Back to Search
Start Over
Class-biased sarcasm detection using BiLSTM variational autoencoder-based synthetic oversampling.
- Source :
-
Soft Computing - A Fusion of Foundations, Methodologies & Applications . May2023, Vol. 27 Issue 9, p5603-5620. 18p. - Publication Year :
- 2023
-
Abstract
- Recent research works have established the importance of sarcasm detection in the domain of sentiment analysis. Automatic sarcasm detection using social media data is a challenging task in the presence of imbalanced classes. Real-life social media data often suffer from this problem of class imbalance resulting in dramatical degradation of the performance of classification models attempting to detect sarcasm. Motivated by this, in the current article, a Bi-LSTM variational autoencoder model has been proposed to alleviate the problem of imbalanced classes in social media datasets targeted to train sarcasm detection models. The proposed BVA model is trained with a large corpus of sarcastic and non-sarcastic tweets to obtain the most suitable latent space representation of the same. These inherently class-biased latent vectors are then oversampled using synthetic minority oversampling techniques. The quality of the proposed method is established by training and testing a set of well-known classifiers in terms of precision, recall, F1-score, AUC, and G-mean. Extensive experiments reveal that the proposed BVA model combined with oversampling techniques can improve classifier performance for sarcasm detection to a greater extent. [ABSTRACT FROM AUTHOR]
Details
- Language :
- English
- ISSN :
- 14327643
- Volume :
- 27
- Issue :
- 9
- Database :
- Academic Search Index
- Journal :
- Soft Computing - A Fusion of Foundations, Methodologies & Applications
- Publication Type :
- Academic Journal
- Accession number :
- 162993139
- Full Text :
- https://doi.org/10.1007/s00500-023-07956-w