Ambiguity and Sarcasm Detection on Social Media Data Using Ensemble Techniques

Conveying things in a positive way while the intention is negative. This intentional ambiguity in the text leads to Sarcasm and it is a tedious task in Sentiment analysis. Sarcasm detection obtains a lot of importance in all the fields such as businesses, politics and academics and various other fields. Numerous researches have been taken place to detect sarcasm in the text. However, existing social media analysis systems are hampered due to the inability to detect and interpret figurative language. Sarcasm detection has always been an ambiguous and difficult task in the field of sentiment analysis. The main aim of the proposed work is to develop a hybrid model to improve the performance of sarcasm detection in Reddit by utilizing beneficial preprocessing techniques and feature engineering along with an ensemble learning model consisting of Stochastic Gradient Descent and Linear Support Vector Classification (Linear SVC) model. Good standards such as Cross-Validation and Early Stopping as well as additional features like Sentiment Score are employed, which result in significantly increased model accuracy of aroundthree per cent than the individualmodel. A comparative analysis has been done to understand the better models among LinearSVC, Stochastic Gradient Descent, Logistics Regression and Gaussian model and obtained higher accuracy using LinearSVC and Stochastic Gradient Descent model. A Sentiment model using Stochastic Gradient Descent algorithmhas been created using Early stopping and obtained an accuracy of 68 per cent before applying Early stopping method and 69% has been obtained after Early Stopping methodon Twitter data. In order to identifyhow the model performs for an independent dataset the proposed work includes Cross validation strategy and obtained an average score of 67% respectively. In order to identify the positive sentiment, a sentiment model has been implemented and obtained an average accuracy of 85 per cent. The final classifier with ensemble model results with an accuracy of 61 per cent in Sarcasm Model implemented using SGD, 64 per cent for Normal LinearSVC Model and 67 per cent for Final Sarcasm Model that is Ensemble model respectively.