Fine-Grained Analysis of Financial Tweets

Abstract

This paper decribes our experimental methods and results in FiQA 2018 Task 1. There are two subtasks: (1) to predict continuous sentiment score between -1 to 1, and (2) to determine which aspect(s) are related to the content of financial tweets. First, we propose a preprocessing procedure for decomposing financial tweets. Second, we collect over 334K labeled financial tweets to enlarge the scale of the experiments. Third, the sentiment prediction task is separated into two steps in this paper, i.e., (1) bullish/bearish and (2) sentiment degree. We compare the results of the CNN, CRNN and Bi-LSTM models. Besides, we further combine the results of the best models in both steps as the model of subtask 1. Finally, we make an investigation of aspects in depth, and propose some clues for dealing with the 14 aspects.