Development of an Interpretable Neural Network Model for Creation of Polarity Concept Dictionaries

Automatic creation of polarity dictionaries is an important issue, as explanations of prediction models are often required in the financial industry. This paper proposes a novel method of developing an interpretable and predictable neural network model. The neural network model we built can extract polarity scores of concepts from documents. Furthermore, we can detect pairwise interactions between concepts, and create polarity concept dictionaries using our neural network model. The model was built using vector representations of words and polarity scores for about 100 words provided by financial professionals, and we obtained about a hundred times more polarity scores for unknown words through backpropagation. First, we analyze the properties of our method from a theoretical point of view. We then confirm its capabilities by conducting simulations of assigning polarity scores to unknown words and detecting interactions using artificial data. We subsequently estimate sentiment tags using real financial textual datasets. Compared with other conventional methods, the proposed approach can forecast sentiments with higher F1 scores. Finally, we develop a polarity concept dictionary based on Yahoo! Finance board textual data.