Cross-Preference Learning for Sentence-Level and Context-Aware Machine Translation

2026-03-26

AI Summary

Existing context-aware machine translation approaches have failed to consistently outperform sentence-level baselines, primarily because they do not explicitly model the varying degrees to which different sentences benefit from contextual information. This work proposes a Cross-Preference Learning (CPL) framework that introduces, for the first time, a cross-conditional preference mechanism. By integrating intra-sentence and cross-conditional preference signals, CPL explicitly guides the model to adaptively determine when and how to leverage context. Trained via preference optimization without any architectural modifications, the method is compatible with large language models such as Qwen3-4B, Qwen3-8B, and Llama-3-8B. It achieves consistent and robust performance gains across multiple public context-aware translation benchmarks, effectively harmonizing sentence-level accuracy with contextual coherence.

Technology Category

• Machine Learning: Learning Preferences or Rankings
• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Humans and AI: Learning Human Values and Preferences

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Context-aware machine translation (MT) leverages document-level information, yet it does not consistently outperform sentence-level MT, as contextual signals are unevenly beneficial across sentences. Existing training objectives do not explicitly model this variability, limiting a model's ability to adaptively exploit context. In this paper, we propose Cross-Preference Learning (CPL), a preference-based training framework that explicitly captures the complementary benefits of sentence-level and context-aware MT. CPL achieves this by integrating both intra- and cross-condition preferences into the preference optimization objective. The introduction of intra- and cross-condition preferences provides explicit supervision on when and how contextual information improves translation quality. We validate the proposed approach on several public context-aware MT tasks using multiple models, including Qwen3-4B, Qwen3-8B, and Llama-3-8B. Experimental results demonstrate consistent improvements in translation quality and robustness across both input conditions, achieved without any architectural modifications.

Problem

• context-aware machine translation
• sentence-level translation
• contextual variability
• preference learning
• translation quality

Innovation

• Cross-Preference Learning
• context-aware machine translation
• preference optimization
• intra-condition preference
• cross-condition preference