LLM-based Translation Inference with Iterative Bilingual Understanding

Abstract

Large Language Models (LLMs) have significantly advanced machine translation (MT), yet they remain susceptible to "Understanding Distortion"—a phenomenon where the model misinterprets source context, leading to semantic hallucinations. This paper introduces Iterative Bilingual Understanding Translation (IBUT), a novel framework that leverages iterative alignment of bilingual contextual understandings. By generating explanations in both source and target languages and refining them through an automated feedback loop, IBUT significantly reduces translation errors in complex domains.

Introduction and Problem Statement

Despite their capabilities, LLMs often fail to capture subtle semantic nuances, particularly in cultural or commonsense contexts. "Understanding Distortion" occurs when a model adopts a literal interpretation of a polysemous word when a metaphorical one is required. For instance, the Chinese term huāpíng (vase) can literally mean a container or metaphorically refer to "eye candy" (someone attractive but lacking ability).

Expand

Source, p. 1

Research indicates that such distortions account for nearly 40% of translation errors in challenging datasets. Existing MT approaches often rely on single-pass reasoning, which lacks the robustness to identify and correct these deep-seated semantic mismatches.

The IBUT Framework

The proposed IBUT framework treats translation as an iterative process of cross-lingual meaning alignment. It operates through four distinct stages:

1. Understanding Generation: The LLM generates detailed explanations for key terms in both the source ($C_s$) and target ($C_t$) languages independently.
2. Alignment Judgment: A dedicated agent compares these bilingual understandings. If inconsistencies are detected (e.g., literal vs. metaphorical), a conflict is flagged.
3. Iterative Refinement: Upon detecting a conflict, the agent provides "verbal feedback"—specific instructions to reconcile the meanings. The model updates its understandings ($C'_s, C'_t$) until alignment is achieved.
4. Understanding-Based Translation: The final translation is produced using the original source sentence $S$ augmented by the refined bilingual insights.

Expand

Source, p. 3

This approach utilizes the principle of semantic invariance: the core meaning of a sentence should remain consistent regardless of the language used to describe it.

Experimental Evaluation

IBUT was evaluated across several rigorous benchmarks, including WMT22/23 general news, Commonsense MT (lexical ambiguity), and Cultural MT (7,000+ cultural-specific items across 6 language pairs).

Key Results

IBUT consistently outperformed established baselines, including MAPS, MAD, and standard few-shot prompting:

• Performance Gains: The method achieved an average improvement of +1.5 COMET points in general news and up to +6.5 COMET points in low-resource settings like Livonian-English.
• Cultural Nuance: In English-to-Tamil translation, IBUT surpassed ChatGPT by +5.5 COMET points.
• Error Correction: The framework successfully resolved approximately 89% of identified "Understanding Distortion" cases.

Discussion and Insights

The success of IBUT highlights several critical aspects of LLM-based translation:

• Bilingual Synergy: Generating understandings in both languages is more effective than monolingual approaches. The target-language understanding often provides a stronger corrective signal for the final output.
• Verbal Feedback as Gradient: Explicit instructions for refinement act as a "semantic gradient," guiding the model toward the correct interpretation more effectively than simple re-prompting.
• Cross-Lingual Knowledge Transfer: Forcing bilingual explanations allows the model to leverage its stronger language to correct "blind spots" in its weaker language.

Limitations and Conclusion

While highly effective, IBUT introduces significant computational overhead, increasing token consumption by approximately 5x compared to standard methods. This limits its current applicability for real-time, high-throughput systems. Future work will focus on optimizing the iterative loop and exploring long-context processing to reduce inference costs.

In summary, IBUT demonstrates that iterative bilingual alignment is a powerful tool for overcoming semantic distortions in machine translation, setting a new standard for context-aware LLM inference.

@inproceedings{chen2025llm,
  author = {Andong Chen and Kehai Chen and Yang Xiang and Xuefeng Bai and Muyun Yang and Yang Feng and Tiejun Zhao and Min Zhang},
  title = {{LLM-based Translation Inference with Iterative Bilingual Understanding}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}