Employing BERT-DCNN with sentic knowledge base for social media sentiment analysis

Machine learning is considered advantageous smart cities through sentiment analysis using social media reviews. Social media reviews can be helpful inside smart cities for various purposes. Primary convolutional neural networks (CNNs) are hard to implement for parallelizing applications and inadequate to understand the contextual semantics in long-term sequences for emotion classification. In this context, this paper presents a Bidirectional Encoder Representations from Transformers (BERT) based Dilated Convolutional Neural Network (BERT-DCNN) model, which leverages BERT as a pre-trained language model to generate word embeddings. Additionally, three parallel layers of Dilated Convolutional Neural Network (DCNN) stacked with a global average pooling layer helps in fine-tuning the model. Our implemented BERT-DCNN model performs dimensionality reduction and assimilates an increase of related dimensions withstanding any information loss. Furthermore, the model can capture long-term dependencies by utilizing various dilation rates. Moreover, the sentic knowledge base is incorporated in our model, enabling it to achieve concept-level sentiment analysis. Our experimental study demonstrates the importance of the implemented model in terms of F-measure, recall, precision, and accuracy with different machine learning models.