Exploring the Impact of Informal Language on Sentiment Analysis Models for Social Media Text Using Convolutional Neural Networks

Absztrakt

The present study sought to investigate the influence of informal language on the effectiveness of sentiment analysis models when applied to social media text. A Convolutional Neural Network (CNN) approach was employed, and the model was developed and trained on three distinct datasets: a sarcasm corpus, a sentiment corpus, and an emoticon corpus. The experimental design involved keeping the model architecture constant and training it on 80% of the data, followed by evaluating its performance on the remaining 20%. The results revealed that the model achieved a high accuracy of 96.47% on the sarcasm corpus, with class 1 exhibiting the lowest accuracy. The sentiment corpus yielded an accuracy of 95.28% for the model. The integration of the sarcasm and sentiment datasets resulted in a slight improvement in accuracy to 95.1%. The inclusion of the emoticon corpus had a marginal positive effect, resulting in an accuracy of 95.37%. These findings suggest that the use of informal language has minimal impact on the performance of sentiment analysis models applied to social media text. Furthermore, the incorporation of emoticon data may lead to a modest improvement in accuracy.