以憂鬱症病患為對象之多標籤對話生成

Abstract

在現今社會中，儘管憂鬱症病患或是有憂鬱傾向的人越來越多，但憂鬱症病患的求助率卻十分低迷，主要是由於憂鬱症病患缺乏一個良好且即時的聊天管道，加上許多應用在醫療領域的對話系統都著重在解決特定任務上，無法解決憂鬱症病患缺乏聊天管道的難題，而此時能夠理解使用者情緒，並且幫助使用者紓解負面情緒或是給予正面積極的回覆的對話系統就顯得極為重要。 因此本論文與本研究室另一位成員Huang(2022)共同設計一個對話系統，能夠分辨使用者輸入文字的態度、意圖與主題，並且透過這些標籤來進行對話生成，希望藉此來控制模型產生出正面積極且相關的回覆給使用者，來與使用者聊天或是給予建議。 本論文使用強大的GPT-2預訓練模型(pre-trained model)並結合多種標籤來訓練模型，分別對英文對話資料集DailyDialog與由本論文與Huang(2022)所共同建構的中文對答資料集ChinesePsyQA進行實驗，將多種維度標籤加入到對話文本前，藉此讓模型產生出相對應的回覆，而最後在人工評估與自動評估兩種評估方式上，本論文所提出的方法皆比其他生成模型有著更好的表現，說明將多種標籤加入到對話中，可以有效控制模型的生成文本。

In today's society, although there are more and more people with depression or depression tendencies, the help-seeking rate of depression patients is still very low, mainly due to the lack of a good and real-time chat channel for depression patients. In addition, many dialogue systems applied in the medical field focus on solving specific tasks, and cannot solve the problem of lack of chat channels for patients with depression. At this time, a dialogue system that can understand the user's emotions and help users relieve negative emotions or give positive responses is extremely important. Therefore, this paper and another member of our laboratory, Huang (2022), jointly designed a dialogue system that can distinguish the attitude, intentions and topics of the user's input text, and use these tags to generate dialogues, hoping to control the model to produce positive and relevant replies to users, to chat with users or give advice. This paper uses the powerful GPT-2 pre-trained model and combines various labels to train the model. Experiments were carried out on the English dialogue dataset DailyDialog and the Chinese QA dataset ChinesePsyQA jointly constructed by this paper and Huang (2022). Various dimension labels were added before the dialogue text, so that the model can generate corresponding responses. Finally, in the human evaluation and automatic evaluation, the method proposed in this paper has better performance than other generative models, indicating that adding multiple tags to the dialogue can effectively control the generated text of the model.