated NTN model is then used to populate the ontology in CS domain. After extracting the words in the key sentence to query the domain ontology for relevant triple retrieval, these Triples are inputted into the slots in the question templates as follow-up questions. Finally, the CNTN sentence matching model is employed to choose the question most related to the user’s Answer sentence as the final follow-up question.
This thesis applied 5-fold cross validation for evaluation. According to the results, the question generation performance in the proposed model is higher than the traditional method. The accuracy of the key sentence selection model achieves about 81.94%, and the accuracy of the sentence matching model achieve about 92.28%. Subject evaluation of the populated Triples found the results to be acceptable. Moreover, the generated follow-up questions are relevant to the selected Triples, which are related to the figurative meaning, beyond the literal content, of the user’s previous answer.

摘要 I
Abstract III
誌謝 V
Contents VII
List of Tables IX
List of Figures X
Chapter 1 Introduction 1
1.1 Background 1
1.2 Motivation 2
1.3 Literature Review 4
1.3.1 Follow-up Question Generation 4
1.3.2 Sentence Generation 4
1.3.3 Method to Construct an Ontology 6
1.4 Problems and Proposed Method 7
1.5 Research Framework 10
Chapter 2 Database 12
2.1 Data Collection 12
2.1.1 Collection of MHMC Interview Corpus 12
2.1.2 Collection of RDF Triple from ConceptNet 14
2.1.3 Collection of Ontology-based A-DQ Data (A,DQ pair) 15
2.2 Data Processing and Analysis 18
2.2.1 Processing and Analysis of MHMC Interview Corpus 18
2.2.2 Processing and Analysis of RDF Triple from ConceptNet 19
2.2.3 Processing and Analysis of Ontology-based A-DQ Data 20
Chapter 3 Proposed Method 21
3.1 Data Preprocessing 22
3.2 Key Sentence Selection 25
3.2.1 Convolutional Neural Tensor Network (CNTN) 25
3.2.2 Key Sentence Selection – CNTN 28
3.3 Ontology Extraction and Population 30
3.3.1 Initial Ontology Extraction 30
3.3.2 Ontology Triple Population 31
3.4 Sentence Similarity Matching 33
Chapter 4 Experimental Results and Discussion 36
4.1 Key Sentence Selection 36
4.2 Ontology Triple Population 38
4.2.1 NTN Model Evaluation 38
4.2.2 Experiment of Ontology Triple Population 40
4.3 Sentence Similarity Matching 43
4.3.1 CNTN Model Evaluation 43
4.3.2 Follow-up Question Generation Evaluation 45
4.3.3 System Output Follow-up Question Evaluation 47
Chapter 5 Conclusion and Future Work 49
參考文獻 51

[1]Y. Mu, and Y. Yin, Task-oriented spoken dialogue system for humanoid robot, in Multimedia Technology (ICMT), 2010, pp. 1-4.
[2]M. Koshinda, M. Inaba, and K. Takahashi, Machine-learned ranking based non-task-oriented dialogue agent using twitter data, in Web Intelligence and Intelligent Agent Technology (WI-IAT), 2015, pp. 5-8.
[3]Maruti Techlabs. What is a Conversational UI and why does it matter, 2017, Available: https://chatbotsmagazine.com/what-is-a-conversational-ui-and-why-it-matters-de358507b9a2
[4]X. Liu, and W. Zhao, Buddy: A Virtual Life Coaching System for Children and Adolescents with High Functioning Autism. in Dependable, Autonomic and Secure Computing, 15th Intl Conf on Pervasive Intelligence ＆ Computing, 3rd Intl Conf on Big Data Intelligence and Computing and Cyber Science and Technology Congress (DASC/PiCom/DataCom/CyberSciTech), 2017, pp. 293-298.
[5]吳元熙. 不只是客服而已, CHATISFY聊天機器人還能幫你下單, Oct. 11, 2017, Available: https://www.bnext.com.tw/article/46496/chatisfy-offered-chat-commerce-service-with-facebook-messenger
[6]P. Cimiano, C. Unger, and J. McCrae, Ontology-based interpretation of natural language, in Synthesis Lectures on Human Language Technologies, 2014, pp.1-178.
[7]T. R. Gruber, A translation approach to portable ontology specifications. in Knowledge acquisition, 1993, pp. 199-220.
[8]R. A. Kadir, and R. A. Yauri, Resource description framework triples entity formations using statistical language model, in Journal of Fundamental and Applied Sciences, 2017, pp. 710-729.
[9]許秩維, 大學個人申請, 通過篩選比率8成1創新高, Mar. 03, 2018, Available: http://www.cna.com.tw/news/firstnews/201803280027-1.aspx
[10]Palladian. Available: http://www.palladiancr.com/
[11]顏聖紘, 高中競相找大學教授模擬面試真的有用嗎, Apr. 03, 2018, Available: https://opinion.udn.com/opinion/story/7492/3068260
[12]劉漢民, 面試實戰技法, 清華大學出版社, 2015.
[13]J. D. Moore, and V. O. Mittal, Dynamically generated follow-up questions, in Computer, 1996, pp. 75-86.
[14]R. Das, A. Ray, S. Mondal, and D. Das, A rule based question generation framework to deal with simple and complex sentences, in Advances in Computing, Communications and Informatics (ICACCI), 2016, pp. 542-548.
[15]M. H. Chu, W. Y. Chen, and S. D. Lin, A Learning-Based Framework to Utilize E-HowNet Ontology and Wikipedia Sources to Generate Multiple-Choice Factual Questions, in Technologies and Applications of Artificial Intelligence (TAAI), 2012, pp. 125-130.
[16]廣義知網知識本體架構2.0. Available: http://ehownet.iis.sinica.edu.tw/index.php
[17]T. H. Wen, M. Gasic, D. Kim, N. Mrksic, P. H. Su, D. Vandyke, and S. Young, Stochastic language generation in dialogue using recurrent neural networks with convolutional sentence reranking, arXiv:1508.01755, 2015.
[18]T. H. Wen, M. Gasic, N. Mrksic, P. H. Su, D. Vandyke, and S. Young, Semantically conditioned lstm-based natural language generation for spoken dialogue systems, arXiv:1508.01745, 2015.
[19]Y. Zhang, Z. Gan, and L. Carin, Generating text via adversarial training, in NIPS workshop on Adversarial Training (Vol. 21), 2016.
[20]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser and I. Polosukhin, Attention is all you need, in Advances in Neural Information Processing Systems, 2017, pp. 6000-6010.
[21]DBpedia. Available: https://wiki.dbpedia.org/
[22]ConceptNet. Available: http://conceptnet.io/
[23]G. Petasis, V. Karkaletsis, G. Paliouras, A. Krithara, and E. Zavitsanos, Ontology population and enrichment: State of the art, in Knowledge-driven multimedia information extraction and ontology evolution, 2011, pp. 134-166
[24]J. A. Khan, ＆ S. Kumar, Deep analysis for development of RDF, RDFS and OWL ontologies with protégé, in Reliability, Infocom Technologies and Optimization (ICRITO)(Trends and Future Directions), 2014, pp. 1-6.
[25]L. De Silva, and L. Jayaratne, WikiOnto: A system for semi-automatic extraction and modeling of ontologies using Wikipedia XML corpus, in Semantic Computing, 2009, pp. 571-576.
[26]Z. Lin, R. Lu, Y. Xiong, and Y. Zhu, Learning ontology automatically using topic model, in Biomedical Engineering and Biotechnology (iCBEB), 2012, pp. 360-363.
[27]X. Qiu, and X. Huang, Convolutional Neural Tensor Network Architecture for Community-Based Question Answering in International Joint Conference on Artificial Intelligence (IJCAI), 2015, pp. 1305-1311.
[28] R. Socher, D. Chen, C. D. Manning, and A. Ng, Reasoning with neural tensor networks for knowledge base completion, in Advances in neural information processing systems, 2013, pp. 926-934.
[29]J. Sun, ‘Jieba’ Chinese word segmentation tool, 2012.
[30]Y. Kim, Convolutional neural networks for sentence classification, arXiv:1408.5882, 2014.
[31]E-HowNet. Available:
http://ehownet.iis.sinica.edu.tw/index.php
[32]C. Xing, W. Wu, Y. Wu, J. Liu, Y. Huang, M. Zhou, and W. Y. Ma, Topic augmented neural response generation with a joint attention mechanism, arXiv preprint arXiv:1606.08340, 2016.