Modelling self-awareness in social robot

• Main Author: Zhang, Jiaheng
• Other Authors: Tan Ah Hwee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/138222
• Institution: Nanyang Technological University

Self-Awareness is a crucial feature for a sociable agent or robot to better interact with humans (Subagdja & Tan, 2017). Given the widespread popularity and large dataset of the Harry Potter Book Series, it is an appealing prospect to try to model Harry Potter’s consciousness to a sociable agent. However, there are few studies conducted on information extraction (including sentiments and emotions) from a fictional novel and the next most relevant study was done on Shakespeare plays 7 years ago (Nalisnick & Baird, 2013). This project was able to improve this study, implementing state-of-the-art methods at each stage of sentiment and information extraction. This project’s main contributions include the implementation of Huggingface NeuralCoref for Coreference Resolution, Window method for Character Relationship Strength Analysis, Stanford CoreNLP Open Information Extraction (and post-processing method) for Direction of Sentiment and Scope of Analysis, VADER Sentiment Analysis for Sentiment Analysis, EmoLex Sentiment Analysis for Emotion Analysis. This project also created new approaches in order to increase accuracy and/or coverage of results: Gender-Specific Proximity Algorithm was used as an alternative in Coreference Resolution and Dependency Algorithm was introduced to increase coverage for Direction of Sentiment and Scope of Analysis results. A human study was also conducted to rate the effectiveness of the above methods, in an effort to compare the above methods with human understanding and judgement of the text. Lastly, these sentiments and emotions were tagged to major events present in the book and a Self- Awareness data model was created to represent the relationships between the extracted data. These methods have been demonstrated to successfully extract sentiments and emotions from the text. Finding Direction of Sentiment and Scope of Analysis with OpenIE was rated at 93.3% accuracy and Depedency Algorithm was rated at 90.7%. These two methods were used to extract phrases with a directed sentiment within the text. Then, sentiment analysis with VADER and emotion analysis with EmoLex can be conducted. The sentiment and emotion values obtained from the models achieved acceptable ratings of 3.51/5.00 and 3.47/5.00 respectively in the human study conducted. An important finding was that the Dependency algorithm was able to extract about 99.1% more sentiments and 100% more emotions than OpenIE.