School of Engineering Department of Computer Science and Engineering 118 Commonsense Reasoning with Knowledge Graphs Supervisor: SONG, Yangqiu / CSE Student: DING, Wenxuan / COMP Course: UROP1100, Spring Zero-shot question answering (QA) is a type of machine learning task in which a model is trained to answer questions about a given domain without being explicitly trained on data in this domain, which places high demands on the generalizability of the model. In this project, we propose to leverage the conceptualization methods to augment the training data on ATOMIC, a commonsense knowledge graph and get MixedATOMIC, which achieved SOTA on various commonsense QA tasks in a zero-shot setting. We further analyze the success of our methods by exploring the training dynamics. Overall, our results suggest that augmented knowledge conditioned on conceptualization is a promising avenue for improving the performance and generalizability of language models in the zero-shot setting. Commonsense Reasoning with Knowledge Graphs Supervisor: SONG, Yangqiu / CSE Student: VUONG, Tung Duong / DSCT Course: UROP1100, Summer In the field of Artificial Intelligence, commonsense knowledge consists of sensible knowledge about the everyday world, and is important yet challenging to be incorporated into AI systems. One challenge is that although many commonsense benchmarks has been introduced, there is no formal definition of the term "commonsense", leading to the question about the purity of these commonsense benchmarks. The problem is significant, as it may make the evaluation of commonsense reasoning ability on these commonsense datasets unreliable. In this work, we discuss the difference in nature of commonsense knowledge comparing with other types of knowledge and propose well-defined criteria which aims at distinguishing commonsense knowledge from the others. We then study the ability of LLMs in recognizing this difference in order to automatically classify commonsense knowledge and other types of knowledge. Finally, we conclude with our plan as a future study which evaluates LLMs’ performance on these types of knowledge, in order to examine the reliability of the evaluation on commonsense datasets. Commonsense Reasoning with Knowledge Graphs Supervisor: SONG, Yangqiu / CSE Student: WANG, Yicheng / COSC Course: UROP2100, Fall UROP3100, Spring In this report, we document the advancements in our research on abductive reasoning, a valuable form of reasoning utilized to derive the most plausible explanation for a given set of observations. We outline the problem formulation, present our data sampling from the YAGO310, DBpedia50, and FB15k-237 datasets, explore the efficacy of a neural method along with two graph search approaches, and assess the results using two evaluation metrics. We discuss the complexity of the two graph search methods and conclude that they are not feasible in practice. Additionally, we present an analysis of the rationale, advantages, and disadvantages of the two metrics used in our evaluation.
RkJQdWJsaXNoZXIy NDk5Njg=