Work

Enhancing Textual Datasets with Large Language Model-based Data Augmentation for Neural Machine Translation

Year: 2023
Type: dissertation
Abstract: [EN] This work proposes a parallel textual data augmentation framework for low-resource languages or low-resource domains using different technologies like Masked Language Models, Large Language Models, and Natural Language Processing tools and resources. The framework detects parts of the target text that can be replaced with new ones with the same grammatical category at the word level. Then, the replaced word is found in the source text using word alignment to replace it with a suitable synonym. The process includes iterative steps of identifying and replacing words until no further changes can occur. Our approach addresses the need for more extensive high-quality datasets in natural language processing tasks for low-resource situations. The proposed process is designed to support multiple languages and various text types. Large Language Models can be used to maintain the augmented data's quality and facilitate diverse textual datasets in both monolingual and multilingual cases. The quality of the augmented data is evaluated by humans based on various criteria such as fluency, coherence, and relevance, and automatic evaluations are performed to check the improvement in the language model's performance. This automated evaluation employs state-of-the-art metrics such as BLEU, TER, and chrF. The proposed framework aims to improve the quality and quantity of textual data that can be used for natural language processing tasks, such as machine translation.
Language: English
Open Access status: green
OpenAlex ID: W7019109500
Cites: 53
Cited by:
Related to: 9