Seisyun Information Entropy

This repository contains the implementation and data for quantifying the concept of "Seisyun" (青春), a term widely used in Japan to describe youth, through a novel measure called Seisyun Information Entropy. The method combines advanced natural language processing techniques and large language models (LLMs) to establish a quantitative framework for analyzing and understanding this abstract concept.

Overview

The study introduces a mathematical and computational framework to measure the "youthfulness" of a text by considering three key factors:

1. Unusualness: Using the probability of word predictions from pre-trained BERT models to calculate unexpectedness in text.
2. Positivity: Sentiment analysis to evaluate the positive nature of the text.
3. Fluency: Evaluating the grammatical correctness of the text.

These metrics are integrated into a single measure, Seisyun Information Entropy, using information theory principles.

Formula

The Seisyun Information Entropy for a given text (S) is defined as:

I_{adolescence}(S) = -\log_2 P_{unusual}(S) - \log_2(1 - P_{positive}(S)) - \log_2(1 - P_{fluency}(S))

Where:

• $P_{unusual}(S)$: The average word prediction probability.
• $P_{positive}(S)$: The probability that the text is classified as positive.
• $P_{fluency}(S)$: The probability that the text is grammatically correct.

Methodology

The framework leverages multiple pre-trained BERT models:

1. Unusualness:

   • Tohoku BERT (Japanese)
   • Predicts word probabilities to assess unexpectedness.

2. Positivity:

   • Sentiment-Enhanced BERT
   • Analyzes sentiment polarity.

3. Fluency:

   • Fluency-Scoring BERT
   • Evaluates grammatical correctness.

Repository Structure

• module/: Contains the core modules for calculating Seisyun Information Entropy.
• main.py: The main script to execute the entropy calculations.
• seisyun.csv: Dataset containing text samples related to "Seisyun".
• goodness.csv: Dataset with human-annotated "youthfulness" scores.
• comparison.csv: Results comparing calculated entropy with human scores.
• plot.png: Visualization of the comparison results.
• requirement.txt: List of dependencies required to run the project.
• .gitignore: Specifies files and directories to be ignored by git.
• LICENSE: License information for the repository.

Getting Started

Prerequisites

• Python 3.8+
• PyTorch
• Hugging Face Transformers

Install dependencies:

pip install -r requirement.txt

Usage

For testing the result of thesis, run:

python main.py 

Results

Experimental results demonstrated a weak positive correlation ((r = 0.333)) between Seisyun Information Entropy and manually generated "youthfulness" rankings of texts. The significance test ((p = 0.035)) confirmed the statistical validity of the measure at a 5% significance level.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Document Availability

The complete manuscript is accessible online at the following address:

https://web.cshe.nagoya-u.ac.jp/support/student/contest/img/2024_kotama.pdf

References

If you use this work, please cite:

@unpublished{kotama:seisyun_entropy_llm,
  author        = {Takanori Kotama},
  title         = {A Quantitative Definition of Youth Using Large Language Models},
  year          = {2025},
  note          = {Award-winning paper of the Nagoya University Student Paper Contest (Encouragement Prize), presented in 2025},
  url           = {https://web.cshe.nagoya-u.ac.jp/support/student/contest/img/2024_kotama.pdf}
}