FIGURE 1. Proposed sentence-level sentiment classification model. This section first discusses preprocessing strategies, different text representations, and different DL models with imple- mentation details. Fig. 1 describes the proposed classification model. Ill. METHODOLOGY TABLE 1. Work done in literature for sentiment analysis. In sentiment classification of sentence, sentiment carrier words are more important than the rest of words. To enhance the weightage of words that play key role in sentiment cat- egorization, the attention mechanism is utilized in combina- tion with BiLSTM. With attention all former states can be retrieved and weighted according to some learned measure of relevance to the current token allowing it to deliver more specific information on distant relevant tokens. FIGURE 3. A segment in LSTM with interacting layers. Single layer Conv1D with 128 filters of multi-size ker- nel(3,4,5) is used to extract features. Convolution Neural Network (CNN) is used to find rela- tionships and patterns between data items according to their relative position. They extract higher level features by con- voluting efficiently. CNN learns spatial features of the data, convolutes down to a smaller subset of the data while trying to learn more features from the already learned features. CNNs apply a layer called the pooling layer, which reduces input by combining multiple related inputs while preserving the information. This process is visualized in Fig. 5. 1) DROPOU FIGURE 6. An illustration of fully connected dense layers. Dropout randomly removes some nodes by setting them to zero during training. It avoids learning the same values repeatedly by the model in case of large set of parameters. C. PARAMETER SETS TABLE 3. Summary of parameters setting for dl models. which different deep learning techniques are compared with different embeddings. These approaches are evaluated for sentence-level classification tasks in the domain of views about current affairs, sports, literature, and health. Accuracy is used as our main evaluation criteria. In the following subsections, the details of the experiments and their results are described. As we applied early stopping to prevent overfitting, so the number of epochs varied for different experiments. TABLE 5. Parameters setting for experiments. FIGURE 7. Dataset statistics based on sentiment labels. The dataset comes as a CSV file, with each line containing a sentence and its label(‘p’ for Positive, ’n’ for Negative). An imbalanced dataset of 6000 sentences is used and a dis- tribution ratio of 80:20 is applied for train and test as shown in Fig 7. TABLE 6. Comparative analysis of DL models embedding wise. detail of results obtained for respective models. The highest values achieved by models among all embeddings are high- lighted in bold font. Based on the results, it can be determined that regardless of embedding, BiLSTM-ATT performed better for sentence level sentiment classification. FIGURE 8. ROC curves for sentiment classification by DL models. Model based on BiLSTM-ATT outperformed all others, by achieving the highest recall, Fl, and accuracy. LSTM achieved the highest precision. Despite the fact that C-LSTM has been found to be useful for classification in other lan- guages, it has not improved in our experiments.
