Sentiment analysis is one of the critical tasks in NLP, and it usually fails in low-resource languages where labeled data is scarce. This work investigates how multilingual embedding models like “paraphrase-multilingual-mpnet-base-v2”, “distiluse-base-multilingual-cased-v1”, and “LaBSE” perform on sentiment classification over a wide variety of languages. We used Hindi, a medium resource language, as our pivot and worked on labeled words. Using this base, we made predictions of how people feel in high-resource languages such as English, French, and Italian and low-resource languages such as Urdu, Bengali, and Marathi. Further, we compare the outputs for each language with those obtained using Hindi to observe the similarity in sentiments obtained and how well the model works across different languages. Results have shown that accuracy is very high in languages with a lot of pre-trained data and more than 70% in languages with fewer resources. This approach takes the advantage of cross-lingual semantic structures so that large labeled datasets become less necessary to make sentiment analysis work in languages that are less studied. These findings show pathways toward making NLP tools more inclusive and performing outreach beyond languages.