This repository contains a usable code from the paper:

G. Marra, A. Zugarini, S. Melacci, and M. Maggini, “An unsupervised character-aware neural approach to word and context representation learning,” in Proceedings of the 27th International Conference on Artificial Neural Networks – ICANN 2018

The structure of the project contains:

1. A data folder, containing the txt files from which to learn the embeddings.
2. A log folder, where the model is saved.
3. The char2word.py script, which is the main routine.
4. The encoder.py script, which contains some utility functions.

For a standard learning procedure do the following.

• Be sure both data and log folder are present.
• Put your training data into the data folder, with files named as data(something).txt, e.g. data01.txt.
• Simply run char2word.py

The script will create a vocabulary.txt file inside the data folder to be used during training. All the configurations are set to the default ones (i.e. the paper ones). The script does not yet provide a command-line configurations (apart for folder configuration, run --helpfor info). The user willing to have a custom configuration should modify the Config configuration class in the char2word.py script.

We will provide a more user-friendly command-line interface as soon as possible, together with more details about the training procedure and how to incorporate the model inside bigger models.

To see a fast way to exploit already trained embeddings look at the SentenceEncoder class together with the test function.

This code has been tested with tensorflow==1.4 and python2.7. Moreover, it has a dependency with the python library nltk .