Reinforcement Learning for Neural Machine Translation
摘要：Neural machine translation (NMT) has achieved great success in recent years, however, it also faces some challenges in both learning and inference procedures. For example, today’s NMT relies on huge amount of bilingual training data, which might not always be available; the beam search based inference procedure is myopic, which may lose the chance of finding the optimal translation. In this talk, we propose employing reinforcement learning (RL) to tackle these challenges. First, we propose a novel RL framework called dual learning, which leverages the symmetric structure of machine translation to effectively learn from monolingual data. Second, we propose an MCTS-based inference method, which looks ahead and searches along the decoding tree, based on a pre-trained value network. Experimental results show that these two technologies can significantly boost the performance of NMT, and relax the requirements on large-scale bilingual data. At the end of the talk, we will mention a few other important directions to explore regarding NMT, where reinforcement learning can also play a critical role.
简历：刘铁岩博士，微软亚洲研究院首席研究员，IEEE院士、ACM杰出会员、CMU客座教授、诺丁汉大学荣誉教授、中科大博士生导师。他被公认为排序学习领域的代表人物，在网络搜索和计算广告学等方向取得了卓越的学术成果。他近期的研究兴趣包括：深度学习、增强学习、分布式机器学习、符号学习等，发表了百余篇学术论文，被他引万余次。发布了微软分布式机器学习工具包（DMTK）、微软图引擎（Graph Engine）等知名项目开源。担任了包括SIGIR、WWW、KDD、NIPS、AAAI、ACL在内的顶级国际会议的程序委员会主席或领域主席；以及包括ACM TOIS、ACM TWEB、Neurocomputing在内的国际学术期刊副主编。