Bypassing Deep Learning based Sentiment Analysis from Business Reviews
Ashish Bajaj, Dinesh Kumar Vishwakarma · 2023
In recent years, online reviews of businesses have grown increasingly significant, as customers and even competitors use them to evaluate a company's quality. Yelp is one of the most popular review websites, and it would be advantageous for them to be capable of predicting the sentiment or even the star rating of a review. Current deep-learning algorithms excel at sentiment classification. With the tremendous performance of models based on deep learning in text-related problems, they are susceptible to adversarial manipulations that result in inaccurate sentiment classification. An adversarial text is created by manipulating just few letters or words in such a manner so that general meaning of the text remains unchanged for humans but fooling a system into making false predictions. This study highlights the shortcomings of sentiment categorization by employing a range of cutting-edge attack techniques to generate perturbed text. We examined the performance of several models, including BERT, an advanced transformer model, and the extensively used LSTM and Word-CNN classifiers trained on the Yelp polarity dataset. For each model, Attack Success Rates (ASR) are calculated as the evaluation metric. Based on the experimental results, we determined which sentiment classifier is more vulnerable to adversarial perturbations and which is more resistant. The results demonstrate that automatic sentiment classification techniques can be circumvented, which has implications for present policy approaches.