在自然語言領域中，對話機器人應用日益發展迅速，其中需要克服問題之一在於自然語言理解，知道使用者在詢問何種問題及判斷文字間隱藏的資訊，對於使機器了解使用者的問題意圖是相當重要。後續的應用例如對話管理以及如何產生相應的答案皆會需要根據意圖理解來做延伸，因此如何達到更好的辨識率將是一大挑戰。 本研究主要針對對話資料訓練深度學習模型並預測對話行為，運用各種神經網路來解決此困難，並比較其之間的差異，同時引入角色資訊於模型中，找出能針對中文短句子特性能夠有效處理的模型，另外，在模型加入預訓練詞向量，能夠更有效處理中文未知詞，以減少錯誤辨識的可能。最終，本論文比較多種深度學習模型並引入角色資訊來辨識對話行為，相比一般的神經網路模型在電信領域對話資料集提升將近1.2%。 ;In the field of natural language processing, the application of dialogue robot is growing rapidly. One of the problems that need to be overcome in the field is natural language understanding. Knowing what kind of question the user is asking and judging the hidden information between the words, and the intention of making the machine understand the problem of the user is very important. Also the follow-up parts such as dialog management and how to produce the corresponding answer will need to be interpreted according to intent to do, so how to catch a better recognition rate will be a big challenge. In this study, we mainly train the deep learning model for dialogue data and predict the dialogue act. We use various neural networks to solve this problem and compare the differences. At the same time, we introduce the role information in the model to adapt the property of short text in Chinese sentence. In addition, adding pre-training word emebdding to the model can deal with unknown Chinese words more effectively, and this could reduce the possibility of misidentification. In the end, this thesis compares many kinds of deep learning models and introduces role information to identify dialog act, which is nearly 1.2% higher than the typical neural network model in the telecome domain dialogue dataset.
