Quality Estimation for Machine Translation with Multi-granularity Interaction

Abstract

Quality estimation (QE) for machine translation is the task of evaluating the translation system quality without reference translations. By using the existing translation quality estimation methods, researchers mostly focus on how to extract better features but ignore the translation oriented interaction. In this paper, we propose a QE model for machine translation that integrates multi-granularity interaction on the word and sentence level. On sthe word level, each word of the target language sentence interacts with each word of the source language sentence and yields the similarity, and the \(L_\infty \) and entropy of the similarity distribution are employed as the word-level interaction score. On the sentence level, the similarity between the source and the target language translation is calculated to indicate the overall translation quality. Finally, we combine the word-level features and the sentence-level features with different weights. We perform thorough experiments with detailed studies and analyses on the English-German dataset in the WMT19 sentence-level QE task, demonstrating the effectiveness of our method.