Contextual Sentence Analysis for the Sentiment Prediction on Financial Data

Newsletters and social networks can reflect opinions about individual stocks from the perspective of analysts and the general public on their products and/or services. Therefore, sentiment analysis of these texts can provide useful information to help investors make investment decisions. In this paper, a hierarchical stack of Transformers model is proposed to identify the sentiment associated with companies and stocks, by predicting a score (of data type real) in a range between -1 and +1. Specifically, we fine-tuned a RoBERTa model to process headlines and microblogs. Furthermore, we combined it with additional Transformer layers to process the sentence analysis with sentiment dictionaries to improve the sentiment analysis. We evaluated it on financial data released by SemEval-2017 task 5. Our proposition outperformed the best systems of SemEval-2017 task 5 and strong baselines. We verified that the combination of contextual sentence analysis with the financial and general sentiment dictionaries indeed provided useful information to our model and allowed it to generate more reliable sentiment scores.