Assignment Option Four - News Categorization using PyTorch
Link to our GitHub repository: https://github.com/lennartmoritz/nlp-news-categorization
- Carlotta Mahncke
- Lennart Joshua Moritz
- Timon Engelke
- Christian Schuler
sudo apt install python3-venvpython3 -m venv venvsource venv/bin/activate- Automatic: ./script-caller.bh`
- Or Manual:
pip install -r requirements.txtpython -m spacy download "en_core_web_sm"
python news-categorization.py -l b t e m
python news-categorization.py -l b t
python news-categorization.py -l b t -e word2vec- Choose from: word, lemma, word2vec, glove
Text categorization using PyTorch
-
A typical workflow in PyTorch.
-
The task is to classify news articles into one of the following categories:
- Business
- Science and Technology
- Entertainment
- Health
As shown in our presentation, the lemma embedding performed the best, followed by the word embedding. Pretrained embeddings like word2vec and glove performed worse than the embeddings we trained ourselves. This is probably because the self-trained embeddings are more specific to our dataset.
For the binary classification tasks, we saw that differentiating between business and science and technology was the hardest task, while health and entertainment were the easiest. This is probably because the business and science and technology articles are very similar in their vocabulary, while the health and entertainment articles are very different.
You can use the script bert_classification.py to fine-tune a pre-trained BERT classifier on our dataset.
The results are slightly better than the results we got with our own models (around 96% accuracy compared to
around 94% accuracy on our model for the multi-class classification task).
News Aggregator Dataset https://www.kaggle.com/datasets/uciml/news-aggregator-dataset