以憂鬱症患者為對象之聊天機器人對話分析

Abstract

在這個科技發達的時代，現代人生活在充滿壓力的快步調社會，這也導致憂鬱症等精神疾病的患病者持續增加，而精神疾病的盛行所導致的自殺率提高，也會使社會負擔加重。雖然心理疾病的患病者越來越多，本研究發現許多患者是不願意求助的，甚至沒有病識感，不認為自己已經患病，他們缺乏一個管道可以抒發情緒，並感受到被理解與陪伴。 為了解決上述問題，本研究與研究室夥伴(Su, 2022)共同提出了一個針對憂鬱症患者的聊天機器人架構，我們認為一個完善的對話系統可以解決他們即時的情緒需求，給予他們所需要的理解與陪伴。為了設計一個符合憂鬱症患者需求的對話系統，本研究認為對話中的情緒、意圖、主題、態度是有助於對話理解的，並能達成同理患者心情的對話生成。本研究主要透過這些標籤進行對話的分析，期望能透過對話準確的標籤分類，生成正面積極的回覆，滿足患者的傾訴需求。除了要回應使用者的需求，本研究也期望能喚起民眾的病識感，許多人在對話中會透露出自殺想法，我們期望能在對話中完成憂鬱程度的判斷，先確立患者的憂鬱程度，再進行對話的分析與生成，一旦判斷出對話中帶有自殺想法，更能達到即時通報，減少自殺事件發生的可能。 本研究主要的目標是達到準確的判斷與生成，因此我們使用BERT預訓練模型來進行對話的分析與判斷。本研究總共使用三個自行整理的資料集來進行憂鬱程度的判斷，分別為RSP、GSN和MTD資料集；並使用與(Su, 2022)共同進行標記的ChinesePsyQA來進行對話中的主題、情緒、意圖、態度四個標籤的分類。本研究針對標籤進行相關性的人工評估，結果顯示資料集的標籤是與大眾的理解相符的，而在對話分類的評估中，結果顯示我們選用的模型在兩個分類任務上，與之前相似的任務相比都取得了相當好的成果，最後從與(Su, 2022)合作完成的研究成果來看，本論文中準確的分類，是影響生成結果好壞的主要因素之一，也說明了主題、意圖、態度這三個標籤對於此架構來說是不可或缺的一部分。

Today, in the age of advances in information technology, people live in a fast-paced society which full of stress, which also leads to a continuous increase in the number of patients with mental illnesses such as depression, and the increase in suicide rates caused by the prevalence of mental illnesses will also increase the social burden. Although there are more and more patients with mental illness, this study found that many patients are reluctant to seek help, and do not even have a sense of illness. They don’t think they are sick. They lack a channel to express their emotions and feel understood and accompanied. To solve the above problems, the study and my lab partners (Su, 2022) co-proposed a chatbot architecture for depression patients. We believe that a complete dialogue system can solve their immediate emotional needs and give them the understanding and companionship they need. To design a dialogue system that meets the needs of patients with depression, this study believes that sentiment, dialogue act, topic, and attitude in dialogue are helpful for dialogue comprehension, and can achieve dialogue generation that empathizes with patients' feelings. This study mainly analyzes the dialogue through these tags, hoping to generate positive responses through accurate label classification of the dialogue to meet the needs of patients. In addition to responding to the needs of users, this study also expects to arouse the public's sense of illness. Many people reveal suicidal thoughts in conversations. We expect to be able to analyze the depression level in the dialogue. First, we determine the depression level of the patients, and then analyze and generate the dialogue. If it is analyzed that there are suicidal thoughts in the dialogue, it will be able to achieve immediate notification and avoid something regrettable happening again. The main goal of this study is to achieve accurate classification and generation, so we use the BERT pre-trained model to analyze the dialogue. This study uses three self-labeled datasets to classify the depression level, the RSP, GSN, and MTD datasets; and uses ChinesePsyQA, which is co-labeled with (Su, 2022), to classify the four labels of sentiment, intention, topic, and attitude in the dialogue. This study conducted a manual evaluation of the confidence of the labels, and the results showed that the labels of the dataset were consistent with the public's understanding. In the evaluation of dialogue classification, the results show that our chosen model achieves fairly good results on both classification tasks compared to previous similar tasks. Finally, from the research results completed in cooperation with (Su, 2022), the accurate classification results in this study is one of the main factors affecting the quality of the generated results, and it also shows that these four labels are an integral part of this architecture.