Extending Word-Level Quality Estimation for Post-Editing Assistance

TL;DR

This paper introduces extended word alignment and a refined word-level quality estimation task to enhance post-editing efficiency by directly identifying editing operations, utilizing supervised mBERT-based methods.

Contribution

It proposes a novel extended word alignment concept and a refined QE task that improves post-editing assistance by directly pointing out editing operations, using mBERT-based models.

Findings

Feasibility demonstrated on two language pairs.

Extended word alignment improves editing operation detection.

Refined QE outperforms original methods.

Abstract

We define a novel concept called extended word alignment in order to improve post-editing assistance efficiency. Based on extended word alignment, we further propose a novel task called refined word-level QE that outputs refined tags and word-level correspondences. Compared to original word-level QE, the new task is able to directly point out editing operations, thus improves efficiency. To extract extended word alignment, we adopt a supervised method based on mBERT. To solve refined word-level QE, we firstly predict original QE tags by training a regression model for sequence tagging based on mBERT and XLM-R. Then, we refine original word tags with extended word alignment. In addition, we extract source-gap correspondences, meanwhile, obtaining gap tags. Experiments on two language pairs show the feasibility of our method and give us inspirations for further improvement.