Enhance Implicit Sentiment Analysis through Multi-Step Prompting

Implicit sentiment analysis, as a challenging branch of sentiment analysis tasks, faces issues such as the lack of explicit sentiment features and the complexity of textual semantics. This study is inspired by the Chain of Thought (CoT) method and proposes a framework called Multi-Step Prompting of Chain of Thought (MSP-CoT). First, the framework employs a multi-step prompting reasoning process that combines the common-sense reasoning capabilities of Large Language Models (LLMs) with CoT method to effectively analyze aspects and implicit opinions within the text. Then, it connects the aspects and implicit opinions derived from the preceding prompts with the context, serving as the input for the final prompt, thereby gradually inducing sentiment polarity. We evaluated the model using the SemEval 2014 dataset, and the experimental results indicate that the proposed method substantially outperforms existing baselines. Compared to baseline models, our MSP-CoT model demonstrates a significant improvement. Notably, ChatGPT + MSP-CoT achieves an improvement of $\mathbf{4 2. 6 1 \%}$ in F1 score in the overall setting compared to traditional models.