Presented under the opening title “Attention Is All You Need,” Rosalia Ignatova’s “How Does Machine Translation Work?” formed part of Thinking Literature in Translation in November 2025. The presentation introduces the Transformer architecture and explains why attention mechanisms became foundational to contemporary machine translation and generative language models.
Ignatova describes how self-attention allows a model to evaluate relationships among all the words in a sequence. Token embeddings and query, key and value vectors enable the system to interpret context and long-range dependencies, while decoder-only models use large-scale multilingual training to generate coherent text without a separate encoder.
English-to-Bulgarian literary examples produced by systems including Claude and ChatGPT demonstrate both the fluency and the limitations of machine-generated translations. The presentation considers terminology, pronoun reference, rhythm, register and stylistic consistency, comparing direct or zero-shot translation with approaches that first provide the model with explanations and contextual guidance.
Poetry and culturally specific writing receive particular attention because technically correct output may still lose ambiguity, sound, humour or imaginative force. Ignatova ultimately distinguishes computational capacity from human creativity: AI can support analysis and produce useful drafts, but literary translation continues to depend on interpretation, aesthetic judgement and purposeful human decisions.
