AN OPTIMIZED DEEP LEARNING FRAMEWORK FOR SARCASM-AWARE ANALYSIS OF CITIZEN FEEDBACK TO SUPPORT LOCAL GOVERNMENT PROCUREMENT DECISIONS

Abstract

The use of social media has become an important way through which citizens give feedback to assess available services, products and initiatives at the local governance level. Although these platforms are ideal sources of qualitative information on what people feel about things, when sarcasm is present in the user generated content, sentiment analysis is likely to be misinterpreted, thus constraining the ability to use data to make decisions. In order to overcome this problem, this paper offers a sentiment analysis structure equipped with sarcasm, which is constructed using an Optimized Deep Belief Network (ODBN) to analyze social media reviews with local self-government issues. The framework suggested is based on linguistic preprocessing, hybrid feature extraction with semantic sentiment detectors and stylistic punctuation markers, and optimized deep hierarchical learning to achieve strong sarcasm classification. Quantitative analysis performed on benchmark sarcasm data show that the proposed ODBN has an accuracy of 93.0, precision of 92.0, recall of 94.0, F1-score of 93.0, and ROC-AUC of 0.96 compared to Support Vector Machine, Random Forest, and Bi-LSTM models by a margin of 4% to 11% on key metrics. On a qualitative axis, the decrease in the false sarcasm misclassification boosts the constructiveness of citizen responses and helps in more credible evaluation of the trends related to the opinion of the people. The findings validate that optimized deep learning can be used effectively in detecting sarcasm and the reliability of sentiment. The proposed strategy has practical value to the local self-government institutions through the ability to evaluate the outcomes of the public services and policy more accurately, transparently, and based on evidence.