預訓練詞向量模型應用於客服對話系統意圖偵測之研究

Study on Pre-trained Word Vector Model Applied to Intent Detection of Customer Service Dialogue System

近年來對話商務的概念在各大科技巨頭間興起，人機互動方式由圖形化介面轉向對話交互介面的方式。因而自然語言成為人機互動介面的關鍵因子。然而教導機器要如何與人類溝通，以完成一項具體任務是相當有挑戰性的。其中一個需要克服的困難是自然語言理解，包含如何辨識使用者在詢問何種問題及如何取得文字間隱藏的資訊。讓機器了解使用者的問題意圖及資訊是相當重要的。本研究主要是針對去識別化後的中文客服對話資料，利用深度學習模型以達到辨識使用者意圖。為了更有效處理中文未知詞以及減少錯誤辨識，本研究比較不同預訓練詞向量模型與深度學習模型來辨識使用者意圖。相較於使用隨機詞嵌入，使用BERT-WWM-Chinese（BWC）模型的正確率提升近10％。這表示BWC模型產生的向量更能抓住用戶問句字詞間的語意關係。使得語意相近的字詞能產生近似的向量進而提升使用者意圖辨識的準確率。

In recent years, the concept of dialogue business has arisen among major technology giants, and the way of human-computer interaction has changed from a graphical interface to a dialogue interaction interface. Therefore, natural language has become a key factor in the human-computer interaction interface. However, teaching the machine to communicate with humans to accomplish a specific task can be quite challenging. One of the difficulties that needs to overcome is natural language understanding, including how to identify what questions users are asking and how to get information hidden between words. It is important to let the machine know the user's intentions and information. The dataset of this study is collected from the dialogue of customer service materials. User’s intents are recognized by deep learning models. In order to process Chinese unknown words more effectively and reduce false recognition, this study compares different pre-training vector models and deep learning models to understand user’s intents. Compared with the use of random word embedding, the correct rate of using BERT-WWM-Chinese (BWC) model is improved by nearly 10%. It shows that the semantic vector generated by BWC model can better represent the relationship between user’s words. The recognition rate of user’s intent raises because similar vectors can be generated from similar words.