BERT-NLP-TweetSentimentClassification: Project Overview

• Determined the sentiment of public towards work from home by analyzing tweets.
• Scraped over 700,000 tweets using snscrape. Performed text-preprocessing and EDA.
• Used AWS sagemaker/kaggle to get the sentiment from a pre-trained model based on DistillBert.
• Imported the pipeline from Hugging Face model hub.

Code and Resources Used

• Python Version: 3.8.5
• Packages: pandas, tensorflow, transformers, matplotlib, seaborn, snscrape, nltk
• GitHub Repo Ref1: https://github.com/debnsuma/Intro-Transformer-BERT/blob/main/BERT-Disaster-Tweets-Prediction.ipynb
• GitHub Repo Ref2: https://github.com/Rohan2002/BERT-natural-disaster-tweet-prediction/tree/master/lib
• Medium article Ref: https://python.plainenglish.io/scraping-tweets-using-snscrape-and-building-sentiment-classifier-13811dadd11d

Web Scraping (snscrape_tweets.ipynb)

Scraped over 700,000 tweets using snscrape. 27 attributes were there, out of which only following were important.

• date
• tweets (content)
• hashtags

Data Cleaning (NLP sentimental analysis.ipynb)

Cleaned the raw text tweets.

• Removed hashtags, @username and other non english words and punctuations from tweets.

EDA

Plotted distribution of tweets (#WFH) before and after covid. Also observed the average words used per tweets:

Model Building (model_building.ipnyb)

• Used transformers to call a pre-trained model from Hugging Face model hub.
• The pipeline used has an autotokenizer and it's called sentiment-analysis.
• Trained the model in AWS sagemaker/Kaggle.

Conclusion and Future Recommendation

• The sentiment of public is almost equally divided between positive and negative for 'work form home'.
• Pipeline was implemented to train the model from Hugging Face model hub
• The accuracy of model can be improved by further training it in a labellled data set of tweets and using other models.