🐉 ChineseIsEasy — Lexical & Multimodal Dataset

ChineseIsEasy-Dataset is the multimodal backbone of the ChineseIsEasy ecosystem. It is a massive, enriched Mandarin-French linguistic resource designed for Machine Learning, NLP research, and advanced Anki deck generation.

The dataset is fully managed via the Hugging Face datasets library, with binary files (images and audio) embedded directly into Parquet files for seamless integration.

---

🗂 Dataset Structure

The dataset is organized into two distinct configurations (subsets):

1. default (Lexicon & Single Words)

Contains primary lexical entries with frequency statistics, AI-generated semantic images, and word-level audio.

• Size: ~29,000 entries.
• Key Features: Word, hf_img_optim, hf_audio_word, Zipf frequency stats.

2. examples (Contextual Sentences)

Contains tens of thousands of example sentences linked to the lexicon, featuring full sentence-level audio.

• Size: ~44,000 sentences.
• Key Features: ch (Chinese), py (Pinyin), fr (French), audio (Audio binary).

---

Usage

Accessing the data is straightforward using the datasets Python library:

Load the main Lexicon (Words)

from datasets import load_dataset

# Load the default configuration
ds_words = load_dataset("AxelDlv00/ChineseIsEasy", "default", split="train")

entry = ds_words[0]
print(f"Word: {entry['Word']}")
# Images and audio are accessible directly as PIL objects or numpy arrays
# optimized_image = entry['hf_img_optim']

Load Example Sentences (with Audio)

# Load the examples configuration
ds_examples = load_dataset("AxelDlv00/ChineseIsEasy", "examples", split="train")

ex = ds_examples[0]
print(f"Sentence: {ex['ch']}")
# Play the audio binary directly
# audio_array = ex['audio']['array']

---

Generation Pipeline

1. Linguistic Enrichment: Batch processing via GPT-4o-mini for pedagogical categories and grammatical explanations.
2. Visual Semantics:

• LLM-driven prompt engineering.
• Local generation using Juggernaut XL v9 (SDXL) to create high-quality semantic anchors.

3. Audio Strategy:

• Words: Human recordings (CC-CEDICT-TTS) supplemented by gTTS fallbacks.
• Sentences: Synthesized using voxcpm with voice cloning from the ST-CMDS-20170001_1-OS corpus for natural diversity.

---

⚖️ License

• Dataset Content: Released under CC BY 4.0.
• Lexical Base: Derived from CC-CEDICT.
• Frequency Stats: Based on the SUBTLEX-CH corpus.

Author: Axel Delaval (陈安思)