Sarcasm Detection in Twitter - Performance Impact While Using Data Augmentation: Word Embeddings

Abstract

Sarcasm is the use of words commonly used to ridicule someone or for humorous purposes. Several studies on sarcasm detection have utilized different learning algorithms. However, most of these learning models have always focused on the contents of expression only, thus leaving the contextual information in isolation. As a result, they failed to capture the contextual information in the sarcastic expression. Moreover, some datasets used in several studies have an unbalanced dataset, thus impacting the model result. In this paper, we propose a contextual model for sarcasm identification in Twitter using various pre-trained models and augmenting the dataset by applying Global Vector representation (GloVe) for the construction of word embedding and context learning to generate more sarcastic data, and also perform additional experiments by using the data duplication method. Data augmentation and duplication impact is tested in various datasets and augmentation sizes. In particular, we achieved the best performance after using the data augmentation method to increase 20% of the data labeled as sarcastic and improve the performance by 2.1% with an F1 Score of 40.44% compared to 38.34% before using data augmentation in the iSarcasm dataset.

Measuring the Quality of Semantic Data Augmentation for Sarcasm Detection

“…Also, augmenting by only replacing one word with the synonym could potentially generate synthetic data that is very similar to the original one, thus risking overfitting the model. In earlier works, Handoyo et al [21] augmented the data by only changing one word to its synonym. demonstrate the overfitting effects, which cause performance to decline as more augmented data are used.…”