Enhancing Sentiment Analysis via a Synonym-Aware Objective Function in Deep Hybrid Framework

Our current explosion in internet-based technologies has given rise to user-generated content, making Sentiment Analysis an essential field within Natural Language Processing. It focuses on interpreting and categorizing opinions expressed in text. Traditional approaches include lexicon-based, rule-oriented, and popular machine-learning techniques. Recent research highlights the success of deep network modeling of CNNs and LSTMs for capturing complex semantic patterns within machine learning-dependent approaches. These neural networks achieve much better performance when functioning within a carefully optimized feature space. Evolutionary algorithms have proven to be highly effective and are extensively utilized to accomplish this optimization. Beyond conventional techniques, hybrid approaches that blend different methodological elements are receiving increasing attention, owing to their ability to deliver higher accuracy and better adaptability in sentiment analysis tasks. Previous research introduces a Hybrid Deep Neurally Connected framework for classifying sentiment, in which an evolutionary technique Dispersive Flies Optimization (DFO), has been enhanced to Neighbour Adjusted Dispersive Flies Optimization (NADFO) by modifying neighboring agent behaviors. This optimization technique refines the feature space using a custom fitness function integrated with the SentiWordNet lexicon, improving the exploration of the solution space and avoiding local optima. Textual features are then transformed into meaningful numerical vectors using pre-trained embeddings and fed into a hybrid CNN-LSTM model to get the sentiment category. In our work, this SentiWordNet-based fitness function is further modified and improved by incorporating synonym information, moving beyond conventional POS-tag-based synset selection. Experimental results demonstrate that these improvements lead to superior sentiment classification performance compared to the baseline.