Sentiment analysis with word embeddings and Convolutional Neural Networks

Datasets:

1. "Twitter" dataset
   • 61529 different tweets and 300 feature dimensions, with a vocabulary of 73562
   • Twitter hashtags crawled: 'ADBE', 'GOOGL', 'AMZN', 'AAPL', 'ADSK', 'BKNG', 'EXPE', 'INTC', 'MSFT', 'NFLX', 'NVDA', 'PYPL', 'SBUX', 'TSLA', 'XEL', 'positive', 'bad' and 'sad'
   • 3 different sentiments: positive, negative or neutral
2. Pre-trained english word vectors trained with fastText on CommonCrawl and Wikipedia
   • 300 Dimensions was reduced down to 200
   • Source link: https://fasttext.cc/docs/en/crawl-vectors.html
3. Pre-trained word vectors trained with GloVe on Twitter data
   • 27B Tokens 200 Dimensions
   • Source link: https://www.kaggle.com/fullmetal26/glovetwitter27b100dtxt

How to run the code:

1. Written in Google Colab Notebook
2. Download the following files:
   • "Twitter" dataset, found in the folder files
   • FastText word vectors (see link, needs to be reduced to 200 dimensions)

   • model = fasttext.load_model('path to dataset')
     fasttext.util.reduce_model(model, 200)
     #test
     model.get_dimension()

   • GloVe word vectors (see link)
   • sentqs_dataset.npz link: https://cloud.fhws.de/index.php/s/8BwgMykHEf9BwAd
3. Open the latest Notebook
   • Experiments: here the Experiments for the thesis were performed
   • ft_text_classification: Demo file for sentiment analysis
4. Link the datasets needed to run the code
5. For Skip-gram sentence embedding use 1/4th of the "Twitter" dataset, to not overload the RAM
   • can be found in the folder files