Gothic Thematic Analysis Engine

An NLP system for analyzing Gothic literature for thematic content and generating new passages with controllable themes.

Status

Phase 1 (Foundation): Complete Phase 2 (Analyzer): Complete - classifier trained, corpus auto-tagged Phase 3 (Generator): In progress - fine-tuning

Data

• Corpus: 38 Gothic novels (1764-1911)
• Passages: ~11,700 chunks (100-500 tokens each)
• Splits: Train 8,906 / Val 1,619 / Test 1,182

Theme Taxonomy

10 themes derived from corpus sampling:

Theme	Description
supernatural	Vampires, ghosts, apparitions, the undead
captivity	Physical captivity, confinement, being chased, escape attempts
secrecy	Hidden information, concealed identities
love	Romantic attachment, courtship, passion
harm	Death, dying, murder, violence, blood, physical brutality
villainy	Scheming, treachery, persecution, antagonistic behavior
family	Parent-child bonds, inheritance, lineage
sorrow	Sadness, despair, loneliness, isolation
anguish	Terror, horror, dread, guilt, insanity, psychological torment
setting	Atmospheric environments—storms, wilderness, Gothic architecture

Labeling Tool

Web UI for manually tagging passages with themes.

Setup

cd gothic-thematic-analysis-engine
python -m venv venv
source venv/bin/activate
pip install -e ".[labeler]"

Run

uvicorn labeler.app:app --reload
# Open http://localhost:8000

Keyboard Shortcuts

• 1-9, 0, -, =, Q - Toggle themes
• Enter - Submit and next
• S - Skip passage
• U - Undo last label

Project Structure

gothic-thematic-analysis-engine/
├── data/
│   ├── gothic_central_novels/   # Source texts (38 novels)
│   ├── chunks/                  # Per-novel chunk files
│   └── passages/                # Train/val/test splits
├── labeler/
│   ├── app.py                   # FastAPI server
│   ├── data/
│   │   ├── passages.json        # Passages for labeling
│   │   └── labels.json          # Labeling progress
│   ├── templates/
│   └── static/
├── scripts/
│   ├── chunk_corpus.py          # Novel chunking
│   └── split_data.py            # Train/val/test splitting
├── theme_taxonomy.yaml
└── pyproject.toml

Design Decisions

Training Data Size

I labeled 353 passages manually instead of the initially planned 550. At this point, the classifier achieved ~60% recall at a 0.25 threshold, with all 10 themes represented in the training data.

The classifier's purpose is auto-tagging the corpus for generator training, not production classification. With 11,000+ passages being tagged, patterns emerge from volume—minor labeling noise gets averaged out during fine-tuning.

This was a deliberate "good enough to test the pipeline" decision. If generator output quality is poor, I can return to labeling, retrain the classifier, and regenerate the tagged corpus without losing progress.

Multi-Machine Training

Generator fine-tuning was attempted across multiple machines based on available hardware:

Machine	Specs	Use Case
M1 Mac Mini	8GB RAM, MPS	Initial attempts—ran out of GPU memory on larger models
ThinkPad T14s	Snapdragon X, 32GB RAM	CPU training—worked but slow (30+ hours estimated)
Desktop	Ryzen 7 1700X, GTX 1050 Ti 4GB	Final training—GPU acceleration, ~18 hours

The GTX 1050 Ti is old but sufficient for Pythia-160M with small batch sizes. Training completed overnight.

---

Multi-Machine Workflow

Keeping Code in Sync

Use Git. Push before switching machines, pull after.

# Before leaving a machine
git add .
git commit -m "WIP: training in progress"
git push

# On the new machine
git pull

Handling Large Files

Models and datasets shouldn't live in Git. Use .gitignore:

models/
data/chunks
data/generator_training.txt
data/tagged_corpus.json
*.pkl

Transfer large files manually:

• USB drive
• Cloud storage (Google Drive, Dropbox)

Environment Consistency

Each machine needs a different setup:

Machine	PyTorch Install
Mac M1	pip install torch (MPS auto-detected)
CPU-only	pip install torch (defaults to CPU)
NVIDIA GPU	pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126

Keep a pyproject.toml for non-torch dependencies. Install torch manually per machine.