DocMTAgent

This repository releases the codes and data for the paper -- DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory.

DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory

📣 News

• [10/10/2024] Our code and dataset for DelTA is released!
• [11/10/2024] Our paper is published on arXiv: arXiv:2410.08143!

🔗 Quick Links

• About DelTA
• File Structure
• Requirements
• Quick Start
• Citation

🤖 About DelTA

DelTA, which is short for Document-levEL Translation Agent, is an online document-level translation agent based on multi-level memory. It consists of the following four memory components:

• Proper Noun Records: Maintain a repository of previously encountered proper nouns and their initial translations within the document, ensuring consistency by reusing the same translation for each subsequent occurrence of the same proper noun.
• Bilingual Summary: Contain summaries of both the source and target texts, capturing the core meanings and genre characteristics of the documents to enhance translation coherence.
• Long-Term Memory: Store contextual sentences in a wide span, from which reletive sentences will be retrieved while translating subsequent sentences.
• Short-Term Memory: Store contextual sentences in a narrow span, which will be utilized as demonstration exemplars while translating subsequent sentences.

The Framework of DelTA

📜 File Structure

Directory	Contents
data/	Experimental Data
eval_consistency/	Scripts of the LTCR-1 metric
infer/	Testing scripts
prompts/	Prompts for LLMs
results/	Testing outputs

🛠️ Requirements

DelTA with Qwen as backbone models is developed with HuggingFaces's transformers, DelTA with GPT as backbone models is developed with OpenAI API

• Python 3.9.19
• Pytorch 2.4.1+cu121
• transformers==4.45
• accelerate==0.34.2
• spacy==3.7.4
• numpy==2.0.2
• openai==1.51.2

🚀 Quick Start

Installation

git clone https://github.com/YutongWang1216/DocMTAgent.git
cd DocMTAgent
pip install -r requirements.txt

Inference with DelTA

(1) GPT as backbone models

• infer/run_infer_gpt.sh

Make sure to fill in the following parameters before running:

lang=en-zh                         # translation direction, choices=[en-zh,en-de,en-fr,en-ja,zh-en,de-en,fr-en,ja-en]
use_model=gpt35turbo               # GPT model, choices=[gpt35turbo,gpt4omini]
src=/path/to/src/file              # path to source document
ref=/path/to/ref/file              # path to reference document (optional, leave blank if not given)
export API_BASE=                   # base url of the API
export API_KEY=                    # API key

(2) Qwen as backbone models

• infer/run_infer_qwen.sh

Make sure to fill in the following parameters before running:

lang=en-zh                         # translation direction, choices=[en-zh,en-de,en-fr,en-ja,zh-en,de-en,fr-en,ja-en]
use_model=qwen2-7b-instruct        # GPT model, choices=[qwen2-7b-instruct,qwen2-72b-instruct]
modelpathroot=/path/to/checkpoint  # path to huggingface model checkpoint
src=/path/to/src/file              # path to source document
ref=/path/to/ref/file              # path to reference document

Calculating LTCR-1 metric scores

• eval_consistency/run_eval.sh

Make sure to fill in the following parameters before running:

lang=en-zh                         # translation direction, choices=[en-zh,en-de,en-fr,en-ja,zh-en,de-en,fr-en,ja-en]
src_file=/path/to/src/file         # path to source document (.src file generated by the inference script)
hyp_file=/path/to/hyp/file         # path to hypothesis document (.hyp file generated by the inference script)
output_dir=result/                 # output path of the evaluation results

📝 Citation

If you find this repo useful, please cite our paper as:

@misc{wang2024deltaonlinedocumentleveltranslation,
      title={DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory}, 
      author={Yutong Wang and Jiali Zeng and Xuebo Liu and Derek F. Wong and Fandong Meng and Jie Zhou and Min Zhang},
      year={2024},
      eprint={2410.08143},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.08143}, 
}