Enhancement learning on financial text data

Abstract

With the fast development of natural language processing (NLP), financial text data processing has gained much attention due to its huge potential business value. Deep learning model based on manual perception-based labeling is commonly used to illustrate implicit meanings behind financial text. However, such manual labeling is costly and subjective, and may not perform well due to its weak link with direct financial trading. This paper therefore proposes a novel learning model called Enhancement Learning (EL) on financial text data by using task-based labeling. Financial text often has its application task. The derived financial trading data from such task called task-based label is objective and links to certain characters of financial text. Compared to model trained only on manual labels, EL consists of two models trained by manual perception-based labels and derived task-based labels respectively. Then, the task-based model will be used as main model to produce initial judgment on text, with the perception-based model as a filter to drop cases which are different from common perception. This task-oriented perception-enhanced approach can improve the performance of financial text data due to its direct link with financial task and further verification from perception. This paper illustrates the proposed Enhancement Learning on financial text data by the case in stock return prediction. Numerical experiments show that EL performs better than both the perception-based model and the task-based model.