RACL: A robust adaptive contrastive learning method for conversational satisfaction prediction
RACL: A robust adaptive contrastive learning method for conversational satisfaction prediction
复制标题
DOI：
10.1016/j.patcog.2023.109386
复制
发表时间：
2023-02
期刊：
影响因子：
--
通讯作者：
Gang Chen;Xiangge Li;S. Xiao;Chenghong Zhang;Xianghua Lu
中科院分区：
文献类型：
--
作者：
Gang Chen;Xiangge Li;S. Xiao;Chenghong Zhang;Xianghua Lu
近年来，会话系统的商业化取得了巨大的进展，例如智能应答系统[1]、会话助手[2]、人工智能聊天机器人[3]以及会话营销系统（以基于直播和基于微信的销售为例）[4,5]。商业会话代理通过与用户进行人工或自动的实时对话，挖掘用户的需求，有效捕捉用户的购买意图。另一方面，用户代理会话的质量反过来影响用户的会话满意度，用户的会话满意度在决定用户参与意愿和购买意愿的强度方面起着决定性的作用。从这个意义上说，准确预测会话满意度会带来多方面的好处。对于零售商而言，会话满意度预测（CSP）可以为促进产品销售和服务质量提供明确而有价值的线索，从长远来看，这必将提升其竞争优势。对于会话代理来说，CSP有助于建立响应更快、更有效的会话系统[7]。对于会话代理管理器，CSP使他们能够有效地监视代理的会话性能和质量。综上所述，CSP对于会话商务中的研究人员、开发人员和管理人员非常重要和感兴趣。CSP的目标基本上可以归结为两个任务，即用户满意度预测和会话满意度分析，这两个任务都受到了现有研究的极大关注。本研究中CSP的目的是揭示用户在特定的会话营销环境中对产品或服务的特定愿望或目标的实现。影响用户满意度形成的因素包括用户的偏好、期望及其固有特征（如个性和情感信号）[8,9]。这些功能的推导依赖于个人数据，而这些数据通常是公众无法获得的。例如，在基于微信和基于呼叫代理的销售服务场景中，用户和会话代理甚至在会话结束时还不了解对方，这给CSP带来了相当大的挑战。
In recent years, the commercialization of conversation systems has gained dramatic progress, such as intelligent answer systems [1], conversational assistants [2], artificial intelligence (AI) chatbots [3], and conversational marketing systems (exemplified by live-streaming-based and WeChat-based sales)[4, 5]. By conversing with users in real time artificially or automatically, the commercial conversation agents are expected to excavate users’ demands and capture their purchase intentions effectively. On the other hand, the quality of the user-agent conversation in turn influences the user's conversational satisfaction, which plays a decisive role in determining the intensity of the user's willingness to engage and purchase. In this sense, predicting conversational satisfaction precisely brings multifaceted benefits. For retailers, conversational satisfaction prediction (CSP) can provide explicit and valuable cues to promote product sales and service quality [6], which, in the long run, is bound to boost their competitive advantages. For conversation agents, CSP is conductive to the establishment of more responsive and effective conversation systems [7]. For conversation agent managers, CSP empowers them with ability to monitor the conversational performance and quality of the agents efficiently. For all above, CSP has been of great importance and interest to researchers, developers, and managers in conversational commerce.The objective of CSP can be essentially boiled down to two tasks, ie, user satisfaction prediction and conversational satisfaction analysis, each of which has drawn considerable attention from extant research. The aim of CSP in this study is to reveal the fulfillment of a specified desire or goal of users toward a product or service within a certain conversational marketing context. Factors contributing to the formation of user satisfaction include users’ preferences, expectations, and their inherent characteristics (eg, personality and emotion signals)[8, 9]. The derivation of these features relies on personal data, which is generally unavailable to the public. For example, in the scenarios of the WeChat-based and the caller-agent-based sales services, users and the conversation agents even do not know each other well by the end of the conversation, which poses considerable challenges to CSP.