Visual Evaluation of Translation Alignment Data

Abstract

Translation alignment plays a crucial role in various applications in natural language processing and digital humanities. With the recent advance in neural machine translation and contextualized language models, numerous studies have emerged on this topic, and several models and tools have been proposed. The performance of the proposed models has been always tested on standard benchmark data sets of different language pairs according to quantitative metrics such as Alignment Error Rate (AER) and F1. However, a detailed explanation on what alignment features contribute to these scores is missing. In order to allow analyzing the performance of alignment models, we present a visual analytics framework that aids researchers and developers in visualizing the output of their alignment models. We propose different visualization approaches that support assessing their own model's performance against alignment gold standards or in comparison to the performance of other models.