ABSTRACT

One of the most common effects in the use of social media today is that people constantly make fun of each other or certain issues or do not take them seriously. Some comments made by sarcastic people in this widespread effect are misunderstood or taken seriously by other users. Some sarcastic comments, especially in the news headlines, create false effects on the readers and create some misunderstandings for people who do not have this sense of humor. Although there are numerous studies on the problem of sarcasm detection, even low performance increment in automatic sarcasm detection is very important and popular task. In this paper, a new hybrid deep neural model is proposed for more efficient automatic detection of sarcastic context. It is aimed to detect sarcasm using a hybrid neural network model CNN+BILSTM and BERT models with bidirectional language processing in a dataset consisting of headlines of The Onion News, which made such sarcastic headlines, and professionally prepared headlines without any sarcastic comments. When the results of this study were examined, it was seen that the model that gave the best results was BERT. In addition, accuracy, precision, recall and F1 score values were checked without using Glove embeddings in the CNN+BiLSTM model, and then the results were compared by applying Glove embeddings. In this comparison, the CNN+BiLSTM model without Glove embeddings gave relatively better results.

Keywords: - Sarcasm Detection, BERT, CNN+BiLSTM, Deep Learning, Hybrid Model