Lexicon-based prompt for financial dimensional sentiment analysis

Abstract

Financial sentiment analysis attempts to identify the sentiment of a sentence using dimensional or categorical approaches. The dimensional approach represents affective states as continuous numerical values and can provide more fine-grained (real-valued) sentiment analysis than the categorical approach, which represents affective states as several discrete classes (e.g., positive and negative). Although categorical approaches have been studied extensively over the past decade, the lack of domain-specific data makes it challenging to use dimensional approaches in financial domains. To solve this problem, we propose a lexicon-based prompt method for domains with no labeled data by introducing domain-specific lexicons to correct mispredicted words. As this method does not apply domain-adaptation techniques, it can be combined with domain-adaptation methods to improve the generalizability of domain-specific sentiment analysis further. Experiments on various pre-training language modes in the finance domain show that the proposed lexicon-based prompt method outperforms common domain adaptation methods, and the best performance is achieved by combining multi-task fine-tuning and domain adaptation methods.