Optimizing Features using BGWO for Aspect Term Extraction

Abstract

There are two major tasks in the area of sentiment analysis- aspect term extraction and determining the sentiment polarity of extracted aspect terms. One of the major challenges in the domain of aspect term classification is selection of optimal features. In this paper, a binary variant of Grey Wolf Optimization (BGWO) is employed for selecting the most valuable features from the identified linguistic features for classifying the token as an aspect term. The process of GWO starts by selecting three best solutions from the binarized population of grey wolves using the fitness function that maximizes the accuracy and minimizes the feature set size. The position of grey wolves is updated using the stochastic crossover operation on the identified three best solutions. To evaluate the performance of the proposed method for optimal feature selection, an experimental study is conducted on two SemEval datasets of restaurant and laptop reviews using three different classifiers- Logistic Regression, Naive Bayes and Support Vector Machine. The results depict that model could achieve an average f1 score of 86.3% and 71% on laptop and restaurant domains respectively using the optimal feature set. The proposed work is further extended to generate an opinion-based summary of domain-specific aspect terms extracted from the review documents using VADER. VADER is a popular rule-based sentiment analyser that does not require prior training and performs efficiently on social media data. The performance analysis of Sentiment Analysis on few test samples of laptop domain indicates the efficacy of the proposed work.