gundruke/da_bert_sentiment_analysis

Danish Sentiment Analysis

Information

• Dataset : DDSC/angry-tweets
• Base model : Danish bert botxo

Approach

• Preprocessing

  • Links and Usernames are replaced with @USER and [LINK], removing those keyholders
  • Removing hashtags as they generally donot contribute to sentiment
  • Removing emoji as models used in this notebook donot take emojis into consideration (replacing with their meaning could also be tested)
  • lowercase
  • Stopwords removal, danish stopwords from NLTK

• Training with HF trainer

• Training with pytorch loop

• Uploading model to Huggingface hub

• FastAPI endpoint

• Packaged the api service as a docker container