在分析語句情緒的研究中，為了提升準確率，通常會加入一些因素規則，比如情緒關鍵字的使用與人工定義的情緒規則；這些自制化的因素，往往會因為需求龐大的數據與漫長的訓練要求，造成系統架構的不靈活性與效能不佳。因此在論文的研究中，將以上述的需求為考量，建立一個能分析文句語意內容，並具有快速特性與一定效能的系統架構。 論文的系統架構分為三大部分，分別為資料訓練：其為情緒及情緒心理學的相關研究，主要根據知網的語料庫 (HowNet) 與中研院中文詞知識庫小組的中文詞類分析技術報告為參考資料生成情緒規則，產生稀疏表示特徵，建立稀疏表示字典，透過解出稀疏係數後，將兩類別各自的字典及係數還原原向量，並與原向量計算誤差，獲得最小誤差者即為所屬類別；再者為議題輸入與評論資料取得描述如何取得時下論壇的熱門討論文章之評論內容；最後為資料分類：可以根據資料訓練之結果分析議題分類的準確度。另外，在研究實驗上，論文將逐一辨識時下的流行論點作為情緒分類模組的實作議題。;In the research of semantic sentiment analysis, it will normally use some factor rules such as the utilization of emotional keywords and the emotional rules defined manually to increase the accuracy. Because of the demand for large amounts of data and the training take lots of time, these manual factors will usually make the construction of system unportable and decrease efficiency. In this thesis, based on the above demands, we propose a semantic sentiment analysis system, and it also have better quality and increase efficiency. The system structure of this thesis is organized as follows. First, the data training: It is the research of emotion and emotion psychology. According to the linguistic definition such as HowNet and CKIP technical report, we could make the emotional rules to generate the sparse representation characteristic, and build the sparse representation dictionary. By solved the sparse coefficient, return the dictionary and coefficient of two categories to original vector respectively. Then calculate the error with original vector, the dependent category which is obtain minimum error. Second, the input topic and the obtainment of comments: It present how to get the comments of the hot topic in the internet forum. Finally, the data classification: we will analyze the accuracy of classified topics by the result of data training. Besides, the experimental results will identify the hot topic as the implementation of semantic classification models.
