Abstract

Social media platforms like Twitter providea continuous stream of public opinion in the form ofvast amounts of informal text, which presents challengesfor automated analysis. This paper outlines a completeend-to-end system for performing sentiment analysis onTwitter data. The system utilizes robust preprocessingtechniques, an LSTM-based deep learning classifier, and isdeployed with a full-stack architecture featuring a Pythonbackend and a React frontend. We focus on practicalaspects such as dataset curation, tokenization, model train-ing, and evaluation. The methodology is detailed, alongwith an experimental evaluation that includes an ablationstudy comparing LSTM with CNN and transformer-basedmodels. A user study was also conducted to assess thesystem’s usability. The paper acknowledges limitations likehandling sarcasm and domain bias and suggests futureenhancements, including transformer model fine-tuningand multilingual support.