M-PIN: Multilingual prompt-based implicit sentiment analysis network

Sentiment Analysis (SA) is a classic task in Natural Language Processing (NLP). Compared to Explicit Sentiment Analysis (ESA), Implicit Sentiment Analysis (ISA) is more challenging but does not achieve enough attention. As a knowledge intensive task, ISA needs more information than ESA, but existing models prefer introducing external knowledge bases with additional computational resources, ignore the prior knowledge in Pre-trained Language Model (PLM). So we propose a Multilingual Prompt-based Implicit sentiment analysis Network (dubbed M-PIN), where the prior knowledge guidance module and context information module are introduced, the guidance is embedded in trainable template, which can effectively activate PLM knowledge with the context information. Experiments on five public datasets demonstrate the M-PIN effectiveness, M-PIN has better performance than general Large Language Model (LLM). In addition, ablation experiment results show that combined with the context information and on the premise of indicating the evaluation direction for PLM, even a simple prompt template can effectively improve the ISA performance.