β†’ Back to Blog

What's Inside the SinoArabic Arabic-Chinese Corpus? A Technical Breakdown of 1.6M Words

A detailed technical walkthrough of the human-verified 1.6M Arabic + 717K Chinese word corpus. See dialect tags, intent flags, cultural annotation layers, and parameter integrity checks used for LLM training and game localization.

Introduction

The SinoArabic corpus is not just a collection of translated sentence pairs. It is a richly annotated, human-reviewed dataset designed specifically for Arabic-Chinese AI systems.

Corpus overview

  • Total size: 1.6M Arabic words + 717K Chinese words
  • Format: JSONL (one JSON object per line)
  • Alignment: Arabic β†’ Chinese (bidirectional)
  • Human verification: 100% reviewed by native speakers
  • Annotation layers: 7

Core structure: What each entry contains

Every line in the corpus is a JSON object with fields for ID, Arabic text, Chinese text, dialect tag, register, intent preserved, cultural note, param integrity, and reviewer score.

1. ID (Unique identifier)

Every entry has a unique ID like SA_001234 for tracking across versions.

2. ar_text (Arabic text)

The Arabic side of the pair. Can be formal (MSA) or dialectal (Gulf, Levantine, Egyptian).

3. zh_text (Chinese text)

The Chinese side of the pair. Simplified Chinese (Mandarin).

4. dialect_tag (Arabic dialect)

Specifies which Arabic dialect or register is used: msa, gulf, levantine, egyptian, or mixed.

5. register (Formality level)

Indicates the formality and context: formal, casual, gaming, or technical.

6. intent_preserved (Semantic equivalence)

Boolean flag for whether the Arabic and Chinese texts have equivalent meaning and intent.

7. cultural_note (Context and nuance)

Free-text annotation explaining cultural context, idioms, or translation challenges.

8. param_integrity (Parameter safety)

Boolean flag for whether dynamic parameters like %s, %d are preserved correctly.

9. reviewer_score (Quality confidence)

A 1β†’5 score from the human reviewer. For production use, filter to entries with score >= 4.

Distribution and statistics

By dialect: MSA 35%, Gulf 25%, Levantine 20%, Egyptian 15%, Mixed 5%. By register: Casual 40%, Formal 35%, Gaming 20%, Technical 5%.

How the corpus is structured for LLM training

The dataset is distributed as JSONL files compatible with Hugging Face datasets library, transformers, and most LLM fine-tuning frameworks.

Use cases

LLM fine-tuning (filter by dialect and register), game localization (use gaming register entries), and evaluation/benchmarking (use high-scoring entries).

Technical sample available on Hugging Face

A 153-row technical sample is publicly available with the same annotation structure. View the sample on Hugging Face.

Conclusion

The SinoArabic corpus is a richly annotated, production-ready dataset for Arabic-Chinese AI systems. Every entry is human-verified, dialect-tagged, and structured for easy filtering and integration.

β†’ Back to Blog