March 3, 2017 @ 12:00 pm – 1:15 pm
Hackerman Hall B17
Neural Machine Translation (NMT) has gained popularity in recent several years thanks to its simple, yet effective, architecture. Despite the success of achieving state-of-the-art performance for various language pairs, there are several inherent weaknesses for NMT: repeating and dropping translations, rare words coverage, lower alignment accuracy, and slow training convergency. In this talk, I’ll show some work to alleviate those issues by integrating the merits of SMT approach. I’ll cover the following work: large vocabulary NMT, coverage embedding models, and supervised attentions. If time permits, I’ll also show some results of layer normalization, fertility models etc.
Dr. Haitao Mi recently joined the AI department of Ant Financial US as a staff engineer. His main research interests include Natural Language Processing, Machine Learning, and Deep Learning Algorithms. Prior to joining Ant Financial US, he worked as a Research Staff Member at IBM Watson Research Center. His main goals at IBM were developing state-of-the-art Statistical Machine Translation, Neural Machine Translation, Parsing, and Question Answering systems. He was a tech. lead at NMT group, took on responsibility of improving NMT training and decoding performance for various language pairs.
Dr. Mi received his Ph.D. degree from Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) in 2009. He has published more than thirty papers in top-tier NLP conferences. He also served as an area co-chair of the North American Chapter of the Association for Computational Linguistics 2015 (NAACL 2015), and the Association for Computational Linguistics 2017 (ACL 2017).