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

Ying Li; Xinglin Lyu; Junhui Li; Jinlong Yang; Hengchao Shang; Min Zhang; Shimin Tao; Daimeng Wei

cs.CLMar 26, 2026

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

Authors: Ying Li, Xinglin Lyu, Junhui Li, Jinlong Yang, Hengchao Shang, Min Zhang, Shimin Tao, Daimeng Wei

Organizations: School of Computer Science and Technology, Soochow University, Suzhou, China · School of Computer Science and Artificial Intelligence, Zhengzhou University, Zhengzhou, China · Jiangsu Key Lab of Language Computing, China · Huawei Translation Services Center, Beijing, China

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, providing explicit supervision to exploit informative context while remaining robust to uninformative context. We validate the proposed approach on several public context-aware MT tasks using multiple models, including Qwen3-4B, Qwen3-8B, and Llama-3-8B-Instruct. Experimental results demonstrate consistent improvements in translation quality and robustness across both input conditions, achieved without any architectural modifications.

Figures & tables

Figure 1: Distribution of COMET differences between sentence-level (Sent) and context-aware (Ctx) English–German translations produced by Qwen3-8B.

Model	BLEU	COMET
Sentence-level	31.72	86.12
Context-aware	31.60	86.13
\hdashedline Oracle	33.91	87.17

Table 1: Overall English-to-German translation quality of sentence-level and context-aware translation. Oracle means we always select the better translation from the two systems for each sentence.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 En $\Rightarrow$ De		\Block [c]1-2 En $\Rightarrow$ Es		\Block [c]1-2 En $\Rightarrow$ Fr		\Block [c]1-2 En $\Rightarrow$ It		\Block [c]1-2 En $\Rightarrow$ Nl		\Block [c]1-2 En $\Rightarrow$ Ru		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-16 Qwen3-4B
1	Sent-Level	82.51	-	84.91	-	83.13	-	86.94	-	82.71	-	83.12	-	83.89	-
2	Sent-Level ${}_{\text{tuned}}$	84.41	-	86.88	-	84.62	-	87.97	-	84.88	-	84.58	-	85.56	-
3	Sent-Level ${}_{\text{CPO}}$	85.14	-	87.43	-	85.18	-	88.66	-	85.88	-	86.19	-	86.41	-
4	Ctx-Aware	-	83.25	-	86.98	-	84.47	-	87.81	-	83.47	-	83.86	-	84.97

Table 2: Performance comparison on COMET. Best and second-best scores are highlighted in dark red and light red . Sent and CTX denote sentence-level input (

x

) and context-aware input (

x

and context

c

) during inference.

Dataset	Sent	Ctx-gold	Ctx-random
NewsCommentary	82.50	82.71	82.52
IWSLT	78.06	78.70	78.02

Table 3: Performance (document-level COMET) on sentences with useful context. Ctx-random replaces the original context with randomly sampled one.

Model	Coh. $\uparrow$	ALTI+ $\uparrow$	Fluency $\uparrow$
Qwen3-8B	73.28	54.49	3.81
Sent+Ctx ${}_{\text{tuned}}$	74.78	54.67	3.90
CPL (ours)	75.24	54.90	4.04

Table 4: Additional quality analysis using coherence (Coh.), ALTI+, and GPT-based fluency.

Model	Entity	Tense	Pronoun	DM	BlonD-D
Qwen3-8B	61.71	65.44	74.41	58.01	64.66
Sent+Ctx ${}_{\text{tuned}}$	77.46	77.10	82.24	76.76	78.35
CPL (ours)	78.22	77.24	82.38	76.99	78.67

Table 5: Average BlonD-D scores on News Commentary X

\Rightarrow

En tasks with context-aware translation.

Model	Sent	CTX
Sent+Ctx ${}_{\text{tuned}}$	85.74	86.31
CPL	87.73	87.89

Table 6: Evaluation results of GPT-as-Judge.

Metric Used	Sent	CTX
BLEU	86.96	87.16
d-COMET	87.97	87.98
s-COMET	87.97	88.03
(d- + s-COMET)/2	87.99	88.06

Table 7: CPL performance (in COMET) using different metrics for selecting intra- and cross-condition preference pairs.

Model	Sent	CTX
CPL	87.99	88.06
CPL w/o $(y_{\text{w}}^{+},y_{\text{l}}^{+})$	87.98	88.01
CPL w/o $(y_{\text{w}}^{+},y_{\text{l}}^{-})$	87.88	87.91

Table 8: Effect of removing individual cross-condition preference constraints.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Data	Task	Test		CPL
Data	Task	#Doc	#Sent	#Doc	#Sent
News (En $\Rightarrow$ X)	En $\Rightarrow$ De	150	5,967	150	5,628
	En $\Rightarrow$ Es	150	5,819	150	5,782
	En $\Rightarrow$ Fr	150	5,795	150	5,890
	En $\Rightarrow$ It	150	5,749	150	5,948
	En $\Rightarrow$ Nl	151	5,992	151	5,925

Appendix

Table 9: Statistics of the test sets and of the data used for CPL preference selection, across the fifteen translation tasks in the News and IWSLT domains. #Doc and #Sent denote the numbers of documents and sentences.

Data	Task	$\mathcal{P}_{\text{s}}$	$\mathcal{P}_{\text{c}}$	$\mathcal{P}_{\text{cr}}$
News (En $\Rightarrow$ X)	En $\Rightarrow$ De	4,143	4,277	8,086
	En $\Rightarrow$ Es	4,963	5,051	8,736
	En $\Rightarrow$ Fr	4,768	4,877	8,829
	En $\Rightarrow$ It	4,923	5,024	8,877
	En $\Rightarrow$ Nl	3,904	4,072	8,056
	En $\Rightarrow$ Ru	4,016	4,116	7,841

Appendix

Table 10: Statistics of the preference pairs in cross-condition preference learning, across the fifteen translation tasks.

\mathcal{P}_{\text{s}}

\mathcal{P}_{\text{c}}

, and

\mathcal{P}_{\text{cr}}

denote the numbers of intra-condition sentence-level, intra-condition context-level, and cross-condition preference pairs, respectively.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 En $\Rightarrow$ De		\Block [c]1-2 En $\Rightarrow$ Es		\Block [c]1-2 En $\Rightarrow$ Fr		\Block [c]1-2 En $\Rightarrow$ It		\Block [c]1-2 En $\Rightarrow$ Nl		\Block [c]1-2 En $\Rightarrow$ Ru		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-16 Qwen3-4B
1	Sent-Level	79.00	-	81.29	-	79.50	-	83.26	-	79.35	-	79.23	-	80.27	-
2	Sent-Level ${}_{\text{tuned}}$	81.55	-	84.25	-	81.74	-	84.81	-	81.94	-	81.02	-	82.55	-
3	Sent-Level ${}_{\text{CPO}}$	82.40	-	84.90	-	82.37	-	85.57	-	82.98	-	82.73	-	83.49	-
4	Ctx-Aware	-	80.26	-	84.46	-	81.59	-	84.38	-	80.29	-	80.22	-	81.87

Appendix

Table 11: Results of different systems on document-level COMET metric. Best scores are shown in dark red and second-best scores are in light red . Sent and CTX denote sentence-level and context-aware inputs, respectively.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 En $\Rightarrow$ De		\Block [c]1-2 En $\Rightarrow$ Es		\Block [c]1-2 En $\Rightarrow$ Fr		\Block [c]1-2 En $\Rightarrow$ It		\Block [c]1-2 En $\Rightarrow$ Nl		\Block [c]1-2 En $\Rightarrow$ Ru		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-16 Qwen3-4B
1	Sent-Level	12.64	-	28.95	-	22.16	-	22.75	-	14.78	-	14.38	-	19.28	-
2	Sent-Level ${}_{\text{tuned}}$	24.92	-	40.05	-	31.94	-	35.38	-	24.38	-	22.20	-	29.81	-
3	Sent-Level ${}_{\text{CPO}}$	28.99	-	41.73	-	35.16	-	36.25	-	28.61	-	26.49	-	32.87	-
4	Ctx-Aware	-	11.09	-	28.98	-	21.21	-	21.35	-	12.72	-	13.05	-	18.07

Appendix

Table 12: Results of different systems on sentence-level BLEU metric. Best scores are shown in dark red and second-best scores are in light red . Sent and CTX denote sentence-level and context-aware inputs, respectively.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 Es $\Rightarrow$ En		\Block [c]1-2 Fr $\Rightarrow$ En		\Block [c]1-2 Nl $\Rightarrow$ En		\Block [c]1-2 Ru $\Rightarrow$ En		\Block [c]1-2 Zh $\Rightarrow$ En		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-14 Qwen3-8B
1	Sent-Level ${}_{\text{tuned}}$	89.13	-	87.64	-	89.85	-	84.76	-	86.61	-	87.60	-
2	Ctx-Aware ${}_{\text{tuned}}$	-	89.29	-	87.79	-	90.08	-	84.94	-	86.85	-	87.79
3	Sent+Ctx ${}_{\text{tuned}}$	89.20	89.39	87.72	87.89	90.01	90.18	84.85	85.00	86.55	86.93	87.67	87.88
4	CPL (Ours)	89.58	89.73	88.20	88.32	90.33	90.49	85.34	85.50	87.23	87.44	88.14	88.30

Appendix

Table 13: Results of X-to-English translations on sentence-level COMET using Qwen3-8B. Best scores are shown in dark red and second-best scores in light red . Sent and CTX denote sentence-level and context-aware inputs, respectively.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 Es $\Rightarrow$ En		\Block [c]1-2 Fr $\Rightarrow$ En		\Block [c]1-2 Nl $\Rightarrow$ En		\Block [c]1-2 Ru $\Rightarrow$ En		\Block [c]1-2 Zh $\Rightarrow$ En		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-14 Qwen3-8B
1	Sent-Level ${}_{\text{tuned}}$	87.36	-	85.81	-	88.43	-	82.69	-	84.30	-	85.72	-
2	Ctx-Aware ${}_{\text{tuned}}$	-	87.66	-	86.10	-	88.74	-	82.98	-	84.62	-	86.02
3	Sent+Ctx ${}_{\text{tuned}}$	87.45	87.74	85.89	86.17	88.59	88.84	82.75	83.00	84.24	84.68	85.78	86.09
4	CPL (Ours)	87.84	88.10	86.41	86.62	88.91	89.13	83.31	83.57	84.96	85.27	86.29	86.54

Appendix

Table 14: Results of X-to-English translations on document-level COMET using Qwen3-8B. Best scores are shown in dark red and second-best scores in light red . Sent and CTX denote sentence-level and context-aware inputs, respectively.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 En $\Rightarrow$ De		\Block [c]1-2 En $\Rightarrow$ Zh		\Block [c]1-2 Zh $\Rightarrow$ En		\Block [c]1-2 En $\Rightarrow$ Fr		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-12 Qwen3-8B
1	Sent-Level ${}_{\text{tuned}}$	83.37	-	82.67	-	82.18	-	84.37	-	83.15	-
2	Ctx-Aware ${}_{\text{tuned}}$	-	83.65	-	83.16	-	82.56	-	84.52	-	83.47
3	Sent+Ctx ${}_{\text{tuned}}$	83.71	83.80	82.81	83.10	82.19	82.67	84.56	84.70	83.32	83.57
4	CPL (Ours)	83.75	84.14	83.83	84.20	83.12	83.56	85.24	85.47	83.98	84.34

Appendix

Table 15: Results of IWSLT2017 translations on sentence-level COMET using Qwen3-8B (up to 256 context tokens). Best scores are in dark red and second-best in light red . Sent and CTX denote sentence-level and context-aware inputs, respectively.

\Block [l]2-1 #	\Block [l]2-1 Model	\Block [c]1-2 En $\Rightarrow$ De		\Block [c]1-2 En $\Rightarrow$ Zh		\Block [c]1-2 Zh $\Rightarrow$ En		\Block [c]1-2 En $\Rightarrow$ Fr		\Block [c]1-2 Average
		Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX	Sent	CTX
\Block [c]1-12 Qwen3-8B
1	Sent-Level ${}_{\text{tuned}}$	78.27	-	76.48	-	77.57	-	80.45	-	78.19	-
2	Ctx-Aware ${}_{\text{tuned}}$	-	78.69	-	77.12	-	78.23	-	80.72	-	78.69
3	Sent+Ctx ${}_{\text{tuned}}$	78.68	78.91	76.77	77.25	77.61	78.38	80.63	80.92	78.42	78.87
4	CPL (Ours)	78.95	79.39	77.64	78.19	78.62	79.22	81.45	81.73	79.17	79.63

Appendix

Table 16: Results of IWSLT2017 translations on document-level COMET using Qwen3-8B (up to 256 context tokens). Best scores are in dark red and second-best in light red . Sent and CTX denote sentence-level and context-aware inputs, respectively.

Figure 3: COMET under different context lengths (256/512/1024 tokens) on En

\Rightarrow

De and En

\Rightarrow

Es, for sentence-level ( Sent , solid) and context-aware ( CTX , dashed) inputs. CPL is shown in green and the Sent+Ctx

{}_{\text{tuned}}

baseline in blue.

Model	Avg. Score
Sent+Ctx ${}_{\text{tuned}}$ (Sent)	73.83
Sent+Ctx ${}_{\text{tuned}}$ (CTX)	73.44
CPL (Sent)	73.89
CPL (CTX)	74.34

Appendix

Table 17: Human evaluation (average total score out of 100) on En

\Rightarrow

Es.

Figure 4: Distribution (%) of sentences by COMET differences between context-aware and sentence-level translations. CB/B/P/W/CW = Clearly Better , Better , On Par , Worse , Clearly Worse .

Model	BLEU	COMET
Sent-Level	36.61	87.99
Ctx-Aware	36.80	88.06
+ Prob.	37.14	88.19
+ COMETKiwi	37.13	88.42
+ Oracle	38.63	88.73

Appendix

Table 18: Performance of reranking sentence-level and context-aware translations produced by CPL.

Panel of Figure 5 — Figure 5: Case studies illustrating that CPL benefits from informative context.

Panel of Figure 6 — Figure 6: Prompt used for context-dependency evaluation of sentence-level translation.

Panel of Figure 7 — Figure 7: Strict JSON schema used to constrain model outputs in batched API inference.

Panel of Figure 8 — Figure 8: Prompt used for Sentence-level translation (Sent-Level).

Panel of Figure 9 — Figure 9: Prompt used for Context-Aware translation (Ctx-Aware).

Panel of Figure 10 — Figure 10: Prompt used for GPT-as-Judge.

Context-Aware MT

See all 12 papers

G^2C-MT: Graph-Guided Context Selection for Document-Level Machine Translation
Jun 2, 2026Baijun Ji, Zixuan Zhou, Xiangyu Duan +4Graph-Based Retrieval Document-Level MT
School of Computer Science and Technology, Soochow University
Scaling Attention Head Analysis via Gradient-Based Attribution in Context-Aware Machine Translation
Sep 23, 2026Paweł Mąka, Yusuf Can Semerci, Jan Scholtes +1Gradient-Based Attribution Attention Head Analysis
Department of Advanced Computing Sciences Maastricht University
Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection
May 28, 2026Yutong Wang, Xuebo Liu, Derek F. Wong +6Adaptive Retrieval Document-Level MT
Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen · 2NLP2CT Lab, Department of Computer and Information Science, University of Macau · 3Huawei Translation Services Center

Jun 2, 2026J/K move · Enter open · S save

G^2C-MT: Graph-Guided Context Selection for Document-Level Machine Translation

Baijun Ji, Zixuan Zhou, Xiangyu Duan, Yu Liu+3

School of Computer Science and Technology, Soochow University

Graph-Based Retrieval Document-Level MT Machine Translation

Open paper AskPDF arXiv

Effective document-level machine translation (DocMT) requires capturing long-range discourse dependencies. Recent work has explored retrieval-based and discourse-aware context selection. However, these approaches often lack an explicit mechanism for modeling structured discourse dependencies between distant paragraphs in a document. In this paper, we propose G^2C-MT (Graph-Guided Context for Machine Translation), which views DocMT context selection as a structured path discovery problem on a lightweight discourse graph, rather than retrieving unstructured context sets or relying on expensive LLM-based discourse modeling. In detail, we represent each paragraph as a node and model the relationship between each pair of nodes, considering their semantic similarity, adjacency, and keyword overlap. Furthermore, we propose a depth-biased random walk over the graph to sample a backward context path for each target paragraph. The context path will be used to prompt a large language model (LLM) for translation. This framework naturally supports multi-path context sampling, which can improve robustness by aggregating diverse translation candidates for discourse-ambiguous inputs. Experiments conducted across various domains show that G^2C-MT outperforms strong baselines on multiple LLMs, including DeepSeek-V3, Gemini-2.5-Flash-lite, and the Qwen-2.5/3 series.

Abstract

Context-Aware MT

Explore similar work

G^2C-MT: Graph-Guided Context Selection for Document-Level Machine Translation

Scaling Attention Head Analysis via Gradient-Based Attribution in Context-Aware Machine Translation

Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection