Optimized neural attention mechanism for aspect-based sentiment analysis framework with optimal polarity-based weighted features
Mekala Ramasamy,Mohanraj Elangovan
DOI: https://doi.org/10.1007/s10115-023-01998-0
IF: 2.7
2024-01-04
Knowledge and Information Systems
Abstract:In recent years, sentimental analysis has been broadly investigated to extract information to identify whether it is positive, negative or neutral. Sentimental analysis can be broadly performed in social media content, survey response and review. Still, it faces issues while detecting and analyzing social media content. Moreover, a social media network contains indirect sentiments and natural language ambiguities make it complicated to classify the words. Thus, the aspect-based sentiment analysis (ABSA) is emerged to develop explicating extraction methods by utilizing the syntactic parsers to make use of the relation among sentiments and aspects terms. Along with this extraction method, the word embedding is performed through Word2Vec methods to attain a low-dimensional vector depiction of text, which could not capture valuable information. Thus, it aims to design a novel ABSA model using the optimized neural network along with optimal text feature extraction. Initially, various data is collected through the benchmark dataset are given to the image pre-processing. Then, it might undergo different techniques like stemming, stop word removal as well as punctuation removal. Then, the preprocessed data are further given into the feature extraction phase to attain adequate extracted aspects. Then, it further undergoes for deep feature extraction stage, where the text conventional neural network and Glove embedding are utilized to obtain the deep features. Further, the feature concatenation is done to attain the optimization for polarity-based weighted features utilized by the enhanced hybrid optimization algorithm called hybrid Chameleon rat swarm optimization (HCRSO) for improving the performance in sentiment analysis. The optimal features are selected by the HCRSO that provides the polarity-based-weight features; thus, it separates the polarity, and the weighted features are occurred by multiplying the weight with polarities. Especially, the optimized features of polarity-based weighted features and also the parameters of epochs and hidden neuron count of neural attention mechanism-based long short-term network (NAM-LSTM) are optimized using the HCRSO algorithm. The weighted feature is applied by incorporating the NAM-LSTM and proposed HCRSO algorithm for improving the model efficiency. The empirical outcome of the recommended method shows 94% and 93% regarding accuracy and specificity. Thus, the experimental outcomes of the proposed ABSA model reveal the model's efficiency while validating with other conventional approaches.
computer science, information systems, artificial intelligence