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

Cross-Preference Learning (CPL) introduces a novel preference-based training framework designed to address a critical limitation in current machine translation (MT) systems: the assumption that contextual information is always beneficial. The paper establishes that context-aware MT does not consistently outperform sentence-level MT, as contextual signals are unevenly beneficial across sentences. Existing models lack an explicit mechanism to decide when to rely on context, leading to suboptimal behavior when context is noisy or redundant. CPL provides a principled solution by explicitly modeling the interactions between these two conditions, enabling a single shared-parameter model to adaptively exploit context and achieve robust performance across both input scenarios.

The Problem of Dual Conditions

Modern MT systems often operate under multiple input conditions: sentence-level translation (only the source sentence x is available) or context-aware translation (the source sentence x plus surrounding document context c is provided). While these two settings share the same output space, they are typically handled by separate models. The paper highlights that context-aware translation does not consistently outperform sentence-level translation, leading to a suboptimal training objective where models are trained under a fixed input condition, implicitly assuming context is always beneficial.

How Cross-Condition Preference Learning Works

CPL builds upon Contrastive Preference Optimization (CPO) by introducing Cross-Condition Preference Learning (Cross-CPO or C-CPO) to explicitly model the interactions between the two translation conditions. The methodology involves several key steps:

• Defining Distributions: The model theta defines two conditional distributions over target sentences: p theta(yx) for sentence-level and p theta(yx, c) for context-aware translation.

• Selecting Preference Pairs: For each candidate pair, preference labels are constructed by scoring candidates using the average of sentence-level COMET and document-level COMET (dCOMET) against the reference. The higher-scoring candidate is designated y+, while the lower-scoring one is y-.

• Applying Loss Functions: The overall CPL objective, LCPL(theta), unifies two types of preference signals:

1. Intra-Condition Preference Learning (L intra(theta)): Standard CPO applied independently to the selected pairs for each condition (P s and P c).

2. Cross-Condition Preference Learning (L cross(theta)): This is the C-CPO loss, which directly compares sentence-level and context-aware translations for the same source sentence, encouraging the model to consistently prefer the winning condition over both strong and weak outputs from the losing condition.

Experimental Evaluation and Results

The approach was validated on multiple public document-level MT tasks using large foundation models, including Qwen3-4B, Qwen3-8B, and Llama-3-8B. The results demonstrate consistent improvements in translation quality across both input conditions:

• CPL achieved substantial gains over baselines like Sent+Ctxtuned+ (which used the same data), yielding improvements of 1.10/1.12 for sentence-level/context-aware translation using Qwen3-4B, for example.

• The model proved robust to context noise; when the original context was replaced with randomly sampled sentences, the performance of context-aware CPL becomes nearly identical to that of sentence-level translation.

• CPL also exhibited strong consistency across all three LLM backbones, suggesting it is a model-agnostic approach.

Conclusion and Future Work

The paper concludes that CPL provides a principled framework for multi-task preference learning, allowing a single shared-parameter model to perform robustly under both settings without requiring any architectural modifications. While limitations exist—such as reliance on automatic metrics (COMET) and increased computational cost during training—the CPL approach offers a significant step toward achieving adaptive, context-aware machine translation.

As a diligent and fastidious researcher, my analysis of this paper identifies Cross-Preference Learning (CPL) not merely as a training objective, but as a fundamental solution to the problem of adaptive context utilization in machine translation. The core innovation is moving beyond the assumption that more context is always better.

Based on the findings presented in this paper, here are specific improvements and capabilities for AI systems, particularly those involved in Document-Level Machine Translation (DocMT) or advanced Natural Language Understanding (NLU):

A. Implementation of Cross-Condition Preference Learning (CPL)

• What to do: Integrate the CPL loss function (LCPL(theta) = Lintra(theta) + Lcross(theta)) into the fine-tuning pipeline for a shared-parameter LLM. This requires defining and collecting three distinct sets of preference pairs:

1. Intra-Condition Pairs (P s, P c): Standard preference pairs derived from within the sentence-level (S) and context-aware (C conditions).

2. Cross-Condition Pairs (P cr: The crucial step: For a given source sentence x with context c, identify the best translation (y+) across all four candidates (ys+, ys-, yc+, yc-. Then, construct pairs that explicitly compare the winning condition's output against both strong and weak outputs of the losing condition.

• What the improved system can do: This allows the system to learn a shared preference structure across conditions. The model will not only learn how to translate well (Intra-Condition) but also when to rely on context versus when to ignore it (Cross-Condition).

B. Adaptive Context Selection and Reranking Mechanism

• What to do: Implement a unified decoding strategy that does not commit solely to the sentence-level or context-aware output. Instead, the model should generate both candidates (y and y') and then apply a dynamic reranking strategy.

• The Reranker: This module must be trained/tuned to evaluate the semantic utility of the context for that specific sentence. It should use metrics like = COMET ctx - COMET sent (as seen in Figure 3) to decide which output is superior for a given input.

• What the improved system can do: The system gains adaptive context utilization. It can intelligently select the translation that maximizes quality, whether that is achieved by leveraging subtle context (High) or by being robust to noisy/irrelevant context (Low/On Par).

A. Robustness to Noisy and Ambiguous Context

• What the improved system can do: Unlike traditional DocMT models that are forced to use all context regardless of utility, a CPL-optimized system is robust against noisy context. If the surrounding sentences are irrelevant or contradictory, the model can correctly revert to its high-quality sentence-level translation, preventing degradation.

• Specific Example: In legal texts (like Case 2 in Figure 4), if a sentence's meaning is ambiguous without context, CPL uses it. If the context is tangential or misleading, CPL ignores the tendency to over-rely on that specific contextual signal.

B. Improved Semantic Precision and Structural Fidelity

• What the improved system can do: The system becomes highly effective at resolving lexical ambiguity and coreference issues. By leveraging cross-condition preference learning, it ensures that the chosen translation not only conveys the literal meaning but also maintains the correct discourse structure (e.g., correctly translating delusions vs. wahnvorstellungen in Case 1).

C. Enhanced Evaluation and Quality Assurance

• What the improved system can do: By integrating metrics beyond standard BLEU/COMET (like coherence, faithfulness, and fluency—Table 4), the system allows for a comprehensive quality assessment framework. This enables continuous improvement loops where translation errors are flagged not just by semantic mismatch, but also by structural inconsistency across document boundaries.

Feature Traditional DocMT Model CPL-Optimized System

:---:---:---

Context Usage Fixed: Always use all available context. (Implicit assumption) Adaptive: Select the best translation based on contextual utility. (Explicit learning)

Robustness Vulnerable to noisy/misleading context, leading to degradation. Highly robust; maintains high performance even when context is irrelevant or contradictory.

Goal Achieve highest possible average score by using maximum available information. Achieve the highest quality translation for each specific sentence, regardless of input condition utility.