This project contains a Keras/TensorFlow Jupyter notebook which implements a DNN aimed at classifying the Wired.it Italian corpus.
The project also contains a Java tool which process the original corpus so that it is better suited to be read by the Python code in the notebook.
Please note that some branches are also available which contains alternative implementations of the notebook.
The notebook has been featured in the essay "What is the best method for Automatic Text Classification?": you should read the eassy in order to better understand the goals, the algorithms and the rationales.
Basic template for the Keras code has been borrowed from the Coursera course "Natural Language Processing in TensorFlow" held by Laurence Moroney.
My advice is to install Tensorflow (possibly with GPU support) and Jupyter via Anaconda:
conda create -n tensorflow_gpuenv tensorflow-gpu jupyterlab matplotlib scikit-learnconda activate tensorflow_gpuenvjupyter notebook, and then open the notebook from the browser web app
In case you want to try the talos branch, you need a different environment:
conda create --name talos_gpu scipy=1.2.0 matplotlib=2.2.3 h5py=2.8.0 tensorflow-gpu jupyter- this holds at the time of writing, due to some issues with dependencies
conda activate talos_gpupip install talosjupyter notebook, and then open the notebook from the browser web app
- clone on your system the project java-ml-text-utils, and execute
git checkout <version needed by this project pom.xml, e.g. v200>mvn clean install
- then, execute the class
com.ml_text_utils.shell.BuildWiredItFileSystemCorpusShell
WARNING: before creating a new corpus split, consider using the one located in corpora folder.