神經網路已經被成功地應用於解決各種分類及函數近似的問題，尤其因為神經網路是個萬能的近似器(universal approximator)，所以對於函數近似的問題效果更為顯著。以往對於此類問題雖然多數以線性的分析工具為主，但是實際上多數問題本質上是非線性的，所以對於非線性分析工具的需求其實是很大的。自1986年起，神經網路本身的運作一直被視為一個黑箱作業，難以判斷網路學習結果的合理性，更無法有效地幫助使用者增進其知識，因此提供一套合理及有效的神經網路分析方法是重要。 本文提出一套分析神網路系統的方法；利用線性規劃的技巧萃取及分析網路中的規則(rule)，而不需要對任何資料集做分析；進而利用統計無母數方法－符號檢定－歸納出網路中的知識。以債券評價為例，驗證此方法的可行性，實證結果亦顯示此方法所萃取出來的規則是合理的，且由這些萃取出的規則中，所歸納出來有關債券評價的知識多數是合理的。 / Neural networks have been successfully applied to solve a variety of application problems including classification and function approximation. They are especially useful for function approximation problems because they have been shown to be uni-versal approximators. In the past, for function approximation problems, they were mainly analyzed via tools of linear analyses. However, most of the function approxi-mation problems needed tools of nonlinear analyses in fact. Thus, there is the much demand for tools of nonlinear analyses. Since 1986, the neural network is considered a black box. It is hard to determine if the learning result of a neural network is rea-sonable, and the network can not effectively help users to develop the domain knowl-edge. Thus, it is important to supply a reasonable and effective analytic method of the neural network. Here, we propose an analytic method of the neural network. It can extract rules from the neural network and analyze them via the Linear Programming and does not depend on any data analysis. Then we can generalize domain knowledge from these rules via the sign test, a statistical non-parameter method. We take the bond-pricing as an instance to examine the feasibility of our proposed method. The result shows that these extracted rules are reasonable by our method and that these generalized domain knowledge from these rules is also reasonable.

近年來，設計已成為提高產品附加價值並增進利潤的利器之一，企業在全球競爭壓力下為了提升競爭力，積極透過設計力開發創新產品。在政府的積極推動下，許多傳統產業與設計公司媒合。然而如何將產品創新需求，轉換並傳達成設計概念，成為極其重要且困難的問題。 本研究為有效傳達設計概念，蒐集2005年至2012年參加德國iF國際產品設計大獎以及RedDot設計獎得獎作品，鎖定所有桌椅櫃類的產品描述，應用文字探勘技術將產品描述過濾並找出對應特徵值亦即設計元素，再利用KNN技術將設計元素分群，試圖從各群中萃取出設計概念。 本研究將260篇桌椅櫃類產品設計文件中分成16群設計概念。分群係以群內平均相似度大於0.05做為門檻以形成設計概念。 本研究結果分為16群設計概念，分別命名為「特色零件多樣感覺概念」、「傳統與現代木椅概念」、「以系統為主的豪華家具」、「波型的時尚概念」、「多樣設計感沙發」、「多造型十字腳椅」、「仿生化人體工學概念」、「親子概念」、「舒適躺臥概念」、「具設計感的室內外用椅」、「注重靠背設計概念」、「多角度對稱概念」、「各式形狀桌面與沙發概念」、「殼形靠背椅」、「中國傳統」、「強調地點取向的概念」等概念，需求者可透過需求之設計元素對應出相關設計概念群與設計者進行有效溝通，更快的了解所想要設計之產品，設計師可以大大縮短在需求階段所消耗的時間以及力氣。最後本研究亦提出一些未來研究方向。 關鍵字：文字探勘、kNN、設計概念、萃取

對於神經網路系統將提出一個法則萃取的方式，並從神經網路中得到相關法則。在這裡我們所提到的方法是根據反函數的觀念而得到的。 / A rule-extraction method of the layered feed-forward neural networks is proposed here for identifying the rules suggested in the network. The method that we propose for the trained layered feed-forward neural network is based on the inversion of the functions computed by each layer of the network. The new rule-extraction method back-propagates regions from the output layer back to the input layer, and we hope that the method can be used further to deal with the predicament of ANN being a black box.

評論是所有人類行為的核心，因為它影響我們行為的關鍵因素。我們都試著從不同型式的評論分析與研究試著從作者字裡行間的文字呈現內容深入推敲及理解，從而要能過濾出能協助決策的有用資訊。在早期的評論研究將評論視為是文本分類問題，直到2000年前後，從分析評論的主觀句子與評論裡形容詞的程度衡量用詞，學者們開始對解構整篇文本的內容，並試著從語言學的角度分析用字遣詞與情感方向之間的關聯。這種從文字語義關聯分析評論的方式，也使文本挖掘技術必需結合自然語言的處理原則，才能更準確地了解評論的內容。隨著許多新興的機器學習演算法與自然語言處理方法不斷地推陳出新，及網路使用行為拓展至電子商務與線上虛擬社群的建立，情感分析研究亦開始不斷地蓬勃發展。 漢文不同於世界其它語言，它擁有許多獨特表徵：無空格區隔、一字一語素、依詞為語言中表達意義的最小獨立單位，也使得在套用源自西方的情感分析原則時更加困難。然而過去的研究者則加以利用這些語言特徵，建立出專屬中文的情感分析原則。我們務實地討論適用於中文情感分析的情境(a)可取得情感分析資源及專家語言智慧，及(b)可取得領域字詞特徵向量定義的兩個前題下，提出適合的中文情感分析策略。在情境(a)中，我們深入討論運用部首資訊至情感分析中的適用性，並且提出一套能精萃出領域評論文本的觀測字詞/部首組的方法。研究中我們萃取出50個部首組，並運用在領域相近的評論裡得到很好的情感分類成效。而在情境(b)中我們提出適合深層類神經網路學習方法的評論字詞的權重過濾原則，不僅能確保評論字詞在學習過程中仍保有能積旋出合適屬性，並且驗證此權重原則在支援向量機的學習方式下亦有相同的優勢。在研究中，我們亦討論此兩種情境下進行情感分析的必要條件與資訊，並為未來更深入的中文情感分析起到墊腳石的作用。 / Opinion is the core of human behaviors, because it directly influences key factor of our behaviors. Despite of personal or organizational decision making processes, we all constantly conduct various kinds of opinion analysis, including explaining and comprehending what users present. At the beginning, opinion studies considered as a text mining problems, and tried to cluster opinions into positive and negative groups. After 2000, researchers intended to decompose sentences from whole opinions by analysing subjective expressing and adjective words presenting within, as well as explained the relationships between semantics and sentiment from linguistics aspect. Therefore, opinion analysis has to incorporate with natural language processing techniques, so we can understand the opinion contents. Nowadays, sentiment analysis grows event booming due to emerging machine learning and natural language processing approaches, as well as the needs of electronic commerce and virtual community on line. Unfortunately, Chinese is quite unlike other language due to non-space separated, one character as one morpheme, and considering words (compositing with several characters) as minimum semantic expression unit. And those language features also bring difficult to adopted sentiment analysis principles from English. Nevertheless, researchers leveraged Chinese language information to propose specific sentiment analysis approaches dedicated to analyze Chinese opinions. In this study, we practically discussed the situations of conducting sentiment analysis: (a) using sentiment analysis resources and experts’ knowledge; and (b) using word feature vector, called word2vec, and deep learning. In (a) scenario, we propose a Chinese radical-based sentiment analysis approach and experiment the applicability. We also proposed a feature extraction method, so we can generate 50 seeds for further analysis. In (b), we compared 4 different feature selection approaches for deep learning, in order to keep accuracy and make sure understandable feature can be generated in neural network. We also tested feature selection approaches in SVM classifier and retrieved similar results. In this study, we also discussed essential constraints and required information in both scenarios, as well as the results of this study can be the foundation of continuing Chinese sentiment analysis studies.

本研究發展可以在任意網頁上進行閱讀標註之合作式閱讀標註系統，並透過探勘集體智慧方式，在合作式閱讀標註系統上發展「優質標註萃取」及「達人標註萃取」機制，來輔助學習者進行數位文本閱讀學習，以達到提昇閱讀理解成效的目的。此外，本研究也進一步探討透過「優質標註萃取」及「達人標註萃取」機制過濾掉一部份品質較差的標註，是否可有效降低閱讀標註文本時產生的認知負荷。 本研究將學習者分成實驗組1（達人標註）、實驗組2（優質標註）與控制組（所有標註）三組，並分別進行約80分鐘的合作式閱讀標註學習活動。其中控制組的成員採用「呈現所有標註之合作式閱讀標系統」支援閱讀學習；而實驗組1的成員則透過「呈現達人標註之合作式閱讀標註系統」來進行閱讀學習；實驗組2則透過「呈現優質標註之合作式閱讀標註系統」來進行閱讀學習。合作式閱讀標註活動要求學習者在指定時間內閱讀本研究指定的文本（化學科普之文章），同時利用「合作式閱讀標註系統」進行閱讀標註撰寫與分享。閱讀標註活動結束後，學習者將進行所閱讀文本之閱讀理解評量以及認知負荷量表填寫，據此瞭解學習者的閱讀理解成效及認知負荷程度。 研究結果顯示，採用具有「優質標註萃取」機制所得標註支援閱讀學習，有助於過濾品質不佳的閱讀標註，並提供更簡潔易找尋之優質標註支援閱讀學習，進而提昇閱讀理解成效，由於閱讀時更容易找到所需的優質資訊，因此亦較有助於提昇學習者不同面向概念的閱讀理解成效；此外，本研究基於每位學習者的有效標註，在考量標註層次及標註數量下，評估每位學習者的“標註能力”，採用優質標註支援閱讀學習的實驗組2（優質標註）學習者中，標註能力越高的學習者，其閱讀理解成效也較佳；而本研究將學習者依照閱讀理解後測成績高低，分成高分組及低分組後顯示，控制組（所有標註）與實驗組2（優質標註）的組別中，均呈現出低分組學習者的認知負荷顯著高於高分組學習者的現象；除此之外，本研究比較三組採用不同標註呈現方式之合作式閱讀標註系統進行閱讀學習之學習者時，結果發現，採用三種不同閱讀標註呈現方式組別學習者之認知負荷無顯著差異。 最後，本研究歸納研究者在研究過程及結果中之發現，提出發展結合合作式閱讀標註的有效閱讀學習策略、探討各類型標註眼動行為對於閱讀理解成效影響與擴展合作式閱讀標註系統支援行動閱讀學習等未來研究議題之初步架構，供後續研究參考以進行更深入之探究。 / A Collaborative Reading Annotation System, which can be randomly proceeded reading annotations on any web pages, is developed in this study. Furthermore, Quality Annotation Extraction and Master Annotation Extraction are developed on the Collaborative Reading Annotation System by mining collective intelligence for assisting learners in proceeding reading digital texts and promoting the reading comprehension performance. The effect of removing some bad-quality annotations through Quality Annotation Extraction and Master Annotation Extraction on reducing the cognitive load when reading annotation texts is further discussed in this study. The learners are divided into Experiment Group 1 (Master Annotation), Experiment Group 2 (Quality Annotation), and Control Group (All Annotation) for 80-minute collaborative reading annotation learning. Control Group uses Collaborative Reading Annotation System with all annotations for promoting reading; Experiment Group 1 proceeds reading through Collaborative Reading Annotation System with master annotations; and, Experiment Group 2 applies Collaborative Reading Annotation System with quality annotations to reading. The learners are requested to read the assigned texts (articles of popular science in chemistry) in the assigned period and write and share the reading annotations with the Collaborative Reading Annotation System. Afterwards, the learners are evaluated the reading comprehension of the texts and fill in the cognitive load scale for understanding the reading comprehension performance and the cognitive load. The research results show that utilizing the annotations acquired by Quality Annotation Extraction for promoting reading could filter out unfavorable reading annotations and provide quality annotations, which are more easily searched for promoting reading, to further enhance the reading comprehension performance. Since the quality information can be more easily searched, it could better assist learners in promoting reading comprehension performance in various aspects. Moreover, based on the valid annotations of each learner, the annotation ability is evaluated the annotation level and quantity. Learners with higher annotation ability in Experiment Group 2 (Quality Annotation) present better reading comprehension performance. Based on the reading comprehension post-test results, the learners are divided into high-score and low-score groups. The cognitive load of low-score learners in both Control Group (All Annotation) and Experiment Group 2 (Quality Annotation) is higher than it of high-score learners. Besides, the cognitive load among the three groups applying the Collaborative Reading Annotation System with different annotations to reading does not appear significant differences. Finally, developing effective reading strategies with Collaborative Reading Annotation, discussing the effects of various annotations on reading comprehension performance, and expanding Collaborative Reading Annotation System for promoting mobile reading are proposed as the preliminary framework for future research, with which in-depth exploration could be preceded in successive research.

本論文應用策略性貿易理論的觀點，分別探討三個獨立的研究主題：各國關稅結構中普遍存在的「關稅調升」現象、外銷比例政策與外籍廠商技術選擇的關係、及關稅與配額政策對於外籍廠商技術授權決策的影響。 壹、 關稅調升與連續性壟斷 「關稅調升」為世界上大部分國家關稅結構中普遍存在的現象，惟這種現象的理論探討卻十分匱乏。因此，本章的目的即在補充「關稅調升」成因的理論探討。藉由連續性壟斷產業模型的建立，我們的研究顯示：政府若對下游進口產品課徵關稅，此一關稅除了具備新貿易政策理論所稱之「利潤萃取」效果之外，尚具備萃取外國生產上游產品廠商部份獨佔利潤的功能，我們稱此為下游產品關稅的「垂直」效果。在連續性壟斷產業的架構之下，隨著生產階段的遞增，下游產品關稅能夠萃取這些上游獨佔利潤的外國廠商家數亦將增加，「關稅調升」現象因而產生。因此，本章發現關稅之「利潤萃取」效果與「垂直」效果的聯合作用是造成關稅調升現象的重要因素。 貳、 外銷比例政策與技術選擇 在實務上，外銷比例政策常為開發中國家對於多國籍廠商在其國內設廠營運時的一項管制措施。在台灣的發展經驗中，外銷比例政策亦常被政府的財經官員認為具有移轉外國優越技術的有效政策工具，其理由乃在於藉由對多國籍廠商內銷比例的管制，誘使其提升在台的生產技術以面對高度競爭性的國際市場，並同時讓本國廠商透過技術擴散或技術移轉的方式獲得多國籍廠商的技術水準。 本章的目的在於探討外銷比例政策是否真能達到提升外籍廠商技術水準的效果。本章的研究結果發現：在外籍廠商獨占本國市場的情況下，除非政府允諾給於外籍廠商高度比例的內銷市場，否則外銷比例政策非但不會促使外籍廠商選擇較為優良的技術，反而會導致其採取較劣等的技術。另一方面，當本國市場有本國廠商參與競爭時，外銷比例對於外籍廠商技術水準的選擇除了受到前述比例值高低之影響外，也受到市場策略性競爭效果的影響。當本國市場的需求函數為線形時，市場競爭的策略性效果會使得外籍廠商在面對外銷比例的管制增加時會選擇較差之技術。因此，一般而言，外銷比例政策並無法確保外籍廠商會使用較為先進之技術水準。 參、 外籍廠商技術授權：關稅與配額政策的比較 貿易保護政策的實施有可能改變廠商海外市場的營運選擇，譬如改採以技術授權的方式間接進入海外市場，因此地主國的貿易保護政策可以促使該國廠商獲取外籍廠商先進技術的授權。 本章的研究乃在於提供貿易保護政策與國際間授權技術選擇關係的理論分析。藉由比較不同的貿易政策對於對於多國籍廠商市場進入方式與授權技術選擇的影響，本章發現對應於一特定之關稅稅率，等量配額政策在市場需求曲線為凹性（凸性）的情況下，將比關稅政策更易於（更不易於）誘使外國廠商授權先進的技術給本國廠商；而當市場需求曲線為線性的情況下，關稅政策與等量配額政策對於外籍廠商授權技術水準的影響是完全相同的。然而，若本國採取的是等率配額政策，則不論其對應之關稅稅率為何，外籍廠商在等率配額限制之下一定會授權給本國廠商最先進之技術。 第一章 緒論 1 第二章 關稅調升與連續性壟斷 6 第一節 本章前言 6 第二節 基本模型 11 （1） 最終財貨關稅 13 （2） 原物料關稅 20 第三節 關稅調升現象 23 第四節 n層次加工產業下的關稅結構 32 第五節 本章結語 35 附錄 37 第三章 外銷比例政策與技術選擇 39 第一節 本章前言 39 第二節 外籍廠商獨佔下的技術選擇 42 第三節 寡占下的外籍廠商技術選擇 48 第四節 本章結語 54 附 錄 56 第四章 外籍廠商技術授權：關稅與配額政策的比較 58 第一節 本章前言 58 第二節 基本模型 61 第三節 外籍廠商在關稅政策下的授權技術選擇 62 第四節 外籍廠商在等量配額政策下的授權技術選擇 70 第五節 外籍廠商在等率配額政策下的授權技術選擇 78 第六節 範例說明 82 第七節 本章結語 86 第五章 結論 87 參考文獻 89

本研究在合作式數位閱讀環境中發展了一套「知識標註學習系統」，可以支援多人同時針對一篇數位文本進行閱讀標註與互動討論，以提升讀者閱讀的深度與廣度。此外，本研究更進一步地以專家評估法設計「知識萃取機制」，用於判斷讀者閱讀標註的重要度。 「知識萃取機制」是基於讀者閱讀標註中所蘊含的閱讀理解策略與閱讀技巧，以及合作式閱讀社群中產生的標註共識，考量了「標註範圍長度」、「標註範圍詞性」、「標註範圍位置」、「標註策略類型」、「標註範圍共識」與「標註喜愛共識」等六項因素，以專家評估法制定的標註重要度模糊隸屬函數來評定各因素的重要度並量化為「標註因素分數」指標，最後將六項因素以模糊綜合評判進行推論，再將推論結果解模糊化而成為代表標註重要度的量化指標「標註分數」。基於「知識萃取機制」所計算代表標註重要度的「標註分數」，可作為讀者進行閱讀標註是否不佳的判斷，並據此提供標註技巧建議與優質標註內容推薦的「標註建議」，以幫助讀者提昇閱讀理解能力。 為了驗證「知識萃取機制」計算「標註分數」的有效性，以及探討未來改善「知識萃取機制」和可加入的考量因素與適性化設計的可能方向，本研究以單組後測設計規劃實驗，並以國立政治大學圖書資訊數位碩士在職專班19位學生作為實驗對象，進行一份數位學習論文的合作式閱讀標註學習，並於實驗後評估實驗對象閱讀文章之後的閱讀理解能力，作為評鑑「知識萃取機制」計算方式是否有效的指標。最後再以問卷蒐集實驗對象對於「知識萃取機制」的意見，歸納成為未來研究改善的參考依據。 研究結果發現，本研究所提出「知識萃取機制」中計算標註重要度的「標註分數」與實驗對象的閱讀理解能力呈現低度正相關，一定程度地證實了「知識萃取機制」計算方式的有效性。而「知識萃取機制」六項考量因素中，「標註範圍長度」與「標註喜愛共識」為分辨實驗對象閱讀理解能力的關鍵因素；「標註策略類型」與「標註範圍詞性」的標註重要度模糊隸屬函數有待修正；「標註範圍共識」與「標註範圍位置」為無效因素，但這可能是受到計算方式錯誤與閱讀文章類型的影響，未來仍有待進一步評估。在未來發展方面，系統操作標註行為頻率越高，實驗對象的閱讀理解能力也有較高的跡象，未來可以將其納入「知識萃取機制」作為考量因素之一；而閱讀理解能力較差的實驗對象，呈現出比較不願意回應「標註建議」與較常使用社群互動的現象。本研究歸納可能原因為實驗對象自身的閱讀素養不成熟，以至於無法判斷「標註建議」的正確性，而需要參考他人閱讀標註。 未來研究可針對本研究的實驗對象與閱讀標註資料進行更深入的分析，並且將改良後的「知識萃取機制」擴大至探討其他類型的數位文本閱讀標註與實驗對象。也可以搭配認知策略教學法建構閱讀教學鷹架，或是將「知識標註學習系統」用於支援數位典藏與數位圖書館閱讀學習，以激發更多不同領域的應用研究。 / Based on the concept of cooperative reading learning, the study presented a cooperative reading annotation system termed as "Knowledge-based Annotation Learning System (KALS)", which can support cooperative reading annotation while reading a common text-based digital material, to accumulate reading knowledge and to promote readers’ reading comprehension abilities. Through KALS, readers could freely increase annotation for any text words on a text-based digital material with HTML format. Readers can also share and discuss the contributed annotation with other readers via interaction interface in KALS. Furthermore, this study also developed an intelligent Knowledge Extraction Mechanism (KEM), which can mine the quality annotation knowledge and annotation skills based on a large amount of readers’ annotation archived on KALS, to further promote reading comprehension of readers via on-line recommending high quality annotation knowledge and good annotation skills to readers. KEM employed fuzzy synthetic decision approach to quantify each reader’s annotation as a numeric index termed as "Annotation Score" under simultaneously considering two annotation consensuses including anchor consensus and favorite consensus, and four annotation features including anchor length, part of speech of anchor word, anchor location and annotation strategy. In a manner, "Annotation Score" can represent the importance of reader's annotation. Thus, KEN uses "Annotation Score" to determine which annotation needs the suggestion of annotation skill tips, and which high-quality annotation can be recommended to readers. At the same time, readers are encouraged to reflect their annotation behavior based on the suggestion of annotation skill tips and high-quality annotation recommended by KEN, and are asked to respond the feedback from KEM. To evaluate the effectiveness of the proposed KALS with KEM, the study designed an experiment to collect readers' annotation behavior after readers read an assigned text-based digital material, and then assessed readers’ reading comprehension ability. Reading comprehension ability was used to verify the effectiveness of "Annotation Score" inferred by KEM and to explore the potential factors that can improve KEM. In the designed experiment, participants were 19 graduate students of E-learning Master Program of Library and Information Studies of National Chengchi University who took the course of Integrating Information Technology into Teaching. All participants were asked to read an academic paper related E-Learning issue based on the support of KALS with KEM during two weeks. Moreover, they had to finish a reading report and accept a test of reading comprehension after finishing reading learning activity. The report and test were served as the measurement of participants' reading comprehension. The experimental results show that there is a low positive correlation between "Annotation Score" and participants' reading comprehension score, thus confirming the effectiveness of the proposed KEM. Furthermore, KEM could be improved by adjusting the annotation importance calculation approach of part of speech anchor word and annotation strategy. This study also confirmed that the considered factors of KEM should eliminate two factors including anchor consensus and anchor location. Additionally, future study should consider adopting frequency of annotation behavior as considered factors of KEM. Moreover, the experimental results also show that participants with low level of reading comprehension ability have higher need of community interaction than participants with high level of reading comprehension ability while using KALS for reading learning, and they are difficult to confirm whether the recommending tips of annotation from KEM is correct or not. Obviously, exploring the difference of participants’ annotation behavior between different levels of reading comprehension abilities provides benefits to develop adaptive functionalities of KEM in the future.