基於深度強化式學習情感理解之富同情心的對話生成聊天機器人

Abstract

近年來，由於人工智慧的發展，許多自動化的客戶服務系統逐漸多元化。人們逐漸將生活中所要完成的事務，仰賴於人機互動裝置系統，例如：聊天機器人。然而，當人們溝通時，總會思考對話者的處境以及心理狀態，提供較適當的回覆，使得人與人之間，能夠互相扶持及關懷。但是，現在的電腦系統，互動上卻無法給人們，如與真人交流的真實感受，因為目前的機器人尚未理解何謂同理心，導致機器給予人們的回應都限制於特定任務。本論文提出一能夠理解對話者情緒，並能流暢回應對方、符合談話主題、且具有同情心回應的文字聊天機器人系統。透過增強式學習算法，我們的系統能夠相當程度地模擬人們真實對話的過程，經由對話者的回應，學習如何選擇最適當的回覆。在文字情感辨識中，我們結合不同標籤域，包含離散情緒標籤以及連續性情緒關係，提升整體的辨識效果，並勝過許多目前情感辨識的架構。我們還使用多任務學習準則，結合情感資訊來建立對話系統，使我們的聊天機器人能夠根據對應情感，給予適當的回應。本研究是目前學界中，首先使用模擬人類學習的對話過程，來訓練聊天機器人的系統。透過實驗，我們的系統於定性 (Qualitative)和定量(Quantitative)的表現，都優於現有其他系統。

With the rapid development of technology, many automated customer service, and medical care are gradually increasing. More and more people start using computer interactive services to complete their daily tasks, e.g. chatbots. However, interaction experience between computers and humans is not as good as that between humans, since the empathy revealed between people is not experienced in Human-Machine Interaction (HMI). When people interacts with each other, they often try to imagine the situations of the other and give a more appropriate response so that the interlocutor can get more psychological assistance and feel more supportive by each other. In this paper, we propose a chatting system to care about others’ mental state and to be with the abilities to give fluent, coherent, and empathetic responses to users. By simulating real conversations, our system can obtain the feedback from the speakers, and determining the most appropriate way to respond through reinforcement learning. In addition, we incorporated labels of different domains, such as the discrete and the continuous emotional labels, to improve the overall performance of emotional classification with text data, outperforming several existing approaches on emotional classified benchmark dataset. Moreover, we also adopted multi-tasks learning guidance on conditional sentence generation from which we can generate a fluent and coherent responses as well as a specific emotional response. Through experiments, all our models perform better than the baseline model quantitatively. Besides, with several examples of responses generated by our system chatting with real people, we believe that our system can provide fluent, coherent, and empathetic responses in real life.