• 강사 : 강동엽 교수 (Dept. of CSE, Univ. of Minnesota, USA)
• 일시 : 2021. 11. 4 (목) 17:00~18:30

NLP systems trained on standard machine learning pipelines, i.e., annotation, learning, and evaluation, are limited to causing problems in numerous aspects. For instance, the dataset collected from crowd workers often contains annotation artifacts or repeating patterns. As the systems are deployed to real-world users, they are not controlled, interpreted, or interacted with real users. During this talk, I will discuss recent work done at the Minnesota NLP group, focusing on data- and human-centric aspects of NLP pipelines. For the human-centric aspect, we contrast human perception of linguistic styles with BERT's interpretation, and we further make the BERT system mimic human perception. Then we develop interactive NLP systems that help scholars better read and write academic papers. In the data-centric NLP, we investigate what makes good annotations and assess the potential impact of them as data points for annotations.