| Author: | Wen, Zhiyuan |
| Title: | Comprehending and reflecting personality in dialog systems |
| Advisors: | Cao, Jiannong (COMP) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Speech processing systems Human-computer interaction Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xix, 149 pages : color illustrations |
| Language: | English |
| Abstract: | The evolution of dialog systems has revolutionized human-computer interactions. Starting with the emergence of the first chatbot, Eliza, in the 1960s, dialog systems have forged ahead from merely generating predetermined responses to offering increasingly human-like experiences. Despite these advancements, most dialog systems are trained on extensive dialog data from diverse speakers, which restricts their capacity to learn specific conversation patterns of individuals. Consequently, they still struggle with limitations like delivering generic responses regardless of user characteristics, displaying unstable and non-interpretable emotional expressions, and generating content in inconsistent language styles. In psychological literature, personality is understood as a constellation of relatively enduring individual traits and characteristics that delineate a person’s unique cognitive, affective, and behavioral patterns. In the context of the dialog system, personality encapsulates user attributes, encompasses the tendency of emotional expressions, and also affects word usage and language tones. Therefore, in this thesis, we focus on comprehending and reflecting personality in dialog systems, which requires understanding the personality manifested in dialog context, incorporating personality traits into dialog systems, and ensuring the responses consistently reflect the specified personality trait. We first identify the key challenges and difficulties within our problem. Then, we propose a research framework that outlines the necessary modules required to tackle the challenges and achieve our target. Finally, we propose different approaches to address these challenges, which are summarized below. Our first challenge is comprehending personality in insufficient data. Personalities describe long-term patterns of behaviors that are difficult to understand in short- term conversation data. Besides, dialog content with precise personality annotations is rare due to privacy concerns and the professional nature of personality analysis. Faced with these concerns, we explored extracting information from other aspects for personality recognition. We first propose Affective Dialog Encoder (ADE), the first model to leverage affective information and emotional interactions in dialog flows to enhance personality recognition in conversation. Besides, we also investigate integrating psycho-linguistic knowledge to fine-tune pre-trained language models with only tens of annotations for personality recognition. Our second challenge is integrating psychological findings into neural network (NN)- based dialog system models. Psychological findings provide theoretical evidence for integrating personality into dialog systems. However, these findings are often from special experiments or questionnaires on small groups; they may be unsuitable for NN-based models trained on the massive general corpus. To tackle this issue, we raise a new research task, Personality-affected Emotion Generation, to explore enabling dialog systems to automatically select appropriate response emotions affected by psychology findings. Specifically, we model the emotion generation process as the mood state transitions affected by specified personality in the Valence-Arousal-Dominance (VAD) space. Then, the linear coefficients of personality effects are modeled as trainable model parameters supervised by large-scale dialog data. Besides, we also construct the Personality Emotion Line Dataset (PELD), an emotional dialog dataset of 6.5k dialogues with personality annotations for speakers to facilitate related research. The third challenge issue is to reflect personality consistently across various dialog contexts. In different conversation scenarios, the dialogue system will generate various response content. Identifying and effectively controlling factors influenced by personality across different dialog contexts is difficult. Faced with this issue, we identify the text style words facilitated with linguistic lexicons and investigate incorporating lexical modification with semantic generation for effective language style modification. Based on this principle, we propose the Decode with Template model to validate our idea on the scenario of text sentiment transfer. We believe this thesis takes a solid step towards creating more humanized conversational agents by incorporating personality traits in dialog systems. Plenty of conversational services can benefit from it, such as empathetic companions for the elderly, entertaining social chatbots, and AI-based mental therapy. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12801