This is an NLP-powered system for analyzing the Black Clover anime. It brings together web scraping, theme extraction, character networks, text classification, and an LLM-based character chatbot — all accessible through a Gradio app.
- 📥 Scrapes character dialogue and episode summaries using Scrapy
- 🧾 Extracts themes from episode transcripts using zero-shot classification
- 👥 Builds a character network from named entity recognition and graph analysis
- 🧪 Trains a custom text classifier to label quotes or scenes
- 🤖 Simulates conversations with Black Clover characters using an LLM
- 🖥️ Runs in a browser via a Gradio web app
.
├── crawler/ # Scrapes Black Clover episode/character data
├── character_network/ # spaCy + NetworkX + PyVis for visual character mapping
├── text_classifier/ # Hugging Face-based custom classifier
├── theme_classifier/ # Zero-shot theme detector
├── character_chat_bot/ # LLM-powered character chatbot
├── gradio_app.py # Runs the full system in an interactive Gradio UI
├── requirements.txt
└── README.md
- Python
- Scrapy – web scraping
- spaCy – named entity recognition
- NetworkX + PyVis – character graph visualization
- Hugging Face Transformers – zero-shot and fine-tuned models
- Gradio – web interface
- PyTorch – model backend
This project includes a custom instruct-tuned model published on Hugging Face:
➡️ Clover_Llama-3.1-8B-Instruct
The model was designed to simulate characters from Black Clover based on structured subtitles and script data. It powers the chatbot and some of the classification features.
- View a relationship graph of major characters in the show
- Classify new quotes by tone, character, or plot category
- Extract themes across arcs (e.g. loyalty, rivalry)
- Chat with characters like Asta, Yuno, or Noelle using real lines and modeled personas
A look at the main Gradio web interface that ties all components together.
Example interaction with the character chatbot built using LLMs and dialogue from the show.
An animated walkthrough showing how the web app processes input and delivers results.
This is a snippet of the raw subtitle file used as input for the NLP pipeline. These lines are cleaned and parsed before being passed to the classifier and NER modules.
A snapshot of the NER results extracted from the subtitle data. This output feeds into the character network builder to visualize relationships and character prominence across episodes.
Example rows from the theme classifier's output. Each episode is analyzed using a zero-shot model to determine the dominant themes, along with confidence scores.
