Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Utterance to Vividity: Training Expressive Subtitle Translation LLM via Adaptive Local Preference Optimization

Chaoqun Cui, Shijing Wang, Liangbin Huang, Qingqing Gu, Zhaolong Huang, Xiao Zeng, Wenji Mao

TL;DR

The paper addresses the challenge of domain specific subtitle translation by training an expressive subtitle translation LLM with Adaptive Local Preference Optimization ALPO, which leverages LLMs as judges for fine grained local preferences. It introduces a multidirectional visual media subtitle corpus MuSC and demonstrates that per segment local alignment improves vividness without sacrificing overall quality. Empirical results across multiple directions show ALPO outperforms strong baselines on accuracy, naturalness, and especially vividness, with human evaluation corroborating the gains. The work releases data and code to support reproducibility and suggests multimodal extensions to further improve subtitle translation in real-world settings.

Abstract

The rapid development of Large Language Models (LLMs) has significantly enhanced the general capabilities of machine translation. However, as application scenarios become more complex, the limitations of LLMs in vertical domain translations are gradually becoming apparent. In this study, we focus on how to construct translation LLMs that meet the needs of domain customization. We take visual media subtitle translation as our topic and explore how to train expressive and vivid translation LLMs. We investigated the situations of subtitle translation and other domains of literal and liberal translation, verifying the reliability of LLM as reward model and evaluator for translation. Additionally, to train an expressive translation LLM, we constructed and released a multidirectional subtitle parallel corpus dataset and proposed the Adaptive Local Preference Optimization (ALPO) method to address fine-grained preference alignment. Experimental results demonstrate that ALPO achieves outstanding performance in multidimensional evaluation of translation quality.

From Utterance to Vividity: Training Expressive Subtitle Translation LLM via Adaptive Local Preference Optimization

TL;DR

The paper addresses the challenge of domain specific subtitle translation by training an expressive subtitle translation LLM with Adaptive Local Preference Optimization ALPO, which leverages LLMs as judges for fine grained local preferences. It introduces a multidirectional visual media subtitle corpus MuSC and demonstrates that per segment local alignment improves vividness without sacrificing overall quality. Empirical results across multiple directions show ALPO outperforms strong baselines on accuracy, naturalness, and especially vividness, with human evaluation corroborating the gains. The work releases data and code to support reproducibility and suggests multimodal extensions to further improve subtitle translation in real-world settings.

Abstract

The rapid development of Large Language Models (LLMs) has significantly enhanced the general capabilities of machine translation. However, as application scenarios become more complex, the limitations of LLMs in vertical domain translations are gradually becoming apparent. In this study, we focus on how to construct translation LLMs that meet the needs of domain customization. We take visual media subtitle translation as our topic and explore how to train expressive and vivid translation LLMs. We investigated the situations of subtitle translation and other domains of literal and liberal translation, verifying the reliability of LLM as reward model and evaluator for translation. Additionally, to train an expressive translation LLM, we constructed and released a multidirectional subtitle parallel corpus dataset and proposed the Adaptive Local Preference Optimization (ALPO) method to address fine-grained preference alignment. Experimental results demonstrate that ALPO achieves outstanding performance in multidimensional evaluation of translation quality.
Paper Structure (59 sections, 17 equations, 8 figures, 14 tables, 1 algorithm)

This paper contains 59 sections, 17 equations, 8 figures, 14 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-as-a-Judge
  4. Language Model Preference Optimization
  5. Empirical Investigation
  6. LLM Is Excellent Translation Evaluator
  7. Parallel Corpora Are Actually Not Parallel
  8. Chat LLMs Favor Literal, Reason LLMs Excel in Liberal
  9. Method
  10. Notations
  11. Overall Framework
  12. Sampling Strategy
  13. Adaptive Alignment Loss
  14. Experiments
  15. Experimental Settings
  16. ...and 44 more sections

Figures (8)

  • Figure 1: Investigation of the consistency of multiple evaluators with Spearman rank correlation $\rho$.
  • Figure 2: Bland-Altman plot comparing Qwen3-14B with human evaluator.
  • Figure 3: Pairwise BLEU scores between translations.
  • Figure 4: The overall framework of ALPO.
  • Figure 5: Human assessment consistency verification.
  • ...and 3 more figures