Multi-Lingual Sentence Alignment with GPT Models
Xiao Liang, Yen-Min Jasmina Khaw, Soung‐Yue Liew, Tien-Ping Tan, Donghong Qin · 2023
This paper investigates sentence alignment technology, an essential element in Natural Language Processing tasks like machine translation, that pairs corresponding sentences within bilingual documents. Despite the existence of standard bilingual alignment corpora, they often fall short in covering low-resource languages and specific domains, underscoring the need for advanced sentence alignment algorithms. We explore the potential of the GPT model for sentence alignment, focusing on aspects like prompt construction, experiment design, and results evaluation. Utilizing the OPUS corpus, our experimental design covers eight commonly used languages in 28 directions. The study reveals that the GPT-aligner outperforms traditional algorithms by 5-30% in most scenarios, with the performance contingent on language, prompting, the GPT system model, and few-shot demonstration strategies. This work illuminates the significant potential of sentence alignment algorithms in the future development of language models.