05 / 分享主题:Neuralizing Symbolic and Statistical Approaches to NLP
分享嘉宾:上海科技大学信息科学与技术学院长聘副教授 屠可伟时间:20:55--21:55分享环节:50分钟主题分享+10分钟互动问答分享摘要:While deep learning and neural approaches become dominant in the field of NLP over the past five years, we argue that traditional symbolic and statistical approaches still have their merits. In this talk, I will discuss our recent effort of integrating traditional symbolic and statistical techniques with modern neural approaches. I will first introduce a novel type of recurrent neural networks that can be converted from regular expressions and deployed in zero-shot and cold-start scenarios. I will then introduce the technique of unfolding statistical inference algorithms as recurrent neural networks and discuss its application to dependency parsing and CRF decoding. Finally, if time permits, I will briefly discuss our work on vectorizing formal grammars.