Semantic Structure from Paraphrase Pairs

Anne Cocos
Ph.D. Student in Computer Science
University of Pennsylvania
SERC 306
Wednesday, October 3, 2018 - 11:00
In order for a computer to answer a question like “What’s a Chinese dish that’s not so hot,” it needs to incorporate multiple types of semantic knowledge. Specifically, it must understand sense distinctions -- Chinese dish here refers to food, not pottery; hypernymy -- sweet & sour pork IS-A Chinese dish; and adjective intensity -- mild is less intense than hot. Previously, these three types of semantic knowledge have been acquired manually, or via automatic methods based on distributional similarity or lexico-syntactic patterns in text corpora. In this talk, I will propose ways that we can derive these three types of semantic structure from a different source: a large, noisy collection of paraphrase pairs. In each case, I will discuss how the paraphrase-based information can complement existing methods for automatically acquiring semantic information.