Classify the Roman ruler on a coin based on the coin's description

Data was taken from eBay Ancient Roman Coin listings. Some of them are tagged with the aspect of a Ruler and some are not. For future usage I needed to fill the gaps for the entries that do not have the ruler specified.

Using CNN classification I was able to reach 95% accuracy.

Based on https://github.com/jiegzhan/multi-class-text-classification-cnn

Why Neural Networks?

From most entries you could deduct the ruler by a simple RegEx. Problem comes with ±10% of the entries where the names are not included in the description at the expected format

For example the entry 'FL IVL CONSTANTIVS NOB C - GLORIA EXERCITVS 334-335 AD' stands for Constantine II. No reasonable RegEx could deudct that, but the CNN had no problem.

Data:

• Input: description

  • Example: "FL IVL CONSTANTIUS NOB C - GLORIA EXERCITVS 334-335 AD"

• Output: ruler

  • Example: Constantine II

Train:

• Command: python3 train.py training_data.file parameters.json
• Example: python3 train.py ./data/title_ruler.csv ./parameters.json

A directory will be created during training, and the trained model will be saved in this directory.

Predict:

Provide the model directory (created when running train.py) and new data to predict.py.

• Command: python3 predict.py ./trained_model_directory/ new_data.file
• Example: python3 predict.py ./trained_model_1521137138/ ./data/small_samples.json

Reference:

• Implement a cnn for text classification in tensorflow