Aspect-Based Sentiment Analysis of Customer Speech Data Using Deep Convolutional Neural Network and BiLSTM

The process of detecting sentiments of particular context from human speech emotions is naturally in-built for humans unlike computers, where it is not possible to process human emotions by a machine for predicting sentiments of a particular context. Though machines can easily understand the content-based information, accessing the real emotion behind it is difficult. Aspect-based sentiment analysis based on speech emotion recognition framework can bridge the gap between these problems. The proposed model helps people with autism spectrum disorder (ASD) to understand other’s sentiments expressed through speech data about the recently purchased product based on various aspects of the product. It is a framework through which different sound discourse documents are characterized into various feelings like happy, sad, anger, and neutral and label the sound with aspect-wise sentiment polarity. This study proposed a hybrid model using deep convolutional neural networks (DCNN) for speech emotion recognition, bidirectional long short term memory (BiLSTM) for speech aspect recognition, and rule-based classifier for aspect-wise sentiment classification. In the existing work, sentiment analysis was carried out on speech data, but aspect-based sentiment analysis on speech data was not carried out successfully. The proposed model extracted standard Mel frequency cepstral coefficient (MFCC) features from customer speech data about product review and generated aspect-wise sentiment label. Enhanced cat swarm optimization (ECSO) algorithm was used for selection features from the extracted feature in the proposed model that improved the overall sentiment classification accuracy. The proposed hybrid framework obtained promising results on sentiment classification accuracy of 93.28%, 91.45%, 92.12%, and 90.45% on four benchmark datasets. The proposed hybrid framework sentiment classification accuracy on these benchmark datasets were compared with other CNN variants and shown better performance. Sentiment classification accuracy of the proposed model with state-of-art methods on the four benchmark datasets was compared and shown better performance. Aspect classification accuracy of the proposed with state-of-art methods on the benchmark datasets was compared and shown better performance. The developed hybrid model using DCNN, BiLSTM, and rule-based classifier outperformed the state-of-art models for aspect-based sentiment analysis by incorporating ECSO algorithm in feature selection process. The proposed model will help to perform aspect-based sentiment analysis on all domains with specified aspect corpus.