Transformer-based deep learning models for the sentiment analysis of social media data

Abstract

Sentiment analysis (SA) is a widely used contextual mining technique for extracting useful and subjective information from text-based data. It applies on Natural Language Processing (NLP), text analysis, biometrics, and computational linguistics to identify, analyse, and extract responses, states, or emotions from the data. The features analysis technique plays a significant role in the development and improvement of a SA model. Recently, GloVe and Word2vec embedding models have been widely used for feature extractions. However, they overlook sentimental and contextual information of the text and need a large corpus of text data for training and generating exact vectors. These techniques generate vectors for just those words that are included in their vocabulary and ignore Out of Vocabulary Words (OOV), which can lead to information loss. Another challenge for the classification of sentiments is that of the lack of readily available annotated data. Sometimes, there is a contradiction between the review and their label that may cause misclassification. The aim of this paper is to propose a generalized SA model that can handle noisy data, OOV words, sentimental and contextual loss of reviews data. In this research, an effective Bi-directional Encoder Representation from Transformers (BERT) based Convolution Bi-directional Recurrent Neural Network (CBRNN) model is proposed with for exploring the syntactic and semantic information along with the sentimental and contextual analysis of the data. Initially, the zero-shot classification is used for labelling the reviews by calculating their polarity scores. After that, a pre-trained BERT model is employed for obtaining sentence-level semantics and contextual features from that data and generate embeddings. The obtained contextual embedded vectors were then passed to the neural network, comprised of dilated convolution and Bi-LSTM. The proposed model uses dilated convolution instead of classical convolution to extract local and global contextual semantic features from the embedded data. Bi-directional Long Short-Term Memory (Bi-LSTM) is used for the entire sequencing of the sentences. The CBRNN model is evaluated across four diverse domain text datasets based on accuracy, precision, recall, f1-score and AUC values. Thus, CBRNN can be efficiently used for performing SA tasks on social media reviews, without any information loss.