Towards humanized open-domain conversational agents

• Main Author: Zhong, Peixiang
• Other Authors: Miao Chun Yan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Document and text processing; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/147791
• Institution: Nanyang Technological University

Language is the hallmark of humanity. Conversation or dialogue is a fundamental arena of language and one of the most commonly used forms by humans. In the field of artificial intelligence (AI) or, more specifically, natural language processing (NLP), a conversational agent (CA), also known as a dialogue system (DS), is an intelligent machine that can converse with humans in natural language. There are primarily three types of CAs: task-oriented CAs, question-answering (QA) systems, and open-domain CAs. In this thesis, we focus on open-domain CAs, also known as chatbots, which are designed to chat with users in any topics engagingly with the aim of establishing long-term relationships. Open-domain CAs are essential in modern conversational user interfaces and have been adopted in numerous business domains such as personal assistant, customer support, education, and healthcare. Building a human-level open-domain CA has been one of the major milestones in AI research. However, existing open-domain CAs often fail to model the intrinsic traits of humans and exhibit the following limitations: 1) they lack emotional intelligence and cannot generate or recognize emotions in conversations, which often lead to dull or generic responses; 2) they lack commonsense knowledge and often produce incoherent or unrelated responses; 3) they lack persona and often produce inconsistent responses; and 4) they lack empathy and often produce non-empathetic responses. Addressing the aforementioned limitations is important for bridging the gap between existing CAs and human-level CAs. These intrinsic traits of humans have been empirically shown to improve the performance of CAs on various tasks, e.g., user satisfaction in customer support, user trust and engagement in education, and mental health of participants in healthcare. Humanization is the process of attributing human traits to an entity. In this thesis, we propose to address the limitations by humanizing open-domain CAs with the following human traits: emotion, commonsense, persona, and empathy. Our thesis makes a step towards humanized open-domain CAs. Specifically, to humanize CAs with emotion and commonsense, we first propose an emotional open-domain CA that can generate natural and emotional responses. We then incorporate commonsense into emotional CAs and propose a conversational emotion recognition model and a commonsense-aware emotional response generation model. Experimental results show that both emotion and commonsense improve response quality and human ratings. In addition, emotion and commonsense are shown to have complementary effects in conversational emotion recognition and generation. To humanize CAs with persona and empathy, we propose a persona-based empathetic CA and investigate the impact of persona and empathy on response quality. Experimental results show that both persona and empathy consistently improve response quality and human ratings. In addition, we investigate the impact of persona on empathetic responding and our results suggest that persona has a larger impact on empathetic conversations than non-empathetic ones. Finally, we propose a humanized open-domain CA (HCA) that possesses all the proposed human traits simultaneously: emotion, commonsense, persona, and empathy. HCA aims to address the aforementioned limitations altogether. Specifically, we adopt a pretrain-and-finetune paradigm to develop a retrieval-based HCA in a multi-task learning setting. Experimental results show that the multi-task performance of HCA is better than its single-task performance, and our HCA outperforms the state-of-the-art CAs for response retrieval across multiple evaluation datasets. Our case study shows that our proposed HCA can demonstrate multiple human traits and produce consistent, informative, and empathetic responses.