Machine Translation (MT) has traditionally relied on parallel corpora, posing significant challenges for low-resource languages. This thesis reimagines Machine Translation as a style transfer problem, introducing a novel framework, CycleGN, to enable both translation and style transfer without requiring parallel data sets. Using monolingual corpora, the proposed method seeks to broaden equitable access to language translation systems, while advancing theoretical insights into style transfer. Another contribution of this work is the development of Tokengram_F, a novel metric that extends n-gram analysis to better capture linguistic and contextual nuances in translation evaluation and can estimate the quality of machine-generated sentences in more than 200 different languages. Furthermore, Embed_llama leverages pre-trained Large Language Model (LLM) embeddings to enhance semantic alignment and evaluation accuracy, deepening the work on transfer learning. This thesis also explores text compression through the development of Llamazip, a lossless compression algorithm that uses the predictive capabilities of LLMs. Beyond achieving excellent compression ratios, Llamazip demonstrates innovative applications, such as identifying training set membership of given target text and benchmarking predictive performance. The research presented in this thesis has led to the publication of four peer-reviewed publications, the submission of another one, and we are also in the process of writing a further paper. Ultimately, this work seeks to democratise access to translation technologies by broadening the scope of accessible training data. It aims to contribute to the evolution of language technologies in a multilingual world.