Since the appearance of Siri from Apple, dialogue systems have become more and more prominent in our lives. Recently, there has been an increasing interest in the Natural Language Processing (NLP) community to design adaptive systems. Initial research has shown that stylized and personalized conversations, tailored to users’ needs and preferences, would help strengthen the connections between dialog systems and human users. As a result, personalized content improves user engagement in conversations, increases communication effectiveness, and develops trust in the systems.
For the selection of user-centred content, psychologically motivated concepts such as emotions and personality have been investigated and incorporated into the development of human-like conversational dialogue systems. In contrast to short-lived emotions and affective states, personality traits are more stable and endurable over time. Therefore, personality is better suited to model long-term user preferences while emotions are for short- and mid-term preferences. Injecting human traits into a system should start with understanding real human interactions. However, there is seemingly a lack of insights from other disciplines in popular research literature. The definitions of “emotions” and “personality” are often data-driven, and so are the responses of the systems to the users. In the PERSONA-CHAT dataset, a persona, or personality, is a list of five random characteristics.
There are three key challenges in delivering stylized and personalized content to users by emotion- and personality-aware dialogue systems.
The first one involves the automatic detection of the user’s affective states and personality traits to build their models. Which emotion and personality inventories should be selected? And what could be used as feature cues for the detection (e.g. texts, speech, body language)? Secondly, with the user data from the previous step, the generated responses should be personalized and stylized to user preferences. Furthermore, the personality of the systems should be consistent throughout the conversations. The last challenge is about the ethical aspects of the dialogue systems. For example, given the users’ distressing emotional states, how should the systems respond? More importantly, when should they try to change user behaviours, and when not?
Using established theories from both psychology and linguistics, and latest model architectures from NLP, the project aims to partly address the second and third challenges. In the Style Transfer task, the GAN architecture has been utilised extensively for the conversion of texts from one style to another according to users preferences. This method lies in the assumption that style and content can be separated completely . However, recent work has proven that such clear separation is not easily attainable, if not impossible, depending greatly on the domains. These findings have motivated us to examine other frameworks for a deeper understanding of their own strengths and weaknesses. Taking a data-centric approach to the challenges, we work to develop flexible ML applications in NLP that can deliver these goals.