Lunch at 12:30pm, talk at 1pm, in 148 Fitzpatrick
Title: Comparative Reasoning Argument Generation via Topic Discovery
Abstract: Comparative reasoning is essential in academic writing, where the authors prove the importance of some point by comparing it to the others. A specific scenario is when an article is citing and making a comparative argument on related works to demonstrate its significance. Existing works on automatic citation text generation define it as a text summarization task. However, it is beyond summarization when focusing on the comparative reasoning part, where the model needs to detect not only the similarities (e.g., the same research topic), but also the differences (e.g., the different computational complexity of the algorithms) between the citing and cited papers. To study this problem, we construct a dataset named “HOWEVER”, containing sentences making a comparative argument when research papers cite their related works, and we propose a new task to automatically generate these comparative sentences. We observe that there is a clustered distribution of the topic of the target texts, however, there is a distribution gap between source and target texts’ topics. We believe bridging the gap by simulating comparative reasoning behavior is the key to improve the generation quality.
In this talk, I will present my ongoing work on this problem. I will first introduce the challenges of the problem, then introduce our possible approach to model the comparative reasoning behavior.
Bio: Mengxia Yu is a 2nd year Ph.D. student at the University of Notre Dame advised by Meng Jiang. Her research focuses on Natural Language Understanding, Reasoning and Generation.