semARTagger

This repository contains the code for semARTagger, a pipeline developed as part of my Master's thesis in Cultural Data & AI at the University of Amsterdam. It automates the tagging of historical artworks based on semantic similarity, using multilingual sentence embeddings and a logistic regression classifier.

The tool was designed to process Dutch art exhibition catalogues and enrich them with descriptive tags, supporting art historical research and data structuring efforts at the RKD (Netherlands Institute for Art History).

---

📌 Project Description

The core pipeline uses LaBSE (Language-Agnostic BERT Sentence Embedding) to convert Dutch (and occasionally French) artwork titles into multilingual sentence embeddings. These embeddings are passed to a trained logistic regression classifier that predicts a controlled vocabulary of tags derived from RKD thesauri.

The interface allows users to review, edit, and export tags via a Streamlit web app, supporting both automated and manual modes of tagging.

---

📁 Project Structure

## 📁 Project Structure

semARTagger/
├── .github/
│   └── workflows/
│       └── keep-site-alive.yml        # GitHub Action to ping the app and create an alert if down
├── .streamlit/
│   └── config.toml                    # Streamlit app configuration (UI settings)
├── pages/
│   └── About.py                       # About page for Streamlit multipage app
├── .gitignore                         # Files and folders ignored by Git
├── Home.py                            # Main Streamlit app interface
├── LICENSE                            # Project license (MIT or similar)
├── README.md                          # Project overview and documentation
├── SUBJECT_all_terms_DUTCH.csv        # Subject vocabulary (Dutch) for tagging
├── SUBJECT_all_terms_ENGLISH.csv      # Subject vocabulary (English) for tagging
├── example_input.csv                  # Sample input CSV for testing
├── labse_label_binarizer.pkl          # Label binarizer for tag prediction decoding
├── labse_logreg_model.pkl             # Trained logistic regression model based on LaBSE
├── packages.txt                       # (Optional) Linux system dependencies (for deployment if needed)
├── pipeline.py                        # Backend tagging pipeline (used internally by the app)
├── requirements.txt                   # Python environment dependencies
└── rkd_aat_term_mapping.csv           # Mapping of RKD subject terms to AAT broader terms

---

How to Use

1. Clone the repository

git clone https://github.com/marnikzaf/semARTagger.git
cd semARTagger

2. Install dependencies

We recommend using Python 3.9+ and a virtual environment:

pip install -r requirements.txt

3. Launch the Streamlit interface

streamlit run Home.py

You’ll be able to upload data, run the tagging pipeline, review predicted tags, edit them, and export results.

---

Interface Overview

• Upload a CSV file with artwork titles (the file should have a column titled "Artwork").
• Run the automated tagging pipeline inside the app.
• Review, edit, and add tags for each artwork entry.
• Save progress and export edited tags as a new CSV file.
• Each session is protected by a password and can be saved and resumed later.

Example Input

An example input file is included under pipeline/example_input.csv:

Artist Name,Artwork
ABRAHAMS (Mej. Anna),Tulpen
ABRAHAMS (Mej. Anna),Pensées
AKKERINGA (Jan),In 't Duin
ALMA TADEMA (Mrs Laura),Well employed

This allows you to immediately test the pipeline without uploading your own data.

---

📄 Requirements

All required packages are listed in requirements.txt. Key libraries:

• sentence-transformers
• scikit-learn (via joblib)
• spacy
• streamlit

---

🎥 Demo Video

Watch a short demo of the tagging tool in action:

▶️ Watch the video

---

✉️ Contact

Maria Elpiniki Zafeiraki
LinkedIn | marnikzaf@gmail.com

---

📝 License

This project is provided for academic purposes. Please contact me if you intend to reuse or adapt it.