Multilingual Whispers: Generating Natural Translation through Multilingual Paraphrases

Machine translationMultilingual NLP

Can multilingual paraphrasing produce more natural translations than direct LLM translation? A controlled comparison with native-speaker evaluation.

About the project

Machine Translation has been the go to method for transforming data into a target language. However it often leads to translationese content which does not demonstrate a good quality sentence/document in the target language. Prior works showcase that paraphrasing can be used to produces diverse multilingual translation data which improves the quality of the produced output sentences/documents. In the era of LLM, such machine translation and multilingual paraphrasing tasks can simply be accomplished through prompting, e.g., “You are a professional translator. Please translate the following sentence into language {lang} while preserving the naturalness and semantic meaning: {sentence}” and “You are a native speaker for language {lang}. Please naturally paraphrase the following sentence in your own language while preserving the meaning: {sentence}”. In this work, we want to analyze and understand whether translation through multilingual paraphrasing can really be a tool for improving the translation quality.

What mentees potentially work on

  • LLM Inference for Machine Translation & Multilingual Paraphrase
  • Designing and running a rigorous experiment to compare the two methods
  • Conduct human evaluation to ensure human alignment to the native speakers

Requirements and recommended background

  • Basic programming
  • Basic knowledge about NLP / AI / LLM
  • Experience and passion for multilingual NLP research

There are no formal eligibility limits. See who can apply; you can still apply if you don't match every item.

Relevant publications

Mentors

Samuel Cahyawijaya
Samuel Cahyawijaya
Cohere
Member of Technical Staff
Clara Vania
Clara Vania
Amazon
Applied Scientist