臺灣博碩士論文加值系統

自然語言檢索是使用人類語言問句搜尋答案。其目的是要擷取適合使用者的答案。相較於一般的短語查詢，自然語言問句查詢則是更能直覺的呈現出使用者的問題。例如，問句「台南哪裡適合約會?」。但現今的搜尋引擎在自然語言檢索方面的表現往往不佳，導致使用者必須花大部分的時間在瀏覽出現部分詞彙網頁答案。在本論文，自然語言問句搜尋被定義為使用者想獲得一系列同性質的實體，並與ROSE等人提出條列式資訊目的定義相符。
在仔細的分析自然語言問句後，問句結構可以被分為問句上下文、問句實體、問句活動以及問句情緒。並且我們發現到了活動跟情緒之間有一些關係存在，提出了一個自動化的方式找出兩者的關聯來擴充問句。我們基於使用者需求分析問句中的活動以及隱含的情緒，並推薦適合的實體。我們使用適合中文的問句分析演算法辨識以上問句結構特徵。至於答案結構可以被分為上下文的證明頁面、實體類型、實體活動以及實體情緒詞總結。我們利用問句結構和答案結構的關係，結合實際上的網路的評價意見，建構出以活動及情緒為基礎的實體排序模型(ASERM)來改善實體搜尋。
實驗結果顯示我們提出的模型ASERM可以幫助使用者找到他們想要的實體。也顯示出我們的系統著實能增進實體搜尋的效能

Natural language search is to use human language questions as query to search answers. The task of natural language search is to extract suitable answers for users. Compare with short query, Natural language query users can directly submit their query intents. For example, the question, “Which places in Tainan are appropriate for dating?” But conventional search engines can’t efficiently process natural language queries and users can’t get good search results. So users need to spend lots of time on browsing and filter the result pages, which may involve some noise information. In those natural language question search, user wants to obtain a list of homogeneous entities. According to Rose et al. proposed list-informational goal definition, natural language question search is properly matched list-informational goal.
After advanced analysis on natural language question structure, the question structure can be divided into four parts, they are question context, question entity, question activity and question sentiment. We noted that there are some relations between activity and sentiment. Then, we proposed an automated method to identify related sentiments and use them to expand questions. We try to recommend suitable entities based on the intent of users by analyzing the activity which user want to do and the sentiments which behind the activity. We used the algorithm of question analysis to identify question features. As to the answer structure, it can be divided into context evidence page, entity type, entity activity and entity sentiment summation to match the question structure. We combine the relationship between the question structure and the answer structure to construct Activity-Sentiment-based Entity Ranking Model (ASERM) to improve entity search.
Experiment result shows that our proposed method ASERM can help user to get entity list which matched their intent. And it shows ASERM really can enhance performance in entity search.

摘要 III
Abstract IV
誌謝 VI
Table of Contents VIII
List of Tables X
List of Figures XII
Chapter 1. Introduction 1
1.1 Motivation 1
1.2 Methodology 4
1.3 Organization of this Dissertation 4
Chapter 2. Related Work 5
2.1 List-Informational Natural Language Search 5
2.2 Question Structure 6
2.3 Opinion and Sentiment 7
2.3.1 Opinion Extraction 7
2.3.2 Opinion Mining and Sentiment Definitions 8
Chapter 3 Method 10
3.1 Observation 10
3.2 Problem Formulation 13
3.3 System Architecture 16
3.3.1 Question Analysis 16
3.3.2 Activity-Sentiment-Entity Base Training 17
3.3.3 Hidden Sentiment and Activity Expansion 18
3.3.4 Activity Sentiment Based Entity Ranking Model 18
3.4 Question Analysis 19
3.4.1 Identification of Question Entity 20
3.4.2 Identification of Question Context 21
3.4.3 Identification of Question Activity and Question sentiment 22
3.5 Activity-Sentiment-Entity Base Training 25
3.5.1 Database Architecture 25
3.5.2 Activity-Sentiment Relation 26
3.5.3 Entity-Activity-Sentiment Training 27
3.6 Hidden Sentiment and Activity Expansion 30
3.7 Hidden Sentiment and Activity Expansion 33
3.7.1 Model Inference 33
3.7.2 Context Evidence Model 35
3.7.3 Entity Analysis Model 35
Chapter 4 Experimental 41
4.1 Experiment Setup 41
4.1.1 Data Set 41
4.1.2 Evaluation Metrics 42
4.2 Evaluation of Question Analysis 43
4.3 Evaluation of Hidden Sentiment and Activity Expansion 50
4.4Evaluation of Activity-Sentiment-based Entity Ranking Model 52
4.4.1 Parameter Estimation 52
4.4.2 Baseline 55
4.4.3 Experimental Result 56
Chapter 5 Conclusion 65
Reference 66

[1] Agichtein E., Castillo C., Donato D., Gionis A., and Mishne G. Finding high-quality content in social media. WSDM, 2008.
[2] Broder A., A taxonomy of web search. SIGIR Forum, 2002
[3] Bierner G., Alternative Phrases and Natural Language Information Retrieval, ACL, 2001
[4] Chu C. J., Prager J., Ravin Y. and Cesar C. A Hybrid Approach to Natural Language Web Search, EMNLP,2002
[5] Chan W., Yang W., Tang J., Du J., Zhou X., and Wang W. Community Question Topic Categorization via Hierarchical Kernelized Classification, CIKM’13
[6] Elkan C., Log-linear models and conditional random fields, CIKM, 2008
[7] Ferret O., Grau B., Hurault-Planet M., Illouz G., Monceaux L., Robba I. and Vilnat A., Finding an Answer based on the Recognition of the Question Focus, 2001
[8] Feng, S., Wang, D., Yu, G., Yang, C., and Yang, N., Chinese blog clustering by hidden sentiment factors, ADMA, 2009.
[9] Huang J.Y, Using Opinion-based Entity Ranking Model to Improve List-Informational Search, Master’s thesis, National Cheng Kung University, Tainan, Taiwan, R.O.C., 2013.
[10] Hu M. and Liu B., Mining and summarizing customer reviews, KDD’04, 2004
[11] Hu M. and Liu B., Mining Opinion Features in Customer Reviews, AAAI, 2004
[12] Ku L.W. and Chen H., Mining Opinions from the Web: Beyond Relevance Retrieval. Journal of American Society for Information Science and Technology, Special Issue on Mining Web Resources for Enhancing Information Retrieval, 2007
[13] Liu B., Hu M., and Cheng J., Opinion Observer: Analyzing and Comparing Opinions on the Web, WWW, 2005
[14] Liu B., Guang Q., Bu J. and Chen C., Expanding Domain Sentiment Lexicon through Double Propagation, IJCAI-09, 2009
[15] Liu, B., Sentiment analysis and opinion mining. Synthesis Lectures on Human Language Technologies, 5(1), 1-167, 2012
[16]Lin C.Y., Cao Y., Duan H., Yu Y. and Hon H.W., Recommending Questions Using the MDL-based Tree Cut Model, WWW, 2008
[17] Lin C.Y., Duan H., Cao Y. and Yu Y., Searching Questions by Identifying Question Topic and Question Focus, In Proceedings of ACL-08, 2008
[18] Liu Q., Agichtein E., Dror G., Maarek Y., and Szpektor I. When web search fails, searchers become askers: Understanding the transition. SIGIR, 2012.
[19] Lin S.J. and Lu W.H., Learning Question Focus and Semantically Related Features from Web Search Results for Chinese Question Classification, In AIRS, 2006
[20] Popescu A.M. and Etzioni O., Extracting Product Features and Opinions from Reviews, EMNLP’05, 2005
[21] Pang B. and Lee L., A Sentimental Education Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts, ACL, 2004
[22] Pang B. and Lee L., Seeing stars Exploiting class relationships for sentiment categorization with respect to rating scales ACL, 2005
[23] Richard C.W., Schlaefer N., William W. C., and Automatic E.N., Set Expansion for List Question Answering. EMNLP, 2008
[24] Rose D.E. and Levinson D., Understanding User Goals in Web Search, WWW, 2004
[25] Su Q., Xu X., Guo H., Guo Z., Wu, X., Zhang, X., and Su, Z., Hidden sentiment association in chinese web opinion mining, WWW, 2008
[26] Turney P., Thumbs Up or Thumbs Down? Semantic Orientation Applied to Unsupervised Classification of Reviews, ACL’02, 2002
[27] Wang S.M., Using Natural Language Question Structure Features to Improve List-Information Search, Master’s thesis, National Cheng Kung University, Tainan, Taiwan, R.O.C., 2012.
[28] Xue X., Jeon J., and W. B. Croft. Retrieval models for question and answer archives. SIGIR Conference, 2008.
[29] Yang, H., Yang, B., Zhou, X., Zhou, C., Wang, Z., and Chai, Z., Community discovery and sentiment mining for Chinese BLOG. IEEE, 2010