Custom NLP Model for Sentiment Analysis

As per Kaggle notebook

Project Summary

Extract, Process and Predict the sentiment of Twitter tweets referencing South Africa's top 5 banks.

Multiple models were created and tested to predict sentiment of tweets as either postive, neutral or negative

Note:

1. Multiple datasets were used as training of the model, with Sentiment 140 being the main contributor of tweets (1.6 million)
2. Proof of concept version was completed using a pretrained model (Textblob)

Datasets used

1. Sentiment140
2. Twitter and Reddit
3. US airlines
4. Apple sentiment

Known issues with sentiment analysis:

• Sarcasm - "thanks FNB, now I cant open my account cause its frozen"
• Comparison of entities - "Capitec is the worst, you should use Standard Bank"
• Training data on non-South African tweets - Jargon and lingo is different
• Language usage - multiple languages are used in South Africa

Reference notebooks:

• Full sentiment analysis
• beginners notebook