School of Engineering Department of Computer Science and Engineering 118 Commonsense Reasoning with Knowledge Graphs Supervisor: SONG Yangqiu / CSE Student: KWONG Hyun Fung / COMP Course: UROP 1000, Summer In this research, we delve into the application of commonsense reasoning in solving high-school level physics word questions. We aim to build a knowledge graph-based model and apply commonsense reasoning to tackle physics word problems that require not only factual knowledge such as formulas but also a nuanced understanding of the context. To understand the ability of Large Language Models (LLMs) to answer these questions, we crafted a test with around 40 questions for OpenAI’s GPT 3.5 Turbo to answer. These questions require commonsense to know what formulas or concepts should be applied. Under zero-shot learning, GPT 3.5 Turbo achieved a score of 65.8%. The research is still in progress, and we are using physics simulator to generate a physics word questions dataset. We aim to build a model that could rely on commonsense reasoning and apply suitable formulas to solve these questions. Commonsense Reasoning with Knowledge Graphs Supervisor: SONG Yangqiu / CSE Student: LI Yu Hong Harry / COMP Course: UROP 1000, Summer This report summarizes the work done during the 2023-24 Summer UROP 1000 course at The Hong Kong University of Science and Technology under the supervision of Professor SONG Yangqiu and Ph.D. candidate WANG Zihao. The project focused on “Commonsense Reasoning with Knowledge Graphs.” Key activities included studying and presenting a postgraduate course on Graph Machine Learning, setting up simulation environments using SAPIEN and iGibson 2.0, and reading relevant research papers. The report details the course content presented, covering topics such as machine learning with graphs, node embeddings, and graph neural networks. The experience provided valuable insights into machine learning, commonsense reasoning, and knowledge graphs, with aspirations for further research in these areas. Commonsense Reasoning with Knowledge Graphs Supervisor: SONG Yangqiu / CSE Student: WANG Yicheng / COSC Course: UROP 4100, Fall This report presents a study on recent advances in graph neural network-based approaches for reasoning over knowledge graphs. It provides an overview of two state-of-the-art methods: GNN-QE, which combines symbolic and neural methods for complex query answering, and NodePiece, a parameter-efficient node embedding technique inspired by subword tokenization. It then analyzes the expressivity of pure transformer models for graph representation learning through the TokenGT framework. The report evaluates the strengths and limitations of each approach, such as GNN-QE’s ability to leverage both symbolic interpretability and neural generalizability, and NodePiece’s reduction in parameters. It also examines theoretical results on the expressive power of transformers. Overall, the study aims to provide insights into graph neural network models for efficient and interpretable reasoning over large knowledge graphs.
RkJQdWJsaXNoZXIy NDk5Njg=