Replacing human works with machine learning for event extractions from economic news

Abstract

Extracting events affect stock prices from economic news is important for algorithm trading. There are two methods of event extraction. The first is pattern-matching using rules prepared by humans. The second is using an estimation model with algorithms such as machine learning. We need maintenance for rules in either way according to the appearance of new keywords and changes in article format. Selecting the first way makes rules management difficult due to the inefficient stacking of similar rules or omissions of considering some patterns. To solving these difficulties, we developed event extraction models using machine learning. We selected ALBERT and BPE-dropout for NLP and tokenization, respectively, and removed unnecessary parts from news to train the model efficiently. We prepared possible news patterns for training and confirmed that the model has sufficient accuracy for use in business.