Translation Agent: Agentic Translation with LLM Reflection Workflow

Introduction

Translation Agent is a Python demonstration of an innovative agentic workflow for machine translation, developed by Andrew Ng and collaborators. This project leverages large language models (LLMs) to perform translations through a unique reflection process. The core idea involves an LLM first translating a text, then reflecting on its own output to identify areas for improvement, and finally using these self-generated suggestions to refine the translation.

This approach offers significant customizability, allowing users to easily modify the output's style (e.g., formal/informal), specify how to handle idioms and special terms, and even target specific regional language uses or dialects. While not yet mature software, Translation Agent is presented as a promising starting point for agentic translations, encouraging further discussion, experimentation, and open-source contributions in the field.

Installation

To get started with translation-agent, follow these steps. The Poetry package manager is required for installation.

1. Install Poetry:

   pip install poetry
   

   For more detailed instructions, refer to the Poetry Installation Guide.

2. Clone the repository and set up:

   A .env file with an OPENAI_API_KEY is required to run the workflow. See the .env.sample file in the repository for an example.

   git clone https://github.com/andrewyng/translation-agent.git
   cd translation-agent
   poetry install
   poetry shell # activates virtual environment
   

Examples

Once installed, you can use translation-agent in your Python projects:

import translation_agent as ta
source_lang, target_lang, country = "English", "Spanish", "Mexico"
source_text = "Hello, how are you?" # Example text
translation = ta.translate(source_lang, target_lang, source_text, country)
print(translation)

For a complete example script, refer to examples/example_script.py within the cloned repository.

Why Use It

Translation Agent presents a compelling alternative to traditional machine translation systems, primarily due to its high steerability and potential for superior translation quality in specific contexts. By integrating LLMs into a reflection workflow, it allows for granular control over translation nuances, such as stylistic choices, consistent handling of specialized terminology via glossaries, and adaptation to regional language variations.

Although its performance can vary, the project has shown instances of producing "fantastic results" that surpass commercial offerings. Furthermore, this agentic approach could serve as a valuable mechanism for automatically generating high-quality parallel text corpora, which can then be used to train and improve traditional, faster machine translation algorithms. The project actively invites the open-source community to explore extensions, including experimenting with different LLMs, developing glossary creation and usage strategies, and conducting more robust evaluations.

Links

• GitHub Repository: https://github.com/andrewyng/translation-agent
• License: Translation Agent is released under the MIT License.