A Novel Method of Translation Memory to Improve Machine Translation

Numerous studies have provided evidence of the performance enhancement achieved in neural machine translation (NMT) through the utilization of translation memory (TM). Diverging from current approaches that rely on bilingual corpora as translation memories (TMs) and employ source-side similarity search for memory retrieval, our proposed framework adopts monolingual memories and incorporates a learnable memory retrieval mechanism that operates in a cross-lingual manner. The primary aspect of our framework is a cross-lingual in-memory retriever that enables translation memory. Tutilization using abundant monolingual data. This allows us to leverage the richness of monolingual resources effectively. Additionally, our framework facilitates the co-optimization of the memory retriever and the neural machine translation model to jointly work towards achieving the ultimate translation objective. The experimental results demonstrate that our model not only surpasses neural machine translation baselines that are augmented with bilingual translation memories but also exhibits superior performance. The ability of our model to leverage monolingual data enables its effectiveness in scenarios with limited resources and domain adaptation requirements. This showcases the versatility of our framework, making it well-suited for low-resource settings and facilitating domain-specific translations.