Designing AI-powered translation education tools: a framework for parallel sentence generation using SauLTC and LLMs

Area of Science:

• Computational linguistics
• Natural Language Processing
• Translation Studies

Background:

• Translation education (TE) is labor-intensive, with AI offering potential efficiency gains.
• High-quality datasets are crucial for developing AI tools in TE, particularly for Arabic.
• Existing resources like the Saudi Learner Translation Corpus (SauLTC) are not optimized for didactic purposes.

Purpose of the Study:

• To transform the SauLTC into a parallel sentence corpus suitable for AI-driven TE.
• To leverage Generative Pre-trained Transformer (GPT) models for corpus enhancement.
• To evaluate the quality of AI-generated parallel sentences for educational applications.

Main Methods:

• Utilized Generative Pre-trained Transformer (GPT) to process and augment the SauLTC.
• Employed cosine similarity with Language-agnostic BERT Sentence Embedding (LaBSE) for quality assessment.
• Incorporated human evaluation to validate the generated parallel sentences.

Main Results:

• Achieved an 85.2% similarity score using GPT and LaBSE, outperforming other embedding models.
• Demonstrated the effectiveness of AI in generating high-quality parallel sentences.
• Indicated promising results for data-driven solutions in translation education.

Conclusions:

• AI, specifically GPT, can effectively address dataset limitations in translation education.
• The proposed method successfully transforms existing corpora into valuable didactic resources.
• This approach supports the development of more efficient and effective AI-powered translation teaching tools.