Machine Translation for Human Translators

While machine translation is sometimes sufficient for conveying information across language barriers, many scenarios still require precise human-quality translation that MT is currently unable to deliver. Governments and international organizations such as the United Nations require accurate translations of content dealing with complex geopolitical issues. Community-driven projects such as Wikipedia rely on volunteer translators to bring accurate information to diverse language communities. As the amount of data requiring translation has continued to increase, the idea of using machine translation to improve the speed of human translation has gained interest. In the frequently employed practice of post-editing, a machine translation system outputs an initial translation and a human translator edits it for correctness, ideally saving time over translating from scratch. While general improvements in MT quality have led to productivity gains with this technique, there has been little work on designing translation systems specifically for post-editing. In this work, we propose improvements to key components of statistical machine translation systems aimed at directly reducing the amount of work required from human translators. We propose casting MT for post-editing as an online learning task where new training instances are created as humans edit system output, introducing an online translation model that immediately learns from post-editor feedback. We propose an extended translation feature set that allows this model to learn from multiple translation contexts over time as data sources become more reliable. We propose an automatic evaluation metric that scores hypothesis-reference pairs according to several statistics that are directly interpretable as measuring of postediting effort. Our metric can be used to optimize translation systems in scenarios where standard metrics break down, select optimal system configurations for post-editing, and provide insight into the properties of translation quality that are most important for minimizing editing effort. Our online translation models and evaluation metrics are compatible with standard decoders and optimization algorithms. To evaluate the impact of our post-editing-targeted translation system, we propose a series of experiments that use a web-based framework to collect several types of highly accurate data from human translators. We discuss MT for post-editing as a distinct task and present the results of initial post-editing experiments. We finally outline an experimental setup for collecting valuable data that will be used to evaluate the impact of our online translation models and optimization metrics on human editing requirements.