Enhancing Machine Translation Models through Optimal Strategies for Prior Knowledge Integration: A Systematic Review

Abstract

Machine Translation (MT) has significantly transformed cross-language communication, becoming a pivotal field in Natural Language Processing (NLP). This systematic literature review and bibliometric analysis explore the optimal strategies for integrating prior knowledge into machine translation models. We delve into the strategic aspects of incorporating linguistic proficiency, external semantic information, and attention mechanisms, aiming to answer the research question: "What is the best strategy for integrating prior knowledge into machine translation?" The study employs a methodically crafted search strategy, focusing on Scopus and IEEE Xplore databases from 2014 to 2024. The rigorous selection process, following PRISMA guidelines, results in 22 articles for comprehensive analysis. The literature review highlights the significance of incorporating linguistic proficiency, semantic understanding, and domain-specific knowledge into machine translation models. Scholars consistently advocate for leveraging both syntactic and semantic structures from the source language to enhance model performance. The analysis of 22 selected articles presents a variety of strategies, including node random dropping, shared vocabulary and cognate lexicon induction, word-level domain-sensitive representation learning, and knowledge distillation. Notably, the "Node Random Dropping Strategy" stands out for capturing both semantic and syntactic representations, while the "Shared Vocabulary and Cognate Lexicon Induction" strategy proves effective for languages with shared linguistic features. In conclusion, this systematic review identifies a consensus among scholars on the importance of integrating syntactic and semantic language structures as crucial forms of prior knowledge. To effectively implement these structures, the study recommends employing a graph neural network from the encoder side. This comprehensive analysis contributes valuable insights for future research in optimizing strategies for prior knowledge integration into machine translation models.