運用多任務方法進行對話失焦與脫身意圖偵測

Abstract

對話系統已成為企業不可或缺的工具，然而維持使用者與系統之間高品質的互動仍是一個重大挑戰。雖然現有的對話系統提升了回應速度和效率，但其無法偵測和應對使用者是否參與在其中可能會對系統的效用產生負面的影響，是影響整體使用者體驗和滿意度的重要因素，倘若能在對話進行的過程中及時偵測到使用者的對話開始失焦並意圖脫離，就能使對話系統快速地作出相應的策略進行調整，改善整體對話的體驗。因此在本研究中，我們的目標是提出一個有效的對話失焦與使用者脫身意圖偵測模型，並利用多任務學習方法，將對話行為分類和情感識別作為輔助任務試圖提升模型在主任務上的表現，這種多任務框架不僅提高了脫離偵測的準確性，還透過共享相關任務的學習結構，提供了對使用者在對話中行為更多的理解。在實驗過程中，我們提出一個特殊的兩階段訓練策略：初始階段專注於輔助任務，隨後整合主任務進行第二階段的訓練。實驗結果顯示，我們提出的多任務學習模型加上兩階段訓練策略的表現，在偵測使用者產生對話失焦和意圖脫身的行為上表現出色，這一方法證明了專注於輔助任務在提升主任務能力方面的有效性。除此之外，我們創建了一個中文對話資料集，模擬真實世界場景，確保研究的實用性，該資料集包括明確的定義與標註指南，為未來持續擴增和對話系統研究提供有用的資源。總之，本論文通過新穎的多任務學習方法推進了對話失焦與使用者脫身意圖偵測的領域，提供了一個能夠滿足現實場景中用戶多樣化需求的強大解決方案。我們的研究為未來發展奠定了堅實基礎，有助於創建更具回應性和適應性的對話系統。

Dialogue systems have become indispensable tools for businesses; however, maintaining high-quality interactions between users and these systems remains a significant challenge. While current dialogue systems have improved response speed and efficiency, their inability to detect and respond to user engagement can negatively impact their effectiveness, influencing overall user experience and satisfaction. Timely detection of users' disengagement intention during a conversation can enable dialogue systems to quickly adjust dialogue strategies, thereby enhancing the overall experience in systems.This research aims to develop an effective model for detecting user disengagement in dialogues, employing a multi-task learning approach, MAD-detect. By incorporating dialogue act classification and sentiment recognition as auxiliary tasks, we seek to improve the model's performance on the main task, disengagement detection. This multi-task framework not only enhances the performance of disengagement detection but also provides a deeper understanding of user behavior by sharing learned structures across related tasks. We propose a unique two-stage training strategy: the initial phase focuses on auxiliary tasks, followed by the integration of the primary task in the second phase. Experimental results demonstrate that MAD-detect combined with the two-stage training strategy outperforms baseline models, achieving higher recall and balanced.This approach proves the effectiveness of emphasizing auxiliary tasks to enhance disengagement detection capabilities. Furthermore, we created a comprehensive Chinese dialogue dataset that simulates real-world scenarios to ensure the practical applicability of our research. This dataset includes clear definitions and annotation guidelines, making it a valuable resource for future expansions and dialogue system research.In summary, this research advances the field of dialogue disengagement detection through an innovative multi-task learning approach, offering a robust solution that meets diverse user needs in real-world scenarios. Our research lays a solid foundation for future developments, contributing to the creation of more responsive and adaptable dialogue systems.