Sentiment Analysis Using 1-of-m and Log-m Character-Level Embedding Algorithms with Deep Learning

Nowadays, most people around the world use social media platforms to express their feelings, ideas, opinions or other sentiments with each other by using their own native languages. To identify of those sentiments, the feature extraction process needs to be performed with help of human experts in the sentiment classification. It is very difficult to perform feature extraction approach without the help of human experts for non-English languages such as Sinhala, Tamil, or any other languages. The purpose of this research is to perform POSITIVE and NEGATIVE sentiment classification of sentences in Sinhala, which is one of the non-English language, using CNN models with the application of character level embedding algorithms named 1-of-m and Log-m. These character-level embedding algorithms construct binary image representation of the sentences before classification. This research uses 9048 Sinhala news comments labeled as POSITIVE (2490), NEGATIVE (2520), and NEUTRAL (4038). However, POSITIVE and NEGATIVE comments were selected due to 0.92 of Cohen's Kappa value. The datasets were split as 80% of training and 20% of testing data. Cross-validation showed that 1-of-m outperformed Log-m with the average testing accuracy of 66.00% and loss value of 0.6461, with precision, recall, F1 scores of 66.22%, 65.41%, and 65.79% respectively.