EMSA: Explainable multilingual sentiment analysis models providing sentiment analysis across multiple languages.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Multilingual sentiment analysis faces challenges from linguistic diversity and lack of model explainability.
• Existing models often lack transparency, hindering user trust and practical application.

Purpose of the Study:

• To develop an Explainable Multilingual Sentiment Analyzer (EMSA) that addresses limitations in cross-lingual sentiment analysis.
• To enhance the interpretability and performance of sentiment analysis models across different languages.

Main Methods:

• Proposed a novel framework, EMSA, integrating large language models with prompt engineering.
• Employed a two-stage process: chain-of-thought prompting for reasoning and explicit interpretability for classification.
• Evaluated EMSA on the GubaSenti (Chinese financial) and SST (English benchmark) datasets.

Main Results:

• EMSA demonstrated superior performance compared to pre-trained models like RoBERTa, XLNet, and ALBERT.
• The framework provided transparent reasoning steps, increasing user trust.
• Achieved improved multilingual sentiment classification accuracy and interpretability.

Conclusions:

• EMSA offers a significant advancement in multilingual sentiment analysis by combining high performance with explainability.
• The proposed framework contributes to developing more trustworthy and practical sentiment analysis systems.
• Highlights the potential of prompt engineering with large language models for interpretable AI.