Predicting E-Learning System Adoption Rates Using a Dense Convolutional Neural Network Integrated with BERT for Sentiment Analysis

Abstract

             The exploration of sentiment extraction within the realm of e-learning systems through the utilization of social media networks has constituted an ongoing subject of investigation within the field of data mining. To discover the significant intention of the student to the e – learning system through sentimental analysis models stands to be inadequate and sparse as most of the data identified to be non actionable on employment of the existing sentiment prediction technique.  The extracting the sentiment is alone not feasible in the forecasting the interest of the student to the e- learning platform. In addition to incorporation of the annotation, domain knowledge, Sensitiveness and subjective information related to forecasting and prediction analysis on the social network data extraction will eliminate the challenges in the traditional sentiment analysis approaches using machine learning algorithms. In this article, deep learning architecture entitled as Dense CNN +BERT model has been constructed on incorporation of annotation, sensitiveness, subjective information’s, sentiments and domain knowledge. Initially the preprocessing of the data is carried out for normalizing and annotating phrases and words in the data. Preprocessed work sequence is employed to BERT model as it considered as a natural language processing approach to extract the local and global features to obtain the opinion lexicon and it is represented as word embedding vector. Those embedding vector is passed to the proposed dense convolution neural network. Convolution layers composed of inbuilt constraints which compute the high level features from the word embeddings to eliminate the inconsistency related with the extracted features. The challenge of correlating dimensions with the max pooling layer has been addressed by implementing down sampling techniques. The utilization of the ReLU Activation Function has been investigated under specific conditions to gauge the effectiveness of actionable attributes' potency. Furthermore, domain knowledge has been applied for the dynamic refinement and validation of optimal features. Fully connected layer of containing softmax function determines the polarity of the feature and polarity of the feature predicts the adoption of the student to e learning in future.  An experimental result has been carried in the twitter dataset to prove that proposed framework outperforms other conventional machine learning techniques with respect to efficiency, Fmeasure, parameter insensitiveness and accuracy.