A Quantitative Analysis of Discourse Phenomena in Machine Translation

Abstract

State-of-the-art Machine Translation (MT) systems translate documents by considering isolated sentences, disregarding information beyond sentence level. As a result, machine-translated documents often contain problems related to discourse coherence and cohesion. Recently, some initiatives in the evaluation and quality estimation of MT outputs have attempted to detect discourse problems in order to assess the quality of these machine translations. However, a quantitative analysis of discourse phenomena in MT outputs is still needed in order to better understand the phenomena and identify possible solutions or ways to improve evaluation. This paper aims to answer the following questions: What is the impact of discourse phenomena on MT quality? Can we capture and measure quantitatively any issues related to discourse in MT outputs? In order to answer these questions, we present a quantitative analysis of several discourse phenomena and correlate the resulting figures with scores from automatic translation quality evaluation metrics. We show that figures related to discourse phenomena present a higher correlation with quality scores than the baseline counts widely used for quality estimation of MT.