Evaluating automatic sentence alignment approaches on English-Slovak sentences

Abstract

Parallel texts represent a very valuable resource in many applications of natural language processing. The fundamental step in creating parallel corpus is the alignment. Sentence alignment is the issue of finding correspondence between source sentences and their equivalent translations in the target text. A number of automatic sentence alignment approaches were proposed including neural networks, which can be divided into length-based, lexicon-based, and translation-based. In our study, we used five different aligners, namely Bilingual sentence aligner (BSA), Hunalign, Bleualign, Vecalign, and Bertalign. We evaluated both, the performance of the Bertalign in terms of accuracy against the up to now employed aligners as well as among each other in the language pair English-Sovak. We created our custom corpus consisting of texts collected in 2021 and 2022. Vecalign and Bertalign performed statistically significantly best and BSA the worst. Hunalign and Bleualign achieved the same performance in terms of F1 score. However, Bleualign achieved the most diverse results in terms of performance.