Enhancing Affective Computing in NLP Through Data Augmentation: Strategies for Overcoming Limited Data Availability

In this chapter, we delve into the realm of affective computing within natural language processing (NLP) and highlight the challenges posed by limited data availability. While supervised machine learning models are trained on annotated samples to predict the class of unseen data, however, these models may fail to generalize and accurately classify testing samples if the available training data for binary or multi-classification problems is imbalanced and does not equally represent all classes. This biased data can lead to lower precision values for minority classes. The proliferation of content on social media platforms necessitates the extraction of insights in various domains of NLP, such as sentiment analysis, hate speech detection, mental health detection, and spam detection. Researchers have explored various machine learning approaches to mine useful information from this unstructured data. However, due to the lack of available data, training sets are often collected from social media platforms and labeled for classification tasks. In many NLP domains, the gathered data is imbalanced and does not contain equal training samples for all classes, posing a significant challenge in achieving accurate classification. This chapter provides an overview of data imbalance problems in classification and addresses major challenges and issues faced by popular NLP applications such as hate speech detection and mental health detection due to the availability of imbalanced datasets for training classifiers. The need for data augmentation to solve this problem is discussed, and popular data augmentation techniques such as easy data augmentation, using thesauruses, using language models, and machine translation are explored. An experimental study of various classification models with and without augmented training samples of minority classes for two popular NLP datasets of hate speech detection is presented. Finally, the chapter concludes by discussing the empirical results achieved to demonstrate the effectiveness of data augmentation methods in handling the problem of data imbalance in classification.