Enhancing Sentiment Analysis : for Improving Sentiment Classification of Bitcoin Related Tweets

Sammanfattning

Background: Cryptocurrency markets especially BTC experiences high volatility and they are highly influenced by the public sentiment. Social media platforms like Twitter generates huge volume of textual data which contains user sentiment. Sentiment analysis has been widely applied in this context. But it lacks systematic research into how the different preprocessing strategies effects the sentiment classification. In addition to this, the comparison of transformer and recurrent neural network architectures in the context are mostly based on accuracy to assess the model performance. Objectives: The main objective of this thesis is to identify the recent preprocessing trends in financial text sentiment analysis and evaluate how they effect model performance in different aspects in the context of BTC sentiment analysis. Another objective of the thesis is to compare transformer based models and recurrent neural network models for BTC sentiment analysis. Methods: A systematic literature review was conducted to identify the frequently used preprocessing techniques in financial/cryptocurrency sentiment analysis. Based on the findings from the literature review three different levels of preprocessing configuration were designed baseline, domain/crypto aware, and noise filtered hybrid preprocessing. These three levels of preprocessing configurations were evaluated using experimentation by training a transformer based FinBERT model. In addition to this, a Bi-GRU model was trained on highest level of preprocessing configurationand compared. Accuracy and macro F1 score were the comparative metrics used in the study. Results: The obtained results from the study states that preprocessing quality will have a greater impact on the sentiment classification performance. The domain aware preprocessing combined with noise filtered and hybrid strategies on top of general baseline NLP preprocessing boosts the performance of the model. The transformer model FinBERT outperformed the bidirectional gated recurrent unit model in terms of accuracy. But the RNN model achieved higher macro F1 score and displayed more balanced performance across the sentiment classes. Conclusions: The preprocessing step should be given priority in BTC sentiment analysis as it plays important role in achieving better performance which is justified by the results of the thesis. The transformer based models ensure overall correctness while recurrent neural network models still give good competition in balanced class performance.