Sentiment Analysis of Multilingual Social Media Texts Using Natural Language Processing and Large Language Models

Abstract

Sentiment analysis becomes difficult when social-media users mix languages, shorten words, attach emojis, or express evaluation through irony and negation. Large language models offer flexible multilingual interpretation, but their output can vary across languages and prompts. This article presents an auditable workflow that combines transparent natural language processing with selective LLM review. A controlled corpus of 180 posts was created in Azerbaijani, Turkish, English, and Italian: 120 balanced training examples and 60 held-out posts covering clean wording, emojis, code-mixing, and negation. The proposed lexicon-character hybrid reached 0.883 accuracy and 0.885 macro-F1, compared with 0.650/0.647 for word TF-IDF and 0.817/0.813 for character TF-IDF. A prespecified gate accepted 58.3% of posts with 94.3% accuracy and sent 25 uncertain cases to LLM or human review. No LLM output entered the numerical score.