Auto-Adaptive Learning from Weak Feedback for Interactive Lecture Translation

提出的项目目标是使统计机器翻译（SMT）能够完成大学讲座翻译的艰巨任务。这是通过一个互利的学习周期来增强SMT来实现的，称为自适应SMT，它结合了人类对SMT输出的后期编辑，系统可以从中立即学习。在“传统”的设置中，后期编辑被指示制作完美的翻译，这可能是非常耗费资源的，不仅在编辑时间方面，而且在用户所需的语言熟练程度方面。在这个项目中，我们的主要目标是探索从弱反馈中学习的方法，而不是从完整的后期编辑中学习。这种反馈可能包括部分修正或仅仅是对SMT输出质量的判断。一个中心点是，为了保证机器的可学习性，人类的反馈需要包含一个足够强的信号来进行统计学习。这意味着我们面临着机器可学习性和人类用户反馈的可获得性之间的权衡，我们将尝试解决这个问题。我们的研究将侧重于设计从弱反馈中进行学习的高效算法，以及支持在大学讲座交互式翻译的现场测试中实际使用算法的前端/后端接口。

The goal of the proposed project is to enable the use statistical machine translation (SMT) for the difficult task of translation of university lectures. This is done by enhancing SMT by a mutually beneficial learning cycle, called auto-adaptive SMT, that incorporates the human for post-editing SMT output from which the system can learn immediately. In the ``traditional'' setup, post-editors are instructed to produce a perfect translation, which can be very resource-intensive, not only in terms of editing time but also in terms of a user's required language proficiency. In this project, it is our main goal to explore ways of learning from weaker feedback than a full post-edit. This feedback could consist of partial corrections or merely judgments on the quality of the SMT output. A central point is that in order to guarantee machine learnability, human feedback needs to contain a signal strong enough for statistical learning. This means that we are faced with a trade-off between machine learnability and elicitability of feedback from human users, which we will attempt to solve. Our research will focus on the design of efficient algorithms that perform learning from weak feedback, and on frontend/backend interfaces that support a practical use of the algorithms in field tests of interactive translation of university lectures.