Emotion Recognition of Students’ Bilingual Tweets during COVID-19 Pandemic using Attention-based Bi-GRU
Ian James H. Recto, Andrea Danelle P. Quilang, Larry A. Vea · 2022
This paper studied the emotions manifested by students from March 2020 to April 2021, a year of the Coronavirus Disease-2019 (COVID-19) pandemic. Our tweet compromises Taglish (Tagalog—English) texts, a low-resource code-switching language. The texts were cleaned and translated from Taglish to English. WordNet Affect was used to annotate the text with Happy, Angry, Sad, Surprise, and Fear as the output. A neural network, Bidirectional Gated Recurrent unit (Bi-GRU) with Attention layer, was used, and it was compared to Bernoulli Naïve Bayes (BNB) and Support Vector Machine (SVM), which are commonly used algorithms for Taglish emotion recognition. A 100-dimensional GloVe word embedding was applied to the data before training. The augmentation method does not affect the model’s performance negatively; instead has helped the Bi-GRU with Attention boost its performance. Bi-GRU with attention has a higher F1-score on all emotions compared to the other three algorithms but, as expected, requires a large amount of data. The results show that the most dominant emotion manifested by students throughout the year is surprise immediately followed by Sad and Fear. The three are close in values.