Political Sentiment and Geography: Deep Learning Insights from Twitter Data During India’s 2024 Election

Abstract

The growth in the use of social media has revolutionized the nature of political communication, bringing new possibilities to conduct analysis of public opinion during elections. This study presents the topic of a geography aware deep learning framework to examine political sentiment from twitter data during instructors go 2024 General Election in India. A large-scale dataset of over 500.000 tweets has been gathered by the Twitter API v2 and processed by an enriched preprocessing pipeline which incorporates noise exclusion, multilingual normalization, named entity recognition and geoparsing. Sentiment classification was set up as a 3 class classification problem (positive, neutral, negative) and two deep learning model architectures were employed, multilingual BERT and Bidirectional LSTM. Experimental results show that the BERT model reached better performance with a level of 88.2% and a macro F1-score value of 0.86 compared to the bi-LSTM baseline. Geographic aggregation and DBSCAN based spatial clustering, have shown the presence of discrete regional class of sentiments with pro government sentiment for Gujarat, Maharashtra and Karnataka and Uttar Pradesh, Bihar and West Bengal showing mixed/opposite trends. Tweet density analysis also showed the concentration of political discourse in major Metropolitan Regions including Mumbai, Delhi, Bengaluru etc. The results warrants demonstration of the effectiveness of combining transformer-based NLP with geospatial analytics in large-scale political opinion mining, and allows for a scalable framework for real-time election intelligence for democratic environments.