A Rising Tide Lifts All Boats: MTQE Rewards for Idioms Improve General Translation Quality

TL;DR

This work quantifies the non-compositional translation gap and offers insights for developing LLMs with stronger cross-cultural and figurative language understanding.

Abstract

Non-compositional expressions (e.g., idioms, proverbs, and metaphors) pose significant challenges for neural machine translation systems because their meanings cannot be derived from individual words alone. These expressions encode rich, cultural meaning, and have both figurative and literal meanings, making accurate translation difficult. Because models are fairly good at translating compositional text, we investigate GRPO-style fine-tuning using Machine Translation Quality Estimation (MTQE) models as reward functions to train models to better translate idioms. Using Chinese and Hindi idiom datasets, we find that idiom translation abilities improve by ~14 points, general, non-idiomatic translation implicitly improves by ~8 points, and cross-lingual translation abilities (trained on one language, evaluated on another) improves by ~6 points. Overall, our work quantifies the non-compositional translation gap and offers insights for developing LLMs with stronger cross-cultural and figurative language understanding.

Figures (11)

• Figure 1: There are three challenges when modeling and translating non-compositional phrases.
• Figure 2: Illustration of the distinction between all three GRPO-QE-* methods. (a) QE-Positive pulls semantically equivalent texts closer. (b) QE-Negative pulls semantically inequivalent texts apart. (c) QE-Constrained balances both. (d) QE-DA uses a ground-truth reference translation to inform MTQE.
• Figure 3: Evaluation of translation abilities of Chinese idioms. Here, we see that the LIA and TrainingFree (TF) baselines do well on Qwen-2.5-3B, but not on Llama-3.1-8B (hence is unreliable). The GRPO based methods are not only performant, but also reliable.
• Figure 4: Evaluation of translation abilities of Hindi idioms. Here, we see that the core translation models (NLLB and Command-R) translate Hindi idioms well, but nChinese idioms (Fig. \ref{['fig: chinese_eval']}). The GRPO based methods are not only performant, but also reliable.
• Figure 5: Evaluation of translation abilities of regular Chinese sentences. Here, it is shown that performance does not deteriorate when models are trained on idiomatic data.