A HYBRID DEEP LEARNING MODEL FOR DETECTING POLARITY SHIFT DUE TO IMPLICIT NEGATION IN SENTIMENT ANALYSIS OF MOBILE PHONE BRANDS’ REVIEWS

Abstract

Sentiment, the emotional tone, attitude, or opinion expressed in a piece of text, is critical as it captures the emotional polarity or subjective judgment embedded in a language. Sentiment analysis allows any brand or company to gain insight into how consumers feel about their products, brands, or the company itself through the variety of feedback that comes from the digital marketing platforms. These sentiments can have a positive, neutral or negative polarity. A problem arises with the existing sentiment analysis model when the sentiment of text is reversed which results in positive polarity being classified as negative polarity and the vice versa. This results into a polarity shift problem. Therefore, the classifier needs to be able to recognize polarity shifts and separate the impact that their scope has on the text's final polarity. Sentiment analysis models are evolving rapidly, with advancements ranging from state of the art data preprocessing techniques, machine learning, deep learning and transformers. Despite these advancements, the current models still face significant hurdles in accurately detecting polarity shift due to implicit negation, as their ability to handle nuanced shifts remains limited. Polarity shifts due to implicit negations present a persistent challenge in sentiment analysis, as they change the accurate interpretation of sentiment in written text. In light of this, this research is based on a pragmatism philosophy; firstly the current sentiment analysis models and the data preprocessing techniques are evaluated. This provides key findings and a conceptual framework that acts as the road map for the subsequent empirical analysis process. The empirical analysis process is guided by a comprehensive methodology based on mixed research methods for triangular validation. Experiments are conducted on the sentiment analysis models that demonstrate methodological rigor and novelty, in a variety of state of the art feature extraction, feature selection and data augmentation techniques. The invaluable insights based on the results’ analysis of the experiments of the state of art model that included SVM, Naïve Bayes, LSTM and BERT models, evinces the significance of the data augmentation process as a critical leverage point in the effective detection of polarity by the superior deep learning models. This observation guided the implementation of a hybrid deep learning model, through a multi-methodological model development methodology. The novelty of the proposed model lies in its comprehensive integration of Seq2Seq perturbations and contextual back translation for data augmentation, combined with FastText and BERT embeddings, and attention mechanisms, enabling it to effectively address polarity shifts due to implicit negation in sentiment analysis, particularly in mobile phone brand reviews. This model was then validated against the existing state of the art models, using similar metrics, in order to evaluate its effectiveness. Statistical significance tests, using the ANOVA and the Kruskal-Wallis H statistic, demonstrated that the new model is superior. This was also evinced by the improved average performance scores, attention heat maps as well as the presence of shorter whiskers and higher median (Q2) values in the whisker-box plots, in a variety of mobile phone reviews datasets. Future work could explore on combining the seq2seq perturbations on back translation with multiple data augmentation techniques to further enhance the model's ability to detect implicit negation and other nuanced shifts in sentiment. Moreover, future research could also explore methods for optimizing the training and inference efficiency of the hybrid model without sacrificing its efficacy. Techniques such as knowledge distillation, pruning, or quantization could help make the hybrid model more computationally efficient, enabling real-time sentiment analysis in resource constrained environments.