Fine- Tuning Pre-trained Models for Aspect-Based Sentiment Analysis of Amazon Food Reviews

The process of classifying consumer responses to a particular product review as neutral, negative, or positive is known as sentiment analysis. Product owners can utilize sentiment analysis for assistance in them decide whether to continue producing their product, increase production, or eliminate production altogether. As a result, they will gain a deeper comprehension of how customers feel about their products. The method used to determine the opinions expressed in the review regarding a specific feature or aspect of the product is called Aspect Based Sentiment Analysis, or ABSA. Numerous machine learning and natural language processing models are currently available for ABSA. The inability to precisely identify and categories the views expressed towards particular features is a major drawback of the aspect-based sentiment analysis (ABSA) techniques presently in usage for Amazon food assessment tasks. The Hybrid Contextual Aspect-Based Sentiment Analyzer (HCABSA), a hybridized framework for aspect-based sentiment analysis, is developed in this research to accurately assess the sentiment of Amazon food reviews through a number of processes. This research starts by collecting dataset for Amazon food reviews from Kaggle, preprocessing the data to eliminate noise and irrelevant information, and then using POS tagging to identify the specific aspects and sentiment words related to those aspects, which are nouns and adjectives in the review using the Lexicon based approach. In the following stage, the frequency-based candidate extraction method is utilized to pick the characteristics that are used to determine the product's internal and exterior features. The raw data is then transformed into features that the machine learning models can access using the Bag of Words feature extraction process. The Gradient Boosting Decision Tree technique is then used to categories the sentiment as either positive or negative. Using performance metrics like accuracy, precision, recall, and f1 score, the model's output is compared to existing algorithms and the comparison clearly indicates that the proposed approach provides the maximum accuracy of 98%.