Predicting and Using Implicit Discourse Elements in Chinese-English Translation

In machine translation (MT) implicitation can occur when elements such as discourse markers and pronouns are not expected or mandatory in the source language, but need to be realised in the target language for a coherent translation. These ‘implicit’ elements can be seen as both a barrier to MT and an important source of information. However, identifying where such elements are needed and producing them are non-trivial tasks. In this paper we examine the effect of implicit elements on MT and propose methods to identify and make them explicit. As a starting point, we use human translated and aligned data to decide where to insert place holders for these elements. We then fully automate this process by devising a prediction model to decide if and where implicit elements should occur and be made explicit. Our experiments compare statistical machine translation models built with and without these explicitation processes. Models built on data marked for discourse elements show substantial improvements over the baseline.