News classification task for the Human Language Technologies course (spring semester 2023-2024).

• Fabrizio De Castelli   - M.Sc. in Artificial Intelligence, University of Pisa
• Francesco Aliprandi   - M.Sc. in Artificial Intelligence, University of Pisa
• Francesco Simonetti  - M.Sc. in Artificial Intelligence, University of Pisa
• Marco Minniti             - M.Sc. in Artificial Intelligence, University of Pisa
• Tommaso Di Riccio    - M.Sc. in Artificial Intelligence, University of Pisa

This project concerns the multinomial classification of HuffPost news articles from Kaggle dataset. We compare the performance of various models. As baseline Naive Bayes, Logistic Regression were considered, as intermediate bidirectional LSTM and as state-of-the-art models the ones from BERT family and LLAMA3 LLM.

Required version of python is python==3.10.12.

To install the required libraries, run the following command:

pip install --update pip

pip install -r requirements.txt

For test execution, each model is implemented in a separate notebook file. Notebooks are located in src/test folder. Each notebook is named with the model name.

For running the bidirLSTM experiment pre-trained embeddings are required. For this analysis several pre-trained embeddings were considered, but only those of Glove 6B with size 300 were tested. For running the code create a folder named embeddings in the root directory and put the glove.6B.300d.txt file in it.

Below is the list of directories in the project and their subfolders.

.
├── README.md
├── config
│   └── config.py
├── dataset
│   ├── News_Category_Dataset.json
│   └── preprocessing                           # Preprocessed datasets (pipeline)
├── embeddings
├── hyperparameters
│   ├── BidirectionalLSTM.yaml
│   ├── Coarse_BidirectionalLSTM.yaml
│   ├── Fine_BidirectionalLSTM.yaml
│   └── Logistic.yaml
├── logs                                        # Tensorboard logs
├── models                                      # Serialized models
├── requirements.txt
├── results                                     # Results of the experiments. Each subfolder contains report and confusion matrix
│   ├── BidirectionalLSTM
│   ├── LLAMA3
│   ├── Logistic
│   ├── Naivebayes
│   ├── bert-base-uncased-ag-news
│   ├── distilbert-base-uncased
│   ├── newsbert
│   └── roberta-base
└── src
    ├── main
    │   ├── models
    │   │   ├── bidirLSMT.py
    │   │   ├── logistic.py
    │   │   ├── model.py
    │   │   ├── naivebayes.py
    │   │   └── transformers.py
    │   ├── pipeline
    │   │   ├── functions.py
    │   │   └── pipeline.py
    │   └── utilities
    │       ├── plotting.py
    │       └── utils.py
    └── test
        ├── *.ipynb
        └── unit_testing
            ├── test_functions.py
            └── test_pipeline.py

Alongside the model notebooks, the src/test repository, includes notebooks containing dataset and pipeline statistics (dataset_stat and pipeline_stat), as well as a notebook showcasing the clustering analysis performed using BERTopic.

Finally, the src/test/unit_testing folder contains python scripts for unit testing of the pipeline and pipeline's functions.

Some directory does not contain any files (or even it is not there), as they are too large to be uploaded to GitHub.