Deep Learning-Based Methods for Sentiment Analysis on Nepali COVID-19-Related Tweets

Table 7

Classwise study of our proposed method using classification performance (%).

CNN	Positive			Neutral			Negative			Overall
	P	R	F	P	R	F	P	R	F	P	R	F	A
	69.4	74.6	71.8	51.8	22.4	31.1	69.5	76.8	72.8	63.5	57.9	58.5	68.1
	68.0	59.6	62.6	12.7	00.2	00.5	59.0	82.7	68.6	46.5	47.5	43.9	61.5
	64.1	57.6	60.5	43.9	04.5	08.2	57.4	78.6	66.3	55.1	46.9	45.0	59.5
	71.8	72.0	71.9	63.0	14.7	23.7	66.8	82.7	73.8	67.2	56.4	56.4	68.7

Note/ P, R, and F denote Precision, Recall, and F1-score for three classes (positive, negative, and neutral), respectively. Note that the hyperparameters used in our models are as follows: learning-rate: 1e − 05, batch-size: 32, epochs: 50, and optimizer: RMSProp. Boldface denotes the highest performance.