School of Engineering Department of Computer Science and Engineering 125 Commonsense Reasoning with Knowledge Graphs Supervisor: SONG Yangqiu / CSE Student: XU Baixuan / COSC Course: UROP1100, Summer Nowadays, Deep learning is getting more and more popular in our daily life. However, how to make the computer understand the commonsense in human daily language remains a tough question in the Natural Language Processing field. To solve this, many computer scientists proposed different heuristic ideas. As a huge breakthrough in the Natural Language Processing field happened in 2018, the commonsense reasoning field was also significantly influenced by the large pre-trained model brought by Google. This report will mainly illustrate the way of making computers understand commonsense in human language using a large pre-trained language model – Bert by introducing the dataset preparation, question formulation and other related works. Vision Language Model is another field I delved deeper in during this UROP project period. After CLIP was proposed, the alignment of image and text embedding space have been advanced far a huge step and after that, lots of Vision Language work like VQA, Captioning have been done based on CLIP and its variant. In this report, I’ll briefly introduce the CLIP and the relevant works in Vision Language prompt learning. Efficient Queries over Database Supervisor: WONG Raymond Chi Wing / CSE Student: JIANG Han / COSC Course: UROP1100, Spring For a database that contains multi-dimensional data tuples, a user may have his/her unique preference for some data over the others. These preferences are expressed by the users’ utility functions. To determine a user’s favorite tuples in the database, the Strongly Truthful Interactive Regret Minimization method asks questions about a user’s preference on small subsets of the database, and tries to learn his/her utility function as a linear function. We study how to linearize utility functions with specific non-linear forms, in order to adapt these functions into the above method. Specifically, we develop a method which gives all linearizations that meet our purpose. We also present some examples using this method in this report.
RkJQdWJsaXNoZXIy NDk5Njg=