MoonAlign

MoonAlign is a MoonBit toolkit for bilingual sentence alignment and offline translation evaluation.

It combines two practical pieces of text-engineering work:

• length-based sentence alignment in the style of Gale-Church
• corpus checks for translation quality with BLEU, ROUGE, and METEOR-style scoring

What it does

• Aligns parallel texts line by line with a deterministic dynamic program
• Scores candidate translations with BLEU-1 to BLEU-4, ROUGE-1, ROUGE-2, ROUGE-L, and METEOR-style exact matching
• Ships with a small CLI that can run a built-in demo or your own files

Input format

The first version keeps the workflow simple:

• one sentence per line
• blank lines are ignored
• token-based metrics expect whitespace-separated tokens

That makes the tool easy to use with already segmented English data, pre-tokenized Chinese data, or cleaned bilingual corpora.

Quick start

Run the built-in demo:

moon run --target native cmd/main --demo

Align two text blocks:

moon run cmd/main --src-text "Hello world" --tgt-text "Hello MoonBit"

Add evaluation text:

moon run cmd/main --ref-text "Hello world" --hyp-text "Hello MoonBit"

Read from files:

moon run --target native cmd/main --src data/demo/source.txt --tgt data/demo/target.txt
moon run --target native cmd/main --ref data/demo/reference.txt --hyp data/demo/hypothesis.txt

Demo files are also included under data/demo/ for a more realistic walkthrough.

Repository Links

• GitLink: https://gitlink.org.cn/Hzf78909/moonalign
• GitHub: https://github.com/hzfhzf89/MoonAlign

Review Notes

The repository keeps a single visible contributor and includes a concise competition note in docs/competition-notes.md.

Project structure

• moonalign.mbt: core alignment and evaluation logic
• cmd/main/main.mbt: thin CLI wiring
• moonalign_test.mbt: black-box tests for public behavior
• moonalign_wbtest.mbt: white-box tests for internal helpers

Competition notes

This repository is intended to satisfy the competition checklist:

• public source code
• readable README
• license
• commit history
• source description
• reusable MoonBit API surface

Design notes

The current implementation favors clarity over heavy dependencies:

• no external NLP package is required
• alignment is length-based and deterministic
• evaluation is self-contained and reproducible

This leaves room for future extensions such as:

• sentence splitting from raw paragraphs
• dictionary-assisted alignment
• better tokenization for Chinese and mixed scripts
• export to CSV or JSON
• alignment visualizations

Source note

The implementation is new code written for this repository.

The algorithmic design is inspired by:

• Gale and Church, A Program for Aligning Sentences in Bilingual Corpora
• standard BLEU, ROUGE, and METEOR definitions

No external source code was copied into the repository.

License

Apache-2.0