curioustauseef / deep-learning-annotation

AUEB MLCA - Deep Learning Text Classification

We are three ambitious students of the Master of Science in Business Analytics programme at AUEB, who are called upon to complete a text classification problem related to Sustainable Development Goals (SDGs).

The main idea is to train a Neural Network with pre-classified data and create an algorithmic model so that, by entering new data, we can achieve the greatest predictability in new and future publications.

Initially, in order to run either of the 2 deep learning notebooks you should produce the datasets. You have two options:

• You can directly download the two CSVs and run the corresponding classification notebooks, from this link (we already uploaded them on Github)
• You can download the parser_eda_extract.ipynb file, along with the research theme folder

The original raw data is included in the research theme folder. Before getting to the analysis, you should also unzip the documents folder in order for the parser_eda_extract.ipynb to perform correctly. Two csv files will be extracted, so that you can import them to their respective notebooks - one for binary and one for multiclass classification purposes. Notebooks are fully commented, however you can always look at the report for further explanations.

On the same note you should also download the heldout csv with the same options as before:

• You can directly download the CSV and run the corresponding classification notebook, from this link (we already uploaded it on Github)
• You can download the heldout_parser_eda_extract.ipynb file, along with the research theme folder

Tip: Run the jupyter notebooks in Google Colab to enhance performance

Keep in Mind

Due to the demanding time restrictions we had to make 2 compromises:

1. The metrics (as well as the Binary Saved Models files) of the binary classification deep learning models differ from those presented in the mlca-annotation-report.pdf. We had to re-calculate overall predictions (initially we calculated a prediction per model) on a heldout dataset at a later time, so we had to rerun the notebooks. The group1_research_theme_binary.csv contains 92 predictions based on our original best trained model (a CNN model with self-trained embeddings) at the time.

2. We planned to transfer the insights gained from our Binary Classification deep learning models (concerning the hyper-parameter variations) and proceed with Multiclass Classification. Alas, that was not possible and since we were not satisfied with the final notebook, the Multiclass Classification deep learning models are more limited in scope.

Enjoy!

• Lianos Alexandros
• Marinos Efstratios
• Lydia Papanikou