Improving Event-Level Financial Sentiment Analysis with Adaptive Reasoning Order Chain-of-Thought Prompting

Event-level financial sentiment analysis (EFSA), an emerging research field of considerable interest, aims to extract quintuples of sentiment elements from given financial news articles. Preliminary studies indicate that the sequential reasoning order in chain-of-thought (CoT) prompting considerably impacts the performance of financial information extraction. We propose an adaptive reasoning order CoT framework (ARO-CoT) that enables large language models to determine the optimal query order for the extraction of financial quintuples (company, industry, coarse-grained event, fine-grained event, and sentiment) from Chinese news articles. Specifically, ARO-CoT trains models to automatically select query sequences through reasoning order adaptation, thereby yielding more precise quintuples. To validate the effectiveness of our framework, we conducted comprehensive experiments on the EFSA benchmark. Compared with the previous best EFSA results obtained by retrieval-augmented multipath chain-of-thought prompting, ARO-CoT achieves a new state-of-the-art F1 score and demonstrates robust performance across all EFSA subtasks.