Please be polite: Towards building a politeness adaptive dialogue system for goal-oriented conversations

Abstract

Politeness embalms interactions, putting the parties in conversation at ease. Even the most well-intended communication can fall through if there is a manifestation of rudeness. Complementarily, even the most angst-prone situation can be communicated with much less hurt if lathered with politeness. In this paper, we propose a novel task named as Politeness Adaptive Dialogue System (PADS) to incorporate politeness feedback of agents’ actions according to user’s mood and demands. As there is no politeness labeled conversational data available, we annotate the recently released MultiDoGO dataset having six domains with appropriate politeness labels. The proposed end-to-end dialogue system comprises of a transformer-based politeness classifier that interacts with a reinforced learning framework with four different politeness-oriented reward algorithms, forcing the agent to adapt to polite actions upon encountering the user’s dissatisfaction in the dialogue. We train and evaluate PADS by building six different user simulators, each corresponding to a domain. Quantitative and qualitative analysis show that our proposed politeness based reward algorithms improve the task success rate and reduce the dialogue length with state-of-the-art results.

Introduction

The design of intelligent assistants interacting with users in natural language ranks high in the agenda of current Natural Language Processing (NLP) research. In recent years, several conversational agents have emerged (e.g., Apple’s Siri, Microsoft’s Cortana, Google’s Allo) due to the phenomenal growth of artificial intelligence. Prior research mostly focused on supervised learning-based approaches [1], [2] for building the dialogue systems. These agents can perform simple tasks, answer factual questions, and sometimes aimlessly chit-chat with the user. However, they still lag far behind in terms of both the variety and complexity of tasks they can perform as human assistants. In particular, they lack the ability to learn from user interactions to improve and adapt to the demands of the user. Lately, Reinforcement Learning (RL) has been investigated in order to maximize user experiences to adapt to different dialogue agents, for chitchat [3], task completion [4] and information access [5].

Reinforcement learning’s ability to treat the dialog planning as a sequential decision problem and the focus on long-term rewards [6] have enhanced the performance of dialogue agents compared to the earlier systems. Progress in technology has shown significant impact in conversational systems providing dialogue agents the capability to behave in a more human-like manner [7], [8], [9], [10]. In every goal-oriented dialogue system, enhancing the user experience and providing customer satisfaction is of utmost importance.

Politeness is a virtue of humans [11], [12], and making a machine behave politely is a difficult task that has not been deeply investigated. Especially in goal-oriented dialogue systems, politeness becomes essential when we request the users for the different details needed to complete their demands. An example of generic and polite interactions between the user and the conversational agent is provided in Fig. 1. From the example, it is evident that the politeness adaptive approach by the dialogue agents is more suited to assist the aggrieved and anguished users and provide better communication with its capability to empathize with the user in a humanly manner as opposed to the non-polite system. Previously, in [13], user’s sentiment was incorporated to make the systems more user-adaptive and effective. Inspired by this, we propose to make a politeness adaptive end-to-end learning framework that can assist humans in their day-to-day chores. To achieve this goal, the agent needs to behave cordially and courteously with the users [8] i.e. a dialogue agent should empathize, appreciate, and apologize to its users as much as possible. To form smooth and engaging conversations, behaving politely with its users is important for every conversational agent [14]. Politeness, in itself, has different facets [15], [16], [17] which are difficult to inculcate in a conversational system. By incorporating politeness as shown in the example, it is evident that there is no loss of information, but the response length increases to induce the politeness quotient in the agent. In real life applications as well, the agent requires to pacify the aggrieved users or persuade the customers for products in a polite manner to ensure healthy relationship. To do this, responses tend to increase in length that is acceptable for e-commerce applications whose ultimate goal is to provide customer satisfaction and increase customer retention.

The contributions and/or key attributes of our current work are fourfold ¹:

• We propose a politeness adaptive dialogue system (PADS) that can interact with the users in a polite manner showcasing empathy. To the best of our knowledge, this is the very first attempt towards this direction;

• Annotate MultiDoGO dataset with the fine-grained politeness information;

• Design four dialogue policies in a RL-cum-supervised setting to ensure politeness in interactions; and

• Thorough empirical evaluation is conducted to show that our proposed framework is capable of achieving better task success rates and shorter dialogue lengths in comparison to the strong baselines.

The remainder of the paper is organized in the following manner. We offer a brief summary of relevant work in Section II. The proposed methodology is explained in Section III. In Section IV, the dataset and experimental setup, as well as the evaluation metrics, are presented. The results, as well as the appropriate analysis, are presented in Section V. Lastly, in Section VI, we discuss future work.