windowtitles-translation-llm 0.1.0

This project aims to forward translate and backward translate window titles using the help of llama model.

• Forward Translation

  • Word-wise translation of input strings from any detected source language → English.
  • Optional token-level mapping (getTokenList) for round-trip tracking.
  • Preserve specified “protected” tokens (e.g. IDs, placeholders, rules).

• Preserving Sementics during Translation

  • Chunk-wise translation of input strings from any detected source language → English.
  • Merges common prepositional phrases, translates whole sentences when possible, then aligns individual parts for better context.
  • Uses BERT-based alignment to keep multi-word units intact.

• Backward Translation

  • Restoration of original text by inverting the mapping produced during forward translation.

---

📦 Installation

To install the package, run

pip install windowtitles-translation@git+https://github.com/paxray/windowtitles-translation-llm.git

to install without cloning the repository. If the repository is already cloned, running

pip install .

in the root folder also works.

Features

• Forward translation using Hugging Face MarianMT.
• Back-translation to reconstruct original window titles.
• Token mapping for granular translation control.
• Preserve placeholders and specified words.

Usage

Forward Translation

Use the getTranslations function from main.py:

from main import getTranslations

payload = {
    "windowTitles": ["Fenêtre de terminal", "Título de documento"],
    "preserveWordsList": ["Word"],
    "getTokenList": True,
    "windowTitlesLanguage": None,
    "translationType": "forward" 
}

translations = getTranslations(payload)
print(translations)

Output example:

{
  "Fenêtre de terminal": {
    "language": "fr",
    "translation": "Terminal window",
    "tokenMapping": { ... }
  },
  "Título de documento": {
    "language": "es",
    "translation": "Document title",
    "tokenMapping": { ... }
  }
} 

Backward Translation

Use the getOriginalData function to reverse a translation:

from main import getOriginalData

payload = {
    "alteredWindowTitles": {
        "alteredWindowTitle": { ... }
    },
    "translationType": "backward"
}

originals = getOriginalData(payload)
print(originals)

CLI

By default, running main.py performs a backward translation on sample data:

python main.py

Results are saved to data/output-data/output_file_name.json.

Configuration Options

Below are the configuration options for both forward and backward translation payloads. Each field’s datatype and whether it is mandatory are listed.

Forward Translation Payload:

• windowTitles (List[str], mandatory):
  • A list of strings representing the window titles to be translated.
• preserveWordsList (List[str], optional, default=[]):
  • Words or tokens you want to remain unchanged during translation.
• getTokenList (bool, optional, default=False):
  • If true, includes a mapping of source-to-target tokens in the output.
• windowTitlesLanguage (str, optional, default=None):
  • ISO 639-1 code (e.g., "fr", "es"); forces the source language detection.
• translationType (str, mandatory):
  • For forward translation use the keyword "forward"

Backward Translation Payload (fields within alteredWindowTitles entries):

• alteredWindowTitles (str, mandatory):
  • The translated window title you wish to revert.
• tokenMapping (Dict[int, int], mandatory):
  • Original-to-translated token index mapping used during forward translation.
• translationType (str, mandatory):
  • For backward translation use the keyword "backward"

Config.py

• device = "cpu" #or "cuda"
• Set to 'CPU' by default. Change to 'Cuda' to use GPU.

Project Structure

└── src
  └── windowtitles_translation
    ├── constants.py
    ├── config.py
    ├── commonMethods.py
    ├── backwardTranslation.py
    ├── forwardTranslation.py
    ├── forwardTranslationWithSemantics.py
    ├── main.py
└── data
    ├── input-data
    └── output-data