TigXMM: A Multimodal Cross-Lingual Transfer LLM Method for Low-Resource Sentiment Analysis: Multimodal Low-Resource Tigrigna Sentiment Analysis

Sentiment analysis in low-resource languages remains underexplored, particularly for Tigrigna, where communication frequently combines text, emojis, and memes. We introduce TigXMM, a crosslingual, multilingual and multimodal LLM method for Tigrigna sentiment analysis, along with the first multimodal sentiment dataset for this language. The dataset, collected from social media, integrates text, emojis, and meme content to capture real-world communication patterns. We benchmark widely used multilingual models, including mBERT, AfriBERTa, XLM-RoBERTa, XLNet, BLOOMZ, and LLaMA, and highlight their limitations in processing multimodal signals. To address these challenges, we design an LLM-based cross-lingual transfer model with multimodal adapters for text, emoji, and meme fusion using hybrid attention and additive hierarchical strategies. Experimental results demonstrate that our approach consistently improves sentiment classification performance: from 78.4% accuracy on text-only inputs to 81.2% with emojis, 86.3% with memes, and 89.7% when combining all modalities, achieving state-of-the-art performance for Tigrigna sentiment analysis. Beyond performance gains, this work contributes the first multimodal dataset and a reproducible framework, providing open resources to advance sentiment analysis for underrepresented African languages.