Evaluation of rule-based, CountVectorizer, and Word2Vec machine learning models for tweet analysis to improve disaster relief

Social media platforms, such as Twitter, are being increasingly used by people as a means of requesting help during disaster events. Machine learning techniques can be used to identify tweets containing requests for help and classify them based on the type of aid that is being requested. In this paper, we build and compare three different models to classify tweets requesting aid into the categories “food,” “water,” “energy,” and “medical” by using the tweet dataset from Hurricane Sandy, which took place in 2012. Our first model uses a rule-based classifier. Our second model is based on the Scikit-learn toolkit's CountVectorizer, and the third model uses the Word2Vec based classifier. We found that the machine learning models based on CountVectorizer and Word2Vec have higher accuracy than the rule-based classifier model. We also show that the model based on Word2Vec provides the highest accuracy among the three models.