A DEEP LEARNING-BASED HYBRID CNN–BILSTM MODEL FOR SENTIMENT ANALYSIS OF ROMAN URDU E-COMMERCE TEXT

Abstract

The paper presents a hybrid deep learning model that was used to analyze the sentiment of the aspect-level customer reviews on Roman Urdu e-commerce reviews by combining CNN with BiLSTM, on 3,923 customer reviews, specifically of the mobile product category. Each review was tagged by the aspects of it, such as price, delivery, service, or product quality, and sentiment polarity, i.e., positive, negative, or neutral. It was first cleaned, normalized, tokenized, and converted into Word2Vec embeddings to account for the occurrence of irregular spellings, code-mixing, and other informal styles of writing common to Roman Urdu, and then trained. The CNN-BiLSTM was trained with the Adam optimizer and tested with the standard measures, such as the accuracy, precision, recall, F1-score, and ROC-AUC. Experiment results also showed an overall accuracy of 72.5 with a macro-F1 of 0.64 and exceeded the baseline models, such as CNN-only, BiLSTM-only, and SVM classifiers. The confusion matrix additionally confirms the stable performance across the sentiment types in each of the categories by showing the highest accuracy of the positive class. It also indicates that the hybrid model proposed in this paper is able to teach representative attributes out of noisy and unstructured Roman Urdu text. In addition, it offers a beneficial domain-specific data set as well as a benchmark to guide future studies on low-resource language sentiment analysis, specifically in the case of South Asian e-commerce websites.