Abstract:
"Text based emotion detection has emerged as a crucial yet challenging area of research
within the domain of Machine learning along with natural language processing. The problem
circulates around any field which involves long text data which the readers are expected only
to identify the positive, negative, or neutral nature of those. It’s basically the emotions that
decide that nature. Inability to identify the emotions conveyed in long text paragraphs in a
short time stands as a problem. Whilst there exist several text-emotion detection models they
can be treated as incomplete due to the lack of quality.
The proposed system implements a novel method which uses two layers of Long Short
Term Memory architecture in addition to the base neural network. Long Short-Term Memory
architecture means a recurrent neural network approach, a class of artificial neural networks
designed for sequential data. The approach focuses on encoding the labels and tokenizing
dataset texts and converting them to sequences before feeding them to the model.
The system implemented had an accuracy of 83% on the test dataset with 15 epochs of
training. With each epoch’s accuracy being greater than the previous epoch’s accuracy value.
The classification report indicated a total support value of 6459. The lowest accuracy of a
label for this classification model was 72% for the label ‘sad’ with a support of 234. The
highest accuracy was obtained for the three labels ‘jealous’, ‘disgusted’ and ‘prepared’ with a
value of 91% for both. Therefore, the rest of the 29 labels were in-between an accuracy of
72% and 91%. The usage of 32 labels for the implementation of the classification model is a
novelty in this research. "
