請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/70805
標題: | 基於深度強化式學習之富同情心的檢索式聊天機器人 Empathetic and Retrieval-based Chatbot using Deep Reinforcement Learning |
作者: | Jiun-Hao Jhan 詹鈞皓 |
指導教授: | 鄭士康(Shyh-Kang Jeng) |
關鍵字: | 聊天機器人,同情心,自然語言處理,文本情境嵌入,深度強化式學習, Chatbot,Empathy,Natural Language Processing,Contextualized embedding,Deep reinforcement learning, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 聊天機器人是一支透過語音或文本訊息來與使用者進行對話的程式。在建立聊天機器人的程序中,對話的連貫性至關重要,而回覆的文法句構也必須重視。但是,隨著對話機器人技術日漸成熟,人們與聊天機器人對話時,越來越重視個人的情緒與感受;他們希望聊天機器人能夠了解使用者心中潛在的感受,期望聊天機器人能更關心他們或是具有同情心。因此,使用者若是收到帶有同情心的回覆,他們會有更好更完全的聊天體驗。舉例來說,以一個學生未通過考試的陳述當作對話開頭,學生說:「我搞砸了我昨天的考試。」在此情況下,「又來?你應該通過考試的。」和「別擔心,你下次會通過考試的。」都是合理且通順的回覆。然而,很明顯地,第一句回覆像是譴責使用者考試的表現,會使使用者的情緒更為低落;而第二句回覆卻像是安慰使用者,所以第二句回覆對於使用者心境來說,較為恰當。本文所提出的系統架構為檢索式聊天機器人。為使對話內容順暢且連貫,我們在訓練模型中運用 BERT 等預訓模型,提升機器人的文字理解能力。為使聊天機器人產生之回覆帶有同情心,我們使用強化式學習,預測回覆必須具備何種情緒,以最大化同情心分數的正向變化值。實驗結果表明,我們提出之檢索式聊天機器人,效能上皆勝過過去其他學者提出的基準模型。由實驗結果之對話紀錄顯示,本文發想之對話機器人,可以在與使用者的聊天過程中,產生順暢,符合邏輯,具有同情心的會話。 A chatbot is a program to conduct a conversation via auditory or textual information. Although the coherence of response plays an important role in chatting with a chatbot, people might place more emphasis on their feeling during interacting with chatbots, which includes but not only caring for people or. By considering the feelings of others, people can have a better interactive and supportive experience. For instance, given “I failed the exam yesterday” as the speaker’s opening sentence, both “Again? You should pass it!” and “Don’t worry. You will pass the exam next time” are relevant and acceptable responses. Considering the emotional state, however, the latter response would make the speaker feel better, whereas the former response seems to blame the speaker, making the speaker feel worse than before. In this study, we present several empathetic chatbots that understand any implied feelings of users and reply to them by arousing empathy. These chatbots are based on the retrieval-based architecture, and were in addition finetuned by deep reinforcement learning. To generate a response with fluency and coherence, the chatbot is trained with the contextualized embedding, such as BERT and. To acquire the capability of replying in an empathetic way, the chatbot takes the current emotional state and the next emotional state into account. Thus, experiment results demonstrate that our chabot outperform the baseline chatbots in objective metrics and subjective ratings. According to examples of conversation between those generated by our model and by real human users, we believe that our chatbots are able to chat with users in real-time, fluently, coherently, and empathetically. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/70805 |
DOI: | 10.6342/NTU202004139 |
全文授權: | 有償授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2008202015162000.pdf 目前未授權公開取用 | 3.55 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。