Real or Not? NLP with Disaster Tweets
This exploratory data analysis of the Real or Not? NLP with Disaster Tweets Kaggle competition dataset looks at possible underlying patterns in the textual data contained in tweets. The analysis includes syntactical features such as punctuation, spelling mistakes and word frequencies, and semantic features such as sentiment, emojis and bigrams. Further, we look at the top splits in a decision tree and perform latent semantic analysis in an attempt to uncover lower-dimensional patterns. The analysis reveals some underlying differences between the classes and demonstrates how machine learning can be used on this classification problem.
This project paper discusses the use of non-neural and neural methods for binary text classification in Tweets. A range of different feature extractors, using latent semantic analysis, are tested with a wide range of models. Bayesian optimisation is used to optimise hyperparameters. Gradient Boosting with pre-trained embeddings performs best out of the non-neural methods, while a 2-layer LSTM RNN produces the best neural cross-validated result. Additionally, the BERT model is finetuned for classification, scoring better than all previous models tested.