A semantic and syntactic enhanced neural model for financial sentiment analysis

This paper studies the methodology of inferring bullish or bearish sentiments in the financial domain. The task aims to predict a real value to represent the sentiment intensity concerning a target (company or stock symbol) in a text. Previous researches have proved the validity of using deep neural networks to automatically learn semantic and syntactic information for sentiment prediction. Despite the promising performance, these approaches implicitly obtain the target-sentiment representation by a sentence-level vector, lacking explicitly modeling the semantic relatedness between a target and its context. In this paper, we tackle the task by a novel semantic and syntactic enhanced neural model (SSENM), which incorporates dependency graph and context words to guide a target representation. In particular, we devise a self-attentive mechanism to capture semantic contextual information and an edge-enhanced graph convolutional network (E-GCN) to aggregate node-to-node features. In addition, the existing FSA is limited in size, which is prone to the overfitting problem for modern neural models. We further develop a Manifold Mixup strategy to generate pseudo data in training. We perform extensive experiments on two public benchmarks, SemEval2017task5 and FiQA challenges. Results show that our model outperforms the state-of-the-art model by 2% wcs scores on SemEval2017task5 and 3% R2 scores on FiQA, respectively. Finally, we present detailed analysis to indicate the effectiveness of each proposed component.