README.md

---
language:
- ja
tags:
- japanese
- text-difficulty
- language-learning
- linguistics
- aozora-bunko
- educational
- curriculum-design
task_categories:
- text-classification
size_categories:
- 1K<n<10K
license: mit
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: text
    dtype: string
  - name: overall_difficulty
    dtype: float64
  - name: difficulty_level
    dtype: string
  - name: kanji_difficulty
    dtype: float64
  - name: lexical_difficulty
    dtype: float64
  - name: grammar_complexity
    dtype: float64
  - name: sentence_complexity
    dtype: float64
  - name: text_length
    dtype: int64
  splits:
  - name: train
    num_bytes: 9611892
    num_examples: 1800
  download_size: 5523038
  dataset_size: 9611892
---

# Aozora Text Difficulty Dataset

This dataset contains Japanese literary texts from the [Aozora Bunko](https://www.aozora.gr.jp/) digital library, enhanced with **jReadability-based difficulty analysis** for Japanese language learning and curriculum development.

## Dataset Overview

- **Source**: Aozora Bunko (青空文庫) - Japan's premier digital library of public domain literature
- **Enhancement**: jReadability-based difficulty scoring using research-backed Japanese readability models
- **Primary Methodology**: [jReadability](https://github.com/joshdavham/jreadability) - A Python implementation of Lee & Hasebe's Japanese readability evaluation system
- **Use Cases**: Japanese language curriculum design, reading level assessment, adaptive learning systems, difficulty-controlled text generation
- **License**: Original Aozora Bunko texts are public domain; analysis code and scores are provided under open source terms

## 📊 Dataset Structure

**Total Records**: 5,000 Japanese texts
**Total Columns**: 21
**Column Categories**: Original Data (3) + Core Difficulty Scores (6) + Detailed Metrics (11) + Legacy Score (1)

---

## 🗂️ Column Descriptions

### **Original Data Columns (3)**

| Column | Type | Description |
|--------|------|-------------|
| **`text`** | string | Full Japanese text content from Aozora Bunko (50-532,561 characters) |
| **`footnote`** | string | Publishing information and bibliographic details in Japanese |
| **`meta`** | string | JSON metadata with work ID, title, author, and readings |

### **Core Difficulty Scores (6) - Main Features for Learning**

| Column | Type | Range | Description |
|--------|------|-------|-------------|
| **`overall_difficulty`** | float64 | 0.0-1.0 | **Primary difficulty score** based on jReadability model |
| **`kanji_difficulty`** | float64 | 0.0-1.0 | Complexity based on kanji grade levels and density |
| **`lexical_difficulty`** | float64 | 0.0-1.0 | Vocabulary complexity using authentic frequency data |
| **`grammar_complexity`** | float64 | 0.0-1.0 | Grammatical structure complexity |
| **`sentence_complexity`** | float64 | 0.0-1.0 | Sentence length and structure variation |
| **`difficulty_level`** | string | categorical | **Curriculum classification**: Beginner/Elementary/Intermediate/Advanced/Expert |

### **Detailed Linguistic Metrics (11)**

| Column | Type | Description |
|--------|------|-------------|
| **`text_length`** | int64 | Total character count including punctuation |
| **`kanji_density`** | float64 | Proportion of Chinese characters (0.0-1.0) |
| **`avg_sentence_length`** | float64 | Average characters per sentence |
| **`joyo_grade_avg`** | float64 | Average educational grade of kanji used (1-9 scale) |
| **`lexical_diversity`** | float64 | Unique words ÷ total words (vocabulary richness) |
| **`non_joyo_percentage`** | float64 | Proportion of advanced kanji beyond standard education |
| **`avg_word_length`** | float64 | Average characters per word |
| **`katakana_percentage`** | float64 | Proportion of katakana (foreign/technical words) |
| **`word_frequency_score`** | float64 | Vocabulary rarity score (0=common, 1=rare) |
| **`sentence_length_variance`** | float64 | Statistical variance in sentence lengths |
| **`grammar_complexity_score`** | float64 | Grammatical pattern complexity score |

### **Legacy Compatibility (1)**

| Column | Type | Range | Description |
|--------|------|-------|-------------|
| **`difficulty_score`** | float64 | 0.0-10.0 | Traditional 10-point difficulty scale for compatibility |


---

## 🎯 Difficulty Calculation Methodology

### **Primary Difficulty Score: jReadability Model**

The `overall_difficulty` score is calculated using the [jReadability](https://github.com/joshdavham/jreadability) Python library, which implements the research-backed Japanese readability model developed by **Jae-ho Lee and Yoichiro Hasebe**.

**jReadability Model Formula:**
```
readability = {mean words per sentence} × -0.056
            + {percentage of kango} × -0.126      # Chinese-origin words
            + {percentage of wago} × -0.042       # Native Japanese words
            + {percentage of verbs} × -0.145
            + {percentage of particles} × -0.044
            + 11.724
```

**Score Normalization:**
- jReadability output: 0.5-6.5 (higher = easier)
- Our normalization: `(6.5 - jreadability_score) / 6.0` → 0-1 scale (higher = harder)

### **Curriculum Level Classification**
- **Beginner** (0.00-0.19): Basic modern Japanese
- **Elementary** (0.20-0.34): Simple literary texts
- **Intermediate** (0.35-0.54): Standard literary works
- **Advanced** (0.55-0.74): Complex literary language
- **Expert** (0.75-1.00): Classical or highly sophisticated texts

### **Supporting Linguistic Metrics**
1. **Kanji Analysis**: 3,003 kanji with official educational grades from [kanjiapi.dev](https://kanjiapi.dev)
2. **Vocabulary Analysis**: [wordfreq](https://github.com/rspeer/wordfreq) library with real corpus data
3. **Grammar Analysis**: Pattern-based complexity scoring using formal Japanese constructions
4. **Sentence Analysis**: Length variation and structural complexity measures

### **Research Foundation**
- **Lee, J. & Hasebe, Y.** *Introducing a readability evaluation system for Japanese language education*
- **Lee, J. & Hasebe, Y.** *Readability measurement of Japanese texts based on levelled corpora*
- Model specifically designed for **non-native Japanese learners** (not native speaker grade levels)

---

## 📈 Dataset Statistics

**jReadability-Based Analysis Results (5,000 texts):**
- **Overall Difficulty**: Mean 0.547 (0.0-1.0 scale)
- **Difficulty Distribution**:
  - Beginner: 97 texts (1.9%)
  - Elementary: 552 texts (11.0%)
  - Intermediate: 2,047 texts (40.9%)
  - Advanced: 1,690 texts (33.8%)
  - Expert: 614 texts (12.3%)

**Text Characteristics:**
- **Text Length**: 50 - 532,561 characters (mean: ~11,285)
- **Kanji Density**: 29.4% average
- **Average Joyo Grade**: 3.71 (elementary-intermediate level)
- **Lexical Diversity**: 0.270 (moderate vocabulary variation)

---

## 🔬 jReadability Advantages

### **Why jReadability?**
1. **Research-Backed**: Based on empirical studies of Japanese learner corpora
2. **Learner-Focused**: Designed specifically for non-native Japanese speakers
3. **Linguistic Sophistication**: Considers Japanese-specific features (kango/wago ratios, particle usage)
4. **Reproducible**: Standardized implementation with consistent results
5. **Validated**: Published research with proven correlation to learner difficulty perception

### **Improvements Over Composite Scoring**
- **Holistic Assessment**: Considers text as a unified linguistic entity rather than separate features
- **Native Speaker Bias Reduction**: Avoids assumptions based on native speaker intuitions
- **Empirical Foundation**: Based on actual learner performance data
- **Standardized Scale**: Consistent 6-level difficulty assessment widely used in Japanese education

---

## 💻 Usage Examples

### **Loading the Dataset**
```python
from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("ronantakizawa/aozora-text-difficulty")
train_data = dataset['train']
```

### **Filtering by Difficulty Level**
```python
import pandas as pd

df = train_data.to_pandas()

# Get beginner-level texts
beginner_texts = df[df['difficulty_level'] == 'Beginner']

# Get texts within specific difficulty range
intermediate = df[
    (df['overall_difficulty'] >= 0.35) &
    (df['overall_difficulty'] < 0.55)
]

# Filter by kanji difficulty for kanji learning
easy_kanji = df[df['kanji_difficulty'] < 0.3]
```

### **Analyzing Text Metrics**
```python
# Examine vocabulary complexity
rare_vocab = df[df['word_frequency_score'] > 0.7]

# Find texts with specific sentence patterns
short_sentences = df[df['avg_sentence_length'] < 30]

# Analyze kanji grade distribution
elementary_kanji = df[df['joyo_grade_avg'] <= 4.0]
```

---

## 🎓 Applications

- **Language Learning**: Personalized reading recommendations based on learner level
- **Curriculum Design**: Structured progression of reading materials using research-backed difficulty assessment
- **Assessment Tools**: Automatic text difficulty evaluation for placement using jReadability standards
- **Research**: Japanese language complexity and readability analysis with validated metrics
- **EdTech**: Adaptive learning system development and content curation
- **Text Generation**: Difficulty-controlled generation of Japanese educational content
- **Reading Comprehension**: Graded text selection for language learning platforms

---

## 🛠️ Technical Implementation

### **jReadability Integration**
```python
from jreadability import compute_readability

# Calculate jReadability score
jreadability_score = compute_readability(japanese_text)

# Normalize to 0-1 difficulty scale (0=easiest, 1=hardest)
overall_difficulty = max(0.0, min(1.0, (6.5 - jreadability_score) / 6.0))
```

### **Batch Processing**
For optimal performance when processing large datasets:
```python
from fugashi import Tagger
from jreadability import compute_readability

# Initialize tagger once for batch processing
tagger = Tagger()

# Process multiple texts efficiently
for text in texts:
    score = compute_readability(text, tagger)  # Reuse tagger
```

### **Dependencies**
- **jreadability**: Research-backed Japanese readability calculation
- **fugashi**: Fast Japanese morphological analysis (MeCab wrapper)
- **unidic-lite**: Japanese linguistic resources
- **wordfreq**: Authentic Japanese word frequency data

---

## 🔗 Related Datasets

- [Japanese Character Difficulty Dataset](https://huggingface.co/datasets/ronantakizawa/japanese-character-difficulty) - Kanji grades used in this analysis
- [jReadability GitHub](https://github.com/joshdavham/jreadability) - Original jReadability implementation

---

## 📚 Acknowledgments

- **Aozora Bunko** for providing the foundational literary corpus
- **kanjiapi.dev** for comprehensive kanji educational data
- **wordfreq project** for authentic Japanese frequency data

---

## 📄 Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{aozora_text_difficulty_2024,
  title={Aozora Text Difficulty Dataset},
  author={Claude Code Analysis},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ronantakizawa/aozora-text-difficulty}
}
```