Language Models Can Predict the Most Suitable Translation Techniques, Study Finds

LLM Predict Translation Techniques

In a March 21, 2024 paper, Fan Zhou and Vincent Vandeghinste from KU Leuven demonstrated that language models can predict the most suitable translation techniques for translation and post-editing tasks. 

The researchers highlighted a set of persistent issues that remain in MT such as word-for-word translation, false friends, ambiguity, information omission or addition, and cultural insensitivity, leading to low-quality translations that may lack clarity and accuracy. These issues arise from the system using incorrect translation techniques, something a translator wouldn’t do. “The human-generated translation process relies on diverse translation techniques, which proves essential to ensuring both linguistic adequacy and fluency,” they emphasized.

Additionally, they highlighted that “utilizing translation techniques is crucial for addressing translation problems, improving translation quality, and ensuring contextually appropriate translations.”

Zhou and Vandeghinste suggested that automatically identifying translation techniques before can effectively guide and improve the machine translation (MT) process. Additionally, these techniques can serve as prompts for large language models (LLMs) to generate high-quality translations.

They distinguished between two translation scenarios: from-scratch translation and post-editing. For each scenario, they investigated whether pre-trained cross-lingual language models — such as mBART, mBERT, and mT5 — can be fine-tuned to accurately predict translation techniques so as to provide guidance for producing good translations in both from-scratch translation and PE or even automatic post-editing (APE) processes. 

 “Utilizing translation techniques is crucial for addressing translation problems, improving translation quality, and ensuring contextually appropriate translations.”

High Predictive Accuracy

To fine-tune the models they used 100,000 data pairs, each containing a source sentence, a target sentence, an aligned word or phrase in both languages, and a label indicating the translation technique used. 

Zhou and Vandeghinste focused on the English-Chinese language pair and considered eleven translation techniques, as defined in the Annotation Guidelines of Translation Techniques for English-Chinese:

  • literal translation: word-for-word translation
  • equivalence: non-literal translation of proverbs, idioms, or fixed expressions
  • transposition: changing grammatical categories (without altering the meaning)
  • modulation: introducing a slight meaning change at lexical level based on context
  • modulation+transposition: combining modulation and transposition
  • particularization: specifying the meaning of a segment in context or translating a pronoun by the thing(s) it references
  • generalization: translating an idiom by a non-fixed expression or removing a metaphorical imagery
  • figurative translation: using an idiom to translate a non-fixed expression, or a metaphorical expression for a non-metaphor
  • lexical shift: changing verbal tense, modality, determiner, singular/plural forms, and other minor changes alike
  • explication: providing clarifications that are implicit in the source text
  • reduction: deliberately omitting certain words in translation

They found that pre-trained models, once fine-tuned for both the from-scratch translation and post-editing scenarios, “can proficiently predict the most suitable translation techniques.” Specifically, the results indicate high predictive accuracy for both scenarios, with 82% for from-scratch translation and 93% for post-editing. Zhou and Vandeghinste noted that “the post-editing process shows even greater promise.”

The authors acknowledged that the current focus was on the models’ ability to accurately predict the most suitable translation techniques for both tasks. In the future, they plan to explore how information about the most suitable translation techniques can guide NMT systems or LLMs to generate better translations.

Zhou and Vandeghinste believe that this study’s findings can “pave the way for future advancements in the field of MT generation.”