Can You be More Polite and Positive? Infusing Social Language into Task-Oriented Conversational Agents

Goal-oriented conversational agents are becoming ubiquitous in daily life for tasks ranging from personal assistants to customer support systems. For these systems to engage users and achieve their goals in a more natural manner, they need to not just provide informative replies and guide users through the problems but also to socialize with users. To this end, we extend the line of style transfer research on developing generative deep learning models to control for a specific style such as sentiment and personality. This is especially useful and relevant to dialogue generation of conversational agents. In this paper, we first apply statistical modeling techniques to understand human-human conversations. We report that social language used by humans is related to user engagement and task completion. After that, we propose a conversational agent model which is capable of injecting social language into agent responses given user messages as input while still maintaining content. This model is based on a state of the art end-to-end dialogue model using a sequence to sequence deep learning architecture, extended with sentiment and politeness features. We evaluate the model in terms of content preservation and social language level using both human judgment and automatic linguistic measures. The results show that the model can generate social responses that enable agents to address users’ issues in a more socially conscious way. 1 Background and Introduction Conversational agents are becoming part of our lives. These systems generally fall into two categories, task-oriented assistants and chatbots [31]. Task-oriented assistants are designed to fulfill a specific task by having single-turn or multi-turn conversations with users to retrieve information from them and complete that task (e.g., Microsoft Cortana, Apple Siri, Google Assistant). Chatbots are designed to have chit chat with users, and the goal is usually to mimic human-human conversations and engage users in those conversations for as long as possible (e.g., ELIZA[35] and XiaoIce). In order to have human-like and extended conversations, some researchers have studied how to incorporate social language into chatbots to generate proper interpersonal responses and build an emotional connection with users [31]. For example, XiaoIce can respond with empathetic language and show caring while chatting with users. However, there are only a handful of studies that focus on incorporating social capabilities into task-oriented assistants [1, 6, 3, 34] even though prior literature has suggested that these factors might play an important role in the process of task-oriented conversations and be associated with user engagement and satisfaction [14, 21, 2, 3]. Thus, we propose this work to answer the following research questions: (1) Can and how do social language used by humans in task-oriented conversations affect user responsiveness and task completion? (2) Can we inject a certain type of social language into the responses of a task-oriented conversational agent? 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada. We focus on the customer service domain and the task of the driver partner on-boarding support in a ride-sharing provider since customer service is a typical application area of task-oriented assistants. Moreover, the driver partner on-boarding support is a closed-domain problem and has a well-defined task. We first conduct an empirical study to quantitatively examine the relationship of social language usage to driver partner responsiveness and the completion of their first trip, based on a dataset of driver partner and human agent conversations. After that, we apply the findings to build an end-to-end deep learning model to generate agent responses given driver partner inquiries. Our aim is to train a task-oriented agent that can produce dialogues with the desired level of social language while still maintaining the necessary content to guide driver partners through the on-boarding funnel and lead them to complete their first trip. The main contributions of this work are shown below: 1. Systematically analyzed the relationship between social language and user responsiveness as well as task completion using a real-world conversation dataset. 2. Proposed a deep learning framework for task-oriented dialogue generation which includes a social language understanding component. 3. Evaluated the model for both content preservation and social language generation. We believe that our findings and approach are also applicable to other deep learning based conversational applications including personal assistants and even chit-chat systems.