garyhsu29/Natural-Language-Processing-with-Disaster-Tweets

Kaggle competition - Natural Language Processing with Disaster Tweets

Kaggle competition - Natural Language Processing with Disaster Tweets

Slide for the project

Attempt 1: Use Keyword and Location as feature

• Keyword Preprocessing: Remove URL encoding
• Location Preprocessing: Remove the Nation/State, only keep the first part of location information
• Results:
  • Model 1: Decision Tree (71.80% validation accuracy)
  • Model 2: Support Vectior Machine (73.46% validation accuracy)
• Kaggle Leaderboard: 788/868

Attempt 2: Use Tweet and Location as feature

• Preprocessing: Different model need to apply different preprocessing steps for tweet
• Paper Research:
  • TWEETEVAL
  • BERTweet
• Raw Transformers model (79.19% validation accuracy)
• BERT related Models:

Model Name	Paper	Validation Accuracy	Test Accuracy(Kaggle Leaderboard)
bertweet-base	BERTweet	84.17%	84.80%
bertweet-covid19-base-uncased	BERTweet	84.24%	82.08%
bertweet-large	BERTweet	84.36%	82.27%
twitter-roberta-base	TWEETEVAL	83.44%	81.77%

• Kaggle Leaderboard: 48/868