Bertalign-Fast

An automatic multilingual sentence aligner optimized for CPU.

Overview

Bertalign-Fast is a lightweight, CPU-optimized version of Bertalign that uses modern SWE (Static Word Embeddings) instead of CWE (Contextualized Word Embeddings).

Key Features

• 🚀 Fast on CPU - No GPU required
• 💡 Lightweight - Use static word embeddings
• 🌍 Multilingual - Support 30+ languages
• 🎯 Accurate - Maintain high alignment quality
• 🖥️ User-friendly GUI - Visual interface for alignment tasks
• 🔄 Multi-version alignment - Align multiple language pairs or document versions

When to Use Which?

• Use Bertalign when:

  • You have GPU available
  • Maximum accuracy is critical
  • Processing time is not a constraint
  • Working with complex literary or highly nuanced texts

• Use Bertalign-Fast when:

  • You need fast CPU inference
  • Processing large volumes of text
  • Running on resource-constrained systems
  • Working with non-literary texts (news, technical documents, etc.)

Installation

git clone https://github.com/bfsujason/bertalign_fast.git

cd bertalign_fast

# Core only
pip install -r requirements.txt

# If you want the GUI
pip install pyqt5 igraph

# Download SWE model
python download_model.py

# Use mirror site if you cannot visit Hugging Face
python download_model.py --mirror hf-mirror.com

Quick Start

from bertalign_fast import BertalignFast

# Initialize aligner
aligner = BertalignFast()

# Load texts to be aligned
src_file = "data/demo/src/001.txt"
tgt_file = "data/demo/tgt/001.txt"
src_text = open(src_file, "rt", encoding="utf-8").read()
tgt_text = open(tgt_file, "rt", encoding="utf-8").read()

# Start aligning
aligner.align_sents(src_text, tgt_text)

# Print sentence indices and bead scores
print(aligner.alignment)

# Print aligned sentences
for src_sent, tgt_sent in aligner.bitext:
    print(f"{src_sent}\n{tgt_sent}\n")

Evaluation

Run on Intel Core i7-11800H CPU @ 2.30GHz, 16G RAM

Test Data

dataset	genre	language	# 1-1 alignment
gov	political	zh-ja	1574 (87.4%)
berg	yearbook	de-fr	678 (74.0%)
mac	literary	zh-en	2628 (59.8%)

python aligner_eval.py --dataset gov

python aligner_eval.py --dataset berg

python aligner_eval.py --dataset mac

Result

The running time includes both embedding and aligning.

dataset	precision	recall	F1	time
gov	0.986	0.989	0.987	13.20s
berg	0.918	0.923	0.921	4.79s
mac	0.867	0.895	0.881	19.49s

GUI Usage

Bertalign-Fast includes a graphical interface for alignment tasks.

The GUI was prototyped and refined using Claude Sonnet 4.5 as a coding assistant.

Launch

python aligner_gui.py

Features

• Automatic language detection with manual override
• Align one source against multiple targets simultaneously
• Interactive table editing: split, merge, delete, and add rows
• Undo/Redo support (Ctrl+Z / Ctrl+Y)
• Mark/Unmark rows with a distinct color for review
• Export to TMX, TSV, and JSON formats
• Project save/load

Editing Actions

Action	How
Edit cell	Double-click the cell
Mark rows	Select rows → right-click → Mark/Unmark
Split text	While editing → right-click → Split at cursor
Move text	While editing → right-click → Move Up/Down
Merge rows	Select rows → right-click → Merge

Demo: Multi-Language and Multi-Version Alignment

The data/demo directory contains a Chinese source file and three target files, which can be used to demonstrate two workflows:

Multi-language alignment: Align a Chinese source against both an English and a Polish human translation (zh-en-pl), producing a trilingual parallel corpus in one pass.

Multi-version alignment: Align a Chinese source against a human and a ChatGPT English translation side by side, useful for comparing translation quality across different versions of the same text.

To try it: load the source file, click "Add Target File" to add multiple target files, then click "Start Alignment."

Citation

If you use Bertalign-Fast in your research, please cite:

@software{Bertalign-Fast,
  author = {Lei Liu},
  title = {Bertalign-Fast: An Accessible Multilingual Sentence Aligner with CPU Optimization and Interactive Proofreading},
  year = {2026},
  url = {https://github.com/bfsujason/bertalign_fast}
}

Acknowledgments

• Hugging Face Static Embedding Models
• Vecalign
• Bleualign
• AIlign

License

This project is licensed under the GPL-3.0 License.

Support

Questions, bug reports, and feature requests are welcome. Please open an issue on GitHub.

Feedback on the following is especially appreciated:

• Language pairs: The evaluation only covers zh-ja, de-fr, and zh-en. If you test other pairs, we'd love to hear your results.

• Operating systems: The GUI has been tested on Windows. Reports from macOS and Linux users would be very helpful.