Emotion-Based Sentiment Analysis Using Conv-BiLSTM with Frog Leap Algorithms

Social media, blogs, review sites and forums can produce large volumes of data in the form of users' emotions, views, arguments and opinions about various political events, brands, products and social problems. The user's sentiment expressed on the web influences readers, politicians and product vendors. These unstructured social media data are analysed to form structured data, and for this reason sentiment analysis has recently received the most important research attention. Sentiment analysis is a process of classifying the user's feelings in different manners such as positive, negative or both. The major issue of sentiment analysis is insufficient data processing and outcome prediction. For this, deep learning-based approaches are effective due to their autonomous learning ability. Emotion identification from the text in natural language processing (NLP) provides more benefits in the field of e-commerce and business environments. In this paper, emotion detection-based text classification is used for sentiment analysis. The data collected are pre-processed using tokenization, stop word discarding, stemming and lemmatization. After performing data pre-processing, the features are identified using term frequency and inverse document frequency (TF-IDF). Then the filtered features are turned into word embeddings by documents as a vector (Doc2Vec). Then, for text classification, a deep learning (DL) based model called convolutional bidirectional long short-term memory (CBLSTM) is used to differentiate the sentiments of human expression into positive or good and negative or bad emotions. The neural network hyper-parameters are optimized with a meta-heuristic algorithm called the frog leap approach (FLA). The proposed CBLSTM with FLA uses four review and Twitter datasets. The experimental results of this study are compared with the conventional approaches LSTM-RNN and LSTM-CNN to prove the efficiency of the proposed model. Compared to LSTM-RNN and LSTM-CNN, the proposed model secures an improved average accuracy of 98.1% for review datasets and 97.5% for Twitter datasets.