Optimal trained long short-term memory for opinion mining: A hybrid semantic knowledgebase approach

Abstract

This study describes a Hybrid Semantic Knowledgebase-Machine Learning strategy for mining the domain feature-level opinions and categorizing them on a multi-point scale. The proposed work is constructed by following the six major stages: “(a) Pre-processing, (b) Domain Feature Extraction, (c) Sentiment Extraction, (d) Domain Feature-Sentiment Association, (e) Domain Feature Polarity, (f) Knowledgebase Enrichment and (g) Opinion Classification with optimized deep learning model”. The proposed method is divided into several levels, each of which is designed to solve the opinion mining challenges at “domain feature level”. Initially, from the collected reviews, the domain features as well as semantics are extracted. Subsequently, in the Domain feature-sentiment association stage, the extracted filtered domain features are associated with their corresponding extracted sentiments acquired from the Sentiment Extraction phase. Once, the sentiments and the features are associated, the polarity of the feature sentiment pairs are computed using the sentiment aggregation function, which assigns a score to the features based on its proximity. Then, in the Knowledgebase enrichment stage, the initial review knowledgebase was used to bootstrap the domain feature extraction process, which is further enriched with new semantic information related to the analyzed review and the corresponding extracted domain features. Eventually, in the Opinion classification phase, a new optimized deep learning classifier, the optimized Long Short-Term Memory (LSTM) is introduced. The optimized LSTM is trained with the semantic information acquired from the enriched knowledge base and the statistical features created by Vector Space Model. Moreover, to enhance the classification accuracy of the LSTM, an Improved Butterfly Optimization Algorithm is introduced to fine-tune the weight of the model. Finally, the performance of the proposed work is evaluated over the conventional models.