SynText - Data Augmentation Algorithm in NLP to Improve Performance of Emotion Classifiers

Emotion classification is a common task in natural language processing (NLP) where the goal is to classify the emotion expressed in a given text, such as joy, sadness, anger, or fear. Emotion Classification in natural language is an integral task in application domains like marketing, data analysis, and pattern recognition. However, building accurate emotion classifier models can be challenging due to several factors, such as the subjective nature of emotions, the complexity of language, and the scarcity of labeled data. To tackle these problems, research related to different methods of data augmentation has been performed, to increase both, the quantity, and the quality of labeled data. The major challenge faced with labeled data is that the preservation of the label and meaning of the sentence is of top priority while performing augmentation, which is what this research aims to tackle. This research aims to improve existing methods of data augmentation in the field of NLP with the aim of enhancing emotion classifier models. Previous methods focus on augmentation but don’t keep the increase of diversity as the main aim of augmentation. We propose a novel data augmentation technique “SynText’’ which performs augmentation by using contextually meaningful synonyms, with a focus on increasing the diversity of emotion-related words, while also preserving the label of the given data point. The technique is applied on some benchmark datasets mentioned in literature, like the GoEmotions Dataset and the SST5 dataset, and it is found that emotion classification models trained on datasets augmented using “SynText’’ tend to give better results for metrics like F1 Score and Accuracy.