現今國家時事已然影響了人民的生活型態，而這些影響則大多來自於民生物資取用的便利性及價錢的漲跌或是政治政策的實施及改變。這些政策事件的新聞發佈後，常常會引發一連串的相關文章或人民評論釋出到各大社群網站與新聞媒體網站，如Facebook、PTT等等。這些文章經過統整後，以情感分析、正負面分析等工具來分析後，再做出各類統計計算得出結論，以利決策者可以得知人民對於某人事物大多是怎樣的傾向，讓決策更能貼近民心，這就是常用到的輿情系統。不過，現今的輿情系統還是會有些許的限制，目前的輿情系統可以用一些自然語言處理方法計算出很多情緒值，像是開心、難過、生氣等等，但就是無法得知新聞文章裡對於文章內所提事件或議題到底是支持還是不支持。
本研究的目的在從新聞對於某事件的支持語句做配對及比較的語句相似度，試著分析此文章是否支持對應事件。研究方法使用孿生神經網路(Siamese Network)將一句短短的事件支持語句和一整篇事件新聞文章匹配在一起，使用長短期記憶神經網路(Long Short-Term Memory)並加入注意力機制(Attention Mechenism)導入孿生神經網路中，以解決兩個長短差距過大的語句差異問題。同時使用幾個不同的距離計算公式做比較，將兩個文本代表特徵做完整的匹配計算，訓練出分析事件支持度模型，以達成網路上所有有關某事件支持度正反面傾向之統計，讓決策者能更清楚的判斷民眾對於某事件的支持與否。
本研究主要以時事新聞事件資料，而這些資料中包含了許多中文專有名詞，所以語詞分割工具的選擇對於本研究的效果具有非常重要的影響。本研究將共計4112篇文章的新聞立場正反面傾向執行了人工標記，訓練詞彙用語資料庫有七萬多篇，驗證文章為300篇近幾個月的新聞時事。

Today, news events affect people’s life style deeply. Series of public opinions woud be released on social media, such as Facebook, PTT, etc., upon political event news being issued. Policy makers can then respond to people’s tendency by emotion analysis on the assembled opinion papers. This is how and what the Public Opinion System functions. However, there is still some limitation; the current public opinion system can provide many emtion analysis by some Natural Language Processing methods, such as happiness, sadness, anger, etc. of an article, but not telling sustain from opposition.
This research aims for providing an overall sustain level of an article on a certain news event and Text similarity by comparing the article with a simple sustain phrase. The contribution of this thesis is to provide a clear sustain level meter by a sustain model for policy makers to make decisions. The model introduces an improved Attention Mechanism with LSTM(Long Short-Term Memory) to a Siamese Network to solve the phrase length differential problem, and also incorporating several different distance formulae to obtain a comprehensive matching on features of any two selected news articles.
This research analyzes news event articles collected online. The articles in the dataset contain large amount of Manderin proper nouns. Selection of word segmentation tools therfore plays an important roll on the sufficiency of this research. All the 4,112 articles collected are manually labelled on their positive/negative tendency, together with more than 70 thousand news articles being used to train the term library. After trained, 300 recent news event articles are tested to validate of the model proposed in the research.

胡元輝, 國立中正大學傳播學系暨電訊傳播研究所教授, 傳播研究與實踐．第 8 卷 第 2 期．頁 43-73．(2018, 07), 造假有效、更正無力？第三方事實查核機制初探
徐美苓, 中華傳播學刊．第二十七期．(2015.06), 影響新聞可信度與新聞素養效能因素之探討
楊登堯, 國立臺灣師範大學資訊工程研究所(2017.07), 利用臉書資訊探討網路新聞的吸引度及極性分析
蔡依霖, 國立臺灣師範大學大眾傳播研究所(2016.07), 以鉅量資料取徑分析facebook候選人網路競選行為及群眾討論行為—2014台北市長選舉個案研究

鴻海工業大數據辦公室：https://www.facebook.com/ibdo.foxconn/
藍星球-蛛思輿情分析平台：https://choose.blueplanet.com.tw/
CKIP-輿情系統：https://ckip.iis.sinica.edu.tw/service/opinion/

英文
[1]Liu Lizhen, (March 2014), A novel feature-based method for sentiment analysis of Chinese product reviews
[2]Harris, Z. S. (1954). Distributional Structure. Word, 10(2/3), 146–162
[3]Tomas Mikolov & Kai Chen & Greg Corrado & Jeffrey Dean, Efficient Estimation of Word Representations in Vector Space
[4]Quoc Le & Tomas Mikolov, (May, 2014), Distributed Representations of Sentences and Documents
[5]Deza, Elena & Deza, Michel Marie.( Springer. 2009) Encyclopedia of Distances.
[6]Williams, Ronald J. & Hinton, Geoffrey E. & Rumelhart, David E.( October 1986) Learning representations by back-propagating errors. Nature.
[7]Hochreiter, Sepp & Schmidhuber, Jürgen. (1997,11,01)Long Short-Term Memory. Neural Computation.
[8]Junyoung Chung & Caglar Gulcehre & KyungHyun Cho & Yoshua Bengio.(Dec, 2014) Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling
[9]Manning, C. D. & Schütze, H. (1999). Foundations of statistical natural language processing. Cambridge, MA: MIT Press.
[10]Jeffrey C. Reynar (1998). "Topic Segmentation: Algorithms and Applications"
[11]Richard Durbin & Sean R. Eddy & Anders Krogh & Graeme Mitchison. (1999)Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cambridge University Press.
[12]jieba 0.42.1 - Project description：https://pypi.org/project/jieba/
[13]pyhanlp 0.1.66 - Project description：https://pypi.org/project/pyhanlp/
[14]ckip 0.1.2 - Project description：https://pypi.org/project/ckip/
[15]J. Bromley, et al, NIPS 1993 “Signature Verfification usign a Siamese Time Delay Neural Network“
[16]Po-Sen Huang & Xiaodong He & Jianfeng Gao & Li Deng & Alex Acero & Larry Heck, Learning Deep Structured Semantic Models for Web Search using Clickthrough Data
[17]Jonas Mueller & Aditya Thyagarajan(2016), Siamese Recurrent Architectures for Learning Sentence Similarity
[18]Paul Neculoiu & Maarten Versteegh & Mihai Rotaru, Textkernel B.V. Amsterdam(2016), Learning Text Similarity with Siamese Recurrent Networks
[19]Ilya Sutskever & Oriol Vinyals & Quoc V. Le, Sequence to Sequence Learning with Neural Networks
[20]Dzmitry Bahdanau & KyungHyun Cho & Yoshua Bengio, (2015) NEURAL MACHINE TRANSLATION BY JOINTLY LEARNING TO ALIGN AND TRANSLATE
[21]Jiasen Lu & Jianwei Yang & Dhruv Batra & Devi Parikh.(Jan,2017) Hierarchical Question-Image Co-Attention for Visual Question Answering
[22]Ashish Vaswani & Noam Shazeer & Niki Parmar & Jakob Uszkoreit & Llion Jones & Aidan N. Gomez & Łukasz Kaiser & Illia Polosukhin.(Dec,2017), Attention Is All You Need
[23]Caiming Xiong & Victor Zhong & Richard Socher.(Mar,2018), DYNAMIC COATTENTION NETWORKS FOR QUESTION ANSWERING
[24]Jacob Devlin & Ming-Wei Chang & Kenton Lee & Kristina Toutanova.(May,2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
[25]Nils Reimers & Iryna Gurevych, (Aug,2019), Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks