Leveraring multilingual models and rank fusion for translation memory retrieval

Abstract

Translation memories (TMs) are crucial components of modern Computer-Assisted Translation (CAT) tools. TMs store translated texts from a source language to a target language, serving as a repository of previously translated segments that are essential for productivity and cost reduction in the translation process. However, current TM databases rely mostly on lexical rules, such as edit distance or n-gram matches, which limit their ability to identify semantically similar translations. Recently, researchers have been exploring the use of neural models for retrieval tasks, but with a limited scope that fails to fully leverage the multilingual nature of Translation Memories (TMs) and available neural models and tools. In this study, we explore the application of state-of-the-art neural models for the Translation Memory Retrieval problem and present our Robust Translation Memory Retrieval (RTMR) pipeline, which combines neural models and information retrieval techniques to achieve state-of-the-art results. Furthermore, we conduct experiments using a wide range of TMs and different languages as source and target, expanding the scope of previous studies that have often been limited to a single TM and language direction. Through extensive experimentation, we demonstrate that neural models not only yield superior candidate translations but also offer greater flexibility and wider applicability compared to conventional lexical approaches. Furthermore, we show that the integration of rank fusion techniques and multilingual neural models results in state-of-the-art performance for Translation Memory Retrieval. Our findings highlight the potential of neural models to significantly enhance the effectiveness of Translation Memory systems.