Multi-Model agentic & reflective translation workflow in Aamzon Bedrock

Inspired by andrewyng/translation-agent，enabling it run in Amazon Bedrock

A reflection agentic workflow for machine translation. The main step are:

1. Prompt an LLM to translation a text from source_lang to target_lang
2. Have the LLM reflect on the translation to come up with constructive suggestions for improving it;
3. Use the suggestions to improve the translation.

Key features

With the support of Amazon Bedrock for multiple models, different models can be used in different steps to fully leverage the advantages of LLM.

Getting Started

Following below steps to get started with bedrock-translation-agent:

Installation:

• Local AWS cerdentials are required. In Amazon EC2, instance profile is the best option; and if in your local environment, your might need to configure the configuration file or environment variables.

• Python package manage tool Poetry is required, you can follow the installation guide to install it. Depending on your local environment, this might work:

curl -sSL https://install.python-poetry.org | python3 -

• After checking out the code, use Poetry to install dependencies:

poetry install

Usage:

• Using default model:

from bedrock_translation_agent.libs.translation import Translation

translation = Translation(
    source_text="Translation text here.",
    source_lang="English",
    target_lang="Chinese",
    country="China"
)

print(translation.translate())

• Using different models for each step:

The models supported in Amazon Bedrock are defined inlibs/bedrock_models.py, and you can use the method Translation.set_models to declare the model for each step of translation.

Note: The code uses system_prompt and user_prompt. If the model you specified does not support system prompt, you need to modify the API parameters in libs/bedrock.py

from bedrock_translation_agent.libs.bedrock_model import BedrockModel
from bedrock_translation_agent.libs.translation import Translation

translation = Translation(
    source_text=long_text,
    source_lang="English",
    target_lang="Chinese",
    country="China"
).set_models(
    init_model=BedrockModel.CLAUDE_3_SONNET_1_0,
    reflect_on_model=BedrockModel.CLAUDE_3_SONNET_1_0,
    improve_model=BedrockModel.CLAUDE_3_SONNET_1_0,
)

print(translation.translate())

Related work

A few academic research groups are also starting to look at LLM-based and agentic translation. We think it’s early days for this field!

• ChatGPT MT: Competitive for High- (but not Low-) Resource Languages, Robinson et al. (2023), https://arxiv.org/pdf/2309.07423
• How to Design Translation Prompts for ChatGPT: An Empirical Study, Gao et al. (2023), https://arxiv.org/pdf/2304.02182v2
• Beyond Human Translation: Harnessing Multi-Agent Collaboration for Translating Ultra-Long Literary Texts, Wu et al. (2024), https://arxiv.org/pdf/2405.11804

License

This library is licensed under the MIT-0 License. See the LICENSE file.