xLiMe-MELT: A Multi-layer Italian Twitter Dataset for Explainable Sentiment Analysis with Systemic Functional Linguistics

We present xLiMe-MELT (Multi-layer, Explainable, Linguistically-Theorized), a new dataset extending the Italian xLiMe Twitter Corpus (Rei et al 2016) with multi-layer sentiment annotations enriched by Systemic Functional Linguistics (SFL) theory (Halliday et al 2004) and the Appraisal Framework (Martin & White 2005) explainability. Building on the original manually annotated 8,601 tweets, we provide three additional parallel annotation layers: (i) transformer-based predictions (XLM-RoBERTa - Barbieri et al 2022), (ii) GPT-5 standard prompting, and (iii) SFL-GPT-5. The SFL-guided layer introduces fine-grained categories (e.g., cautionary, neutral with humour, slight sadness) that systematically unpack cases otherwise collapsed into “neutral.” Each of these annotations is further accompanied by a natural-language explanation, linking the label to explicit linguistic cues such as evaluative lexis, intensifiers, or irony. With over 34,000 annotations and explanations, xLiMe-MELT provides a rich resource for studying the limitations of polarity-based schemes, the interpretive gains of linguistic theory, and the comparative behaviour of transformer versus LLM-based classification. Beyond sentiment classification, the dataset contributes to explainable and sustainable NLP as well as NLP for languages other than English. We release xLiMe-MELT to support future research in sentiment analysis, stance and emotion detection, explainable AI, and theory-driven annotation workflows in Italian and beyond.