Gemini Book Translator 2.0: A High-Fidelity Literary Translation System Based on Agentic Workflow

Gemini Book Translator 2.0 is an open-source project for book translation using the Google Gemini API. It adopts an Agentic NLP pipeline architecture and solves the challenges of preserving the original style, emotional tone, and cultural connotations in literary translation through multi-stage collaboration and controlled LLM workflows, achieving high-quality and high-fidelity translation. It is suitable for scenarios that require maintaining the original style and context.

Machine translation technology has evolved from rule-based methods to statistical methods and then to neural networks. However, literary translation requires accurately conveying the literal meaning while preserving the original style, emotion, and cultural connotations, which has always been a difficult point for machine translation. Gemini Book Translator 2.0 proposes an innovative solution based on Agentic workflow to address this challenge.

Traditional machine translation is mostly an end-to-end single-stage model. The Agentic translation pipeline draws on the idea of software engineering pipelines, decomposing the translation task into multiple specialized stages, each handled by a dedicated 'agent'. Core advantages:

• Specialized division of labor: Different agents focus on specific tasks such as term extraction, initial translation, and polishing
• Controllable and interpretable: Outputs of each stage can be reviewed and intervened, facilitating debugging and optimization
• Quality accumulation: Multiple rounds of iterative feedback improve translation quality
• Flexible adaptation: Pipeline configuration can be adjusted according to text types

First Stage: Text Preprocessing and Structure Analysis

Identify text structure (chapters, paragraphs, etc.), extract key terms and cultural expressions, analyze register style, and establish a glossary and translation memory.

Second Stage: Initial Translation Generation

Generate initial translation based on the Gemini API, refer to the glossary, maintain the original structure, and mark uncertain translation points.

Third Stage: Style and Context Optimization

Adjust word order and sentence structure to conform to target language habits, maintain literary style and rhythm, handle cultural expressions such as metaphors and puns, and ensure consistent dialogue personalization.

Fourth Stage: Quality Check and Feedback

Evaluate the translation from accuracy, fluency, style consistency, and term consistency, and trigger feedback loops to correct problems if found.

In-Depth Use of Gemini API

• Large context window: Supports processing of entire chapters/books to maintain translation consistency
• Multilingual understanding: Gemini's multilingual capabilities lay the foundation for high-quality translation
• Controllable generation: Guide the model to output as expected through prompts

Modular Design

Each agent is developed and tested independently with clear input/output interfaces, supporting hot-swapping and replacement for easy community contributions.

Human-Machine Collaboration Mechanism

Supports manual review and modification of intermediate results, allows expert intervention in decision-making, and learns from human feedback to optimize strategies.

Literary Publishing

Accelerate the translation process to shorten cycles, provide high-quality first drafts to reduce manual workload, and maintain consistent style for series works.

Academic Research

Accurately handle professional terms and concepts, maintain academic rigor, and support the dissemination of multilingual academic achievements.

Personal Learning

Generate bilingual parallel versions, provide translation notes and cultural background explanations, and support personalized difficulty adjustment.

Dimension	Traditional Machine Translation	Gemini Book Translator 2.0
Translation Quality	Suitable for general texts, weak literary quality	Optimized for literary translation, better style preservation
Controllability	Black-box operation, difficult to intervene	Transparent pipeline, each stage can be reviewed
Consistency	Poor consistency for long texts	Uses large context window to maintain consistency
Customizability	Fixed model, hard to adjust	Modular design, supports customization
Human-Machine Collaboration	Limited	Comprehensive human-machine collaboration interface

Current Limitations

• Computational cost: High cost due to multi-stage LLM API calls
• Latency issue: Long time-consuming for complex text translation processes
• Cultural depth: Still limited for texts extremely dependent on cultural backgrounds

Future Directions

• Efficiency optimization: Reduce cost and time through caching and parallelization
• Multimodal expansion: Combine images and audio to support translation of illustrated books and audiobooks
• Personalized learning: Optimize translation strategies based on user feedback
• Community collaboration: Establish a sharing mechanism for translation memory and glossaries