Personalization in Dialog System

Overview

One of the fundamental challenges in artificial intelligence (AI) is endowing machines with the ability to converse with humans using natural language. Personalization in dialog systems gives persistent personalities to dialog systems so that they can produce more personal, specific, consistent, and engaging responses than persona-free models. Personality can include factual attributes such as age, gender, and workplace, or psychological personality traits, such as the Big Five traits: openness, conscientiousness, extraversion, agreeableness, and neuroticism. This project targets coherent response generation and explores the personalization of psychological traits. Persona information can be categorized into factual attributes and psychological personality traits. The second layer covers general integration approaches in dialog system design. The two mainstream methods are (1) persona-guided generation, which incorporates persona representations or key-value pairs into dialog system design, and (2) generation and modification, which first outputs a general response trained on a large-scale dialog corpus and then modifies it with desired personal information. The third layer of the framework covers personalized response representation methodologies. Since the ultimate target of personalization is to maintain consistency in response generation, the two main aspects needed to achieve this target are factual consistency maintenance and consistent personalized style rendering.

Project Framework

Achievements

• Wen Z, Cao J, Yang R, et al. Decode with template: Content preserving sentiment transfer[C]//Proceedings of The 12th Language Resources and Evaluation Conference. 2020: 4671-4679
• Zhiyuan Wen, Jiannong Cao, Ruosong Yang, Shuaiqi Liu, and Jiaxing Shen: Automatically Select Emotion for Response via Personality-affected Emotion Transition. Accepted by Findings of ACL-IJCNLP 2021

Members

Zhiyuan Wen