IS 400 Colloquium: Liri Fang

The iSchool Colloquium (IS 400) is a venue for presentation and discussion of research and professional activities by faculty, students, staff, and guest speakers. The colloquium is hosted in person and virtually. All are welcome to attend.

Abstract: With the advent of large language models (LLMs), recent works have studied how to use LLMs to accomplish text-attributed graph (TAG) related tasks like node classification and link prediction. 
In this work, we aim to expand the task scope and investigate how LLMs can help TAG’s completion, taking taxonomy expansion as a specific task. As a kind of TAGs, taxonomy is a hierarchical organization of concepts that can serve various applications, and when new concepts emerge, updating the outdated taxonomy by inserting new concepts appropriately attracts much academic and industrial interest. The corresponding task can be referred to as taxonomy expansion. To this end, we first analyze why LLMs (e.g., Llama and DeepSeek) can not directly and solely finish taxonomy expansion tasks with case studies. Second, we propose a simple and effective way of distilling the knowledge of LLMs through LMs and geometric deep learning methods for the hierarchical and semantic taxonomy graphs, such that we (1) avert the potential hallucination, long-context input limit, and not well-instructed output problems of LLMs dealing with entire taxonomy graphs and (2) achieve the LLM-to-LM domain adaptation for a domain-specific taxonomy expansion, via a proposed structure-semantic monotonization learning process and the taxonomy-augmentation by a prompted LLM. Finally, we design extensive taxonomy expansion experiments to show the effectiveness of our method with SOTA algorithms.

Questions? Contact Yingjun Guan.