Sentiment Lexicon Induction and Interpretable Multiple-instance Learning in Financial Markets

Abstract

Sentiment analysis has been widely used in the domain of finance. There are two most common textual sentiment analysis methods in finance: \textit{dictionary-based approach} and \textit{machine learning approach}. The dictionary-based method is the most convenient and efficient method to extract sentiments from the text, but the words in the dictionary are limited and cannot capture the full scope of a particular domain. Additionally, it is expensive and unsustainable to manually create and maintain domain-specific dictionary using expert opinions. Deep learning models become mainstream methods in sentiment analysis because of their better performance by utilizing extra information on a larger corpus and more complex model structures. However, deep learning models often suffer from the interpretability problem.

This thesis is an attempt to address the issues of both methods. It proposes a machine learning method to do a corpus-based sentiment lexicon induction, which extends the sentiment dictionary that is customized to analyze corporate conference calls. The new extended dictionary is shown to have a better performance than the original dictionary in terms of the three-day returns of the companies in the MSCI universe. It also proposes a highly interpretable attention-based multiple-instance learning model to perform sentiment classification. It also shows that the newly proposed model has comparable accuracy performance to the state-of-the-art sequential models with better interpretability. A keyword ranking is also generated by the model as a by-product. A new sentiment dictionary is also generated by the deep learning method and shows even better performance than both the extended dictionary and the original dictionary.