DRT-o1 - A series of AI models for literary translation developed by Tencent Research Institute
DRT-o1 is a series of AI models developed by Tencent Research Institute that significantly improves the translation quality of literary works through Long Chain Thinking (CoT) technology, especially when dealing with rhetorical devices such as metaphors and allegories. The model can understand...
What is DRT-o1?
DRT-o1 is a series of AI models developed by Tencent Research Institute that significantly improves the translation quality of literary works through Long Chain Thinking (CoT) technology, especially in handling rhetorical devices such as metaphors and allegories. The model can understand the text more deeply, preserving the emotional tone and cultural connotations of the original. The DRT-o1 series includes two sizes: DRT-o1-7B and DRT-o1-14B. The BLEU score improved by 7.33 to 8.26, and the CometScore improved by 1.66 to 3.36. DRT-o1-7B outperforms the larger model QwQ-32B, demonstrating its powerful ability to handle complex language structures.
Main functions of DRT-o1
- Long Chain Thinking and Reasoning (CoT)DRT-o1 improves the ability to understand rhetorical devices such as metaphors and allegories in literary works through long thought chain technology, thereby enhancing translation quality.
- Multi-agent frameworkThe model employs a multi-agent framework comprising three roles: translator, consultant, and evaluator. The translator is responsible for the initial translation, the consultant provides suggestions for revision, and the evaluator scores the translation quality based on preset metrics.
- Translation WorkflowThe workflow consists of three main steps: keyword translation, initial translation, and a translation refinement cycle. In the refinement cycle, consultants evaluate the translation from the previous step and provide feedback. Evaluators give an overall score based on predefined scoring criteria, and translators provide new translations based on the feedback and score.
- Iterative optimizationThrough repeated iterations, the model continuously improves translation quality until it reaches a predefined threshold or the number of iterations reaches its maximum value.
- Performance improvementThe experimental results show that both the BLEU score and the CometScore were significantly improved.
- Handling complex language structuresDRT-o1 performs exceptionally well in handling complex language structures, outperforming the larger model QwQ-32B.
DRT-o1 Technical Principles
- Deep Thinking TrainingResearchers selected 400 public domain English books from Project Gutenberg, extracted 577,600 sentences, and selected 63,000 sentences containing similes and metaphors to train a model for "deep thinking".
- ExplainabilityBecause CoT provides visibility into the reasoning process, it helps improve the interpretability of model decisions.
- Logical reasoning and context utilizationCoT can help models perform complex logical reasoning, using contextual information to solve problems through step-by-step reasoning.
DRT-o1 project address
- Github repository:https://github.com/krystalan/DRT-o1
- HuggingFace model library:
- DRT-o1-7B:https://huggingface.co/Krystalan/DRT-o1-7B
- DRT-o1-14B:https://huggingface.co/Krystalan/DRT-o1-14B
- arXiv technical paper:https://arxiv.org/pdf/2412.17498
Application scenarios of DRT-o1
- Literary translationThe DRT-o1 model is particularly suitable for translating literary works, especially texts rich in cultural and contextual meanings. Through Long Chain Thinking (CoT) technology, it can more accurately convey the original text's connotations and emotional nuances.
- Cross-cultural communicationDRT-o1 can effectively handle cultural differences and has the potential to be applied in cross-cultural communication, helping people from different language and cultural backgrounds to better understand and appreciate foreign literary works.
- EducationIn the field of education, DRT-o1 can be used as an auxiliary tool to help students understand the complex rhetorical devices in foreign literary works and improve their language learning and cultural understanding abilities.
- Multilingual content creationThe DRT-o1 supports cross-language writing and translation, facilitating global dissemination and making it suitable for content creators to create and localize multilingual content.
- Customer support and personalized servicesThe DRT-o1 can be used to build intelligent and efficient customer service systems, providing personalized experiences, especially for customer support in multilingual environments.