本研究主軸為開發任務導向(Task-Oriented)對話系統基礎平台及命名實體識別模型之改良與研發,其目標為使非人工智慧研究領域的使用者能夠直接使用對話系統平台建立對話系統以及克服需要使用者自行填寫實體標籤的問題。底層架構包括自然語言前處理、意圖分類、對話狀態追蹤等子模組,使對話系統能夠精準回應最佳的句子,而本研究將針對「對話系統平台建置及改善語意理解命名實體識別模型」進行探討。 本研究以任務導向型的對話系統作為對話系統的媒介,輔助每位使用者能夠建立各式各樣的任務導向對話機器人,不侷限在特定的對話情境,使用者可建立符合特定場景及特定目的服務的對話機器人且基於樣板式回應範本建置任務導向對話系統。然而面對該平台存在著需要使用者自行填寫實體標籤的問題,本研究利用知識濃縮的語言模型DistilBERT訓練中文命名實體模型,並將該模型整合至對話系統使使用者減少需要耗費大量時間建置實體標籤語料,而該模型訓練的資料集包含Weibo NER、OntoNotes 4.0、MSRA及Resume等資料集,整合先進的遷移式學習技術,來訓練各種資料的預訓練模型並融合了Early Stopping技巧來找到最佳的epoch參數。研究結果顯示命名實體辨識模型在精準度、召回率、F1分數各方面指標接近了中文命名實體辨識最頂尖的模型,但是考量對話系統實際應用的情境,本研究所提出的優化模型兼顧了準確度與速度,實驗結果顯示在效能方面F1分數未經過遷移式學習F1分數為65.42%而經過遷移式學習後F1分數則達到93.49%的表現,而在速度方面DistilBERT模型預測速度快Glyce+BERT模型3倍,而相比Lattice LSTM模型則快17倍,明顯表示出本研究模型反應回饋速度即佳,因此本研究使用的DistilBERT模型在效能與速度方面更適合應用在實際的場景。
This study aims to apply and improve the Named Entity Recognition (NER) model with better performance on speed and accuracy in a Task-oriented Dialogue System platform. This system will enable the general users to build a dialogue system directly and only manually label fewer entities for their corpus. The proposed system not only can assist users to build a variety of task-oriented chatbots that are not limited to a specific situation but also lower the manpower and expertise as building the dialogue system. In general, the dialogue system includes submodules such as natural language pre-processing, NER, intention classification, and conversation status tracking. The submodules enable the dialogue system to accurately provide the best response. In particular, this study will focus on improving the NER model for better semantic understanding. In this study, we mainly use the DistilBERT to train the Chinese NER model and integrate it into the dialog system to reduce the users' time-consuming for the entity labeling. In addition, the advanced transfer learning technology is used to train the NER models by various datasets such as OntoNotes 4.0, MSRA, and Resume. In addition, the early stopping technique is used to find the best epochs for training the NER models. The experimental results show that the F1 score of DistilBERT without transfer learning is 65.42%, while the F1 score of DistilBERT after transfer learning is 93.49%. The results show that the proposed NER model using DistilBERT is close to the best Chinese NER models (Glyce+BERT and Lattice LSTM) in terms of accuracy, recall rate, and F1 score. In terms of speed, the DistilBERT model is 3 times faster than the Glyce+BERT model and 17 times faster than the Lattice LSTM model. The experimental results indicate that the DistilBERT model is faster than the Glyce+BERT and Lattice LSTM models. Considering the application in the real world, the proposed NER model using DistilBERT takes into account both accuracy and speed. Therefore, DistilBERT is more suitable to apply in the applications.