TelMeMood: A Multimodal Dataset for Sentiment Analysis of Telugu Memes

Memes have emerged as a compelling medium for expressing opinions on social media, making their analysis an important research area in several domains, such as politics, entertainment, and public opinion analysis. Although most of the existing work regarding the analysis of meme sentiments is limited to English and other resource-rich languages, resource-poor Indian indigenous languages such as Telugu are still underexplored in spite of the increasing quantity of regional users throughout social networks. The absence of benchmarked datasets poses a significant challenge in advancing multimodal sentiment analysis research in Telugu. This paper addresses this research gap by introducing TelMeMood, a comprehensive multimodal dataset designed for sentiment analysis of Telugu memes, which comprises 4800 memes annotated with three sentiment categories: positive, negative, and neutral. The study also provides well-defined annotation guidelines to facilitate the development of future resources in this area. Furthermore, the dataset has been evaluated using a range of unimodal (textual and visual) and multimodal (text+image) models, revealing that the incorporation of multimodal features improves sentiment classification accuracy by 1.46% compared to unimodal approaches. This work establishes a benchmark for Telugu meme sentiment analysis and contributes significantly to extending multimodal research in resource-poor languages.