More about HKUST
A Survey of Abstraction Knowledge: Philosophy, Acquisition, Resources, and LLM Reasoning Applications
PhD Qualifying Examination Title: "A Survey of Abstraction Knowledge: Philosophy, Acquisition, Resources, and LLM Reasoning Applications" by Mr. Zhaowei WANG Abstract: Abstraction is about finding common properties among different things and forming a broader concept, like the concept "furniture," subsuming "sofa," and "table." This ability is a key aspect of human cognition, allowing people to manage daily situations effectively by learning from past experiences and generalizing to new circumstances with minimal practice. In the NLP community, integrating abstraction knowledge into NLP systems has long been a classic challenge. In this survey, we review dozens of papers on abstraction and provide a comprehensive overview. We first discuss the cognitive and linguistic background of abstraction, as well as its comparison with similar linguistic concepts. Next, we explore recent efforts to build abstraction knowledge bases and group existing resources into entity-level and event-level categories. For each category, we synthesize representative studies and highlight their strengths and weaknesses. We then examine the application of abstraction knowledge in various areas, including enhancing the conceptual understanding of LLMs, strengthening LLMs' reasoning ability, and improving performance on downstream tasks. Finally, we summarize the progress and remaining challenges in the whole field, reflecting on future directions for abstraction knowledge in NLP. Date: Wednesday, 21 August 2024 Time: 10:00am - 12:00noon Venue: Room 3494 Lifts 25/26 Committee Members: Dr. Yangqiu Song (Supervisor) Dr. Long Chen (Chairperson) Dr. Junxian He Dr. Xiaojuan Ma