philosopher-style-classification

This repo contains code and a short paper on cross-work authorship attribution for philosophical texts. I fine-tune a DistilBERT classifier to distinguish between Immanuel Kant vs Friedrich Nietzsche's writings and test whether performance persists under semantic control (topic modeling + embedding similarity). I also run a small LIME interpretability analysis to inspect token-level drivers of predictions.

What’s in here

• Data prep: Gutenberg cleaning + token chunking (125 tokens)
• Model: distilbert-base-uncased fine-tuned for binary classification
• Evaluation
  • H1: cross-work test (held-out books)
  • H2: semantic control
    • BERTopic topic overlap diagnostic + topic-controlled subset
    • cosine similarity control (opposite-author nearest neighbor)
• Interpretability: LIME explanations + aggregated feature plots/tables

Reproducibility (typical workflow)

1. Install dependencies
2. Run preprocessing (download/clean/chunk)
3. Train + evaluate (H1)
4. Run semantic controls (H2)
5. Run LIME analysis (H3)
6. Build figures/tables used in the paper

Note: Results depend on the exact corpus versions, random seeds, and filtering thresholds.

Note: Not all notebook cells are executed with saved outputs. Some experiments require GPU access, which may not be available by default on Colab or Kaggle; users are expected to rerun the notebooks in their own environment to reproduce full results.

Dependencies

• Python 3.8+
• transformers, datasets, torch
• pandas, numpy, matplotlib
• bertopic, sentence-transformers
• scikit-learn
• lime

Outputs

Key outputs are saved as:

• confusion matrix / metrics summaries
• topic diagnostics + topic-controlled subset metrics
• similarity curve figure
• LIME plots + summary tables