merriamtao/BiLit-CEFR

Balanced Bilingual Literary Complexity Dataset (EN + ZH)

Description

This dataset contains paragraph-level literary segments labeled with CEFR-style difficulty levels (A/B/C).

Languages:

• English (CommonLit-based scoring)
• Chinese (heuristic labeling)

Data Fields

• text: paragraph
• difficulty: A/B/C
• language: en/zh
• open_source_books: original source

Labeling Method

• English: CommonLit readability model,syntactic complexity, the proportion of rare vocabulary and stylistic modification
• Chinese: heuristic rules (lexical hierarchy and syntactic structure difficulty) Sourcecode avaliable under file/src directory.

Dataset Statistics

English Subset

The English dataset contains approximately 1,026 literary segments with an imbalanced CEFR-style distribution:

• A: 242 samples
• B: 354 samples
• C: 430 samples

Train/test split:

• en_train.jsonl: 820 samples
• en_test.jsonl: 206 samples

Chinese Subset

The Chinese dataset contains 600 literary segments with a balanced CEFR-style distribution:

• A: 200 samples
• B: 200 samples
• C: 200 samples

Train/test split:

• zh_train.jsonl: 480 samples
• zh_test.jsonl: 120 samples

Source Literature

The dataset contains sampled literary excerpts from 39 English Literature books and 16 Chinese literature books rather than complete books.

English Sources

This dataset incorporates subsets from literature by ACOSharma https://huggingface.co/datasets/ACOSharma/literature and works from Gutenburg Project to balance the distribution of labels. Example source works include:

• Ulysses
• Wuthering Heights
• Treasure Island
• The War of the Worlds
• Metamorphosis
• A Room with a View
• McGuffey_First_Reader
• Mother_Goose_Rhymes
• The_Tale_of_Peter_Rabbit

Chinese Sources

Example source works include:

• 红楼梦
• 西游记
• 水浒传
• 三国演义
• 官场现形记
• 呐喊
• 背影
• 史记
• 狂人日记
• 朝花夕拾
• 骆驼祥子
• 寄小读者

Only short segmented excerpts are included for research and educational purposes. The dataset does not contain complete original works.

Combined Dataset

The final multilingual dataset combines English and Chinese literary excerpts for CEFR-style readability classification and multilingual text difficulty modeling.

Use Cases

• Text classification
• Readability estimation
• Multilingual NLP

Limitations

• Chinese labels are heuristic
• Literary bias