Sentiment analysis is a text-classification task aimed at evaluating the polarity (positive or negative) of the opinions expressed within a single document. It serves to analyze and identify the type of sentiments in text-based user reviews [1]. With the development of internet and social media, people use social sites, websites, blogs, and forums more often as the primary mediums for expressing their comments, views, opinions on services, issues, ideas, and various other things [2]. Thus, sentiment analysis plays an important role in analyzing reviews, views, and opinions presented by users in order to evaluate the feedback on a specific aspect [3]. The results of this analysis are very useful to service providers in improving the quality of their respective services. Indirectly, these results can also help service providers increase the profitability of their respective organizations.

Sentiment analysis originates from text mining technology, natural language processing, and text classification methods [4]. It helps classify text-shaped information such as customer reviews of services, politics, products, and others such as positive or negative sentiments [5–8]. Sentiment analysis can be divided into three techniques: lexicon-based techniques, machine learning-based techniques, and hybrid techniques [9]. Machine learning technique is one of the most popular methods for classifying sentiments. This technique is combined with language methods to identify the categories of sentiments present in a text [1]. Three main processes in sentiment analysis were implemented in this study, namely data processing, feature selection, and sentiment classification.

One of the key problems in the study of sentiments is managing user feedback on social media, websites, forums, and blog texts that contain noisy and unimportant details. Such data need to be managed and processed regularly before the performance of sentiment analysis can be carried out [10]. From a study conducted by [10], it was found that text-processing techniques play an important role in ensuring the accuracy of sentiment prediction and also in influencing the computation time during sentiment classification. Text processing was developed to analyze and prepare the text for sentiment classification [11].

In addition, the presence of large feature size is a common problem in sentiment analysis. Large feature size affects the performance of sentiment classification and its processing time. However, this problem can be overcome using machine learning methods that use appropriate feature set selection techniques to eliminate noisy and irrelevant features [1,5,12–16]. Feature selection is a volatile polynomial problem and requires efficient algorithms such as metaheuristic algorithms to assist [17–19].

According to [20], most of the existing feature selection algorithms require specific parameters. Algorithms that use specific parameters need to adjust the value of those parameters to an exact value in order to get good results. For example, genetic algorithms require specific parameters such as cross-probability, mutation rate, and selection method parameters. Accordingly, [20] introduced the teaching–learning-based optimization (TLBO) algorithm that does not require specific parameters to function properly.

Typically, the reviews in datasets containing user reviews have a lot of noise and irrelevant information that makes it difficult to process and analyze information accurately and in detail for identifying customer sentiments regarding service quality and the types of services preferred and disliked by customers. This information that is obtained from social sites and the internet is constantly evolving and changing over time, which renders the manual analysis of reviewing information to become ineffective, irrelevant, and one that requires a lot of effort, often resulting in inaccurate findings. This is a major problem faced by service providers in identifying customer sentiments towards the services and customer feedback provided as a whole. Therefore, a mechanism needs to be developed to analyze user reviews to identify customer sentiments. Therefore, in this study, we have identified a great opportunity to develop a feature selection and feature reduction algorithm for sentiment classification. For this, we used a combination of feature selection based on RST and TLBO, which is referred to as RSTLBO, to reduce the dimension size of features and select a subset of quality features for sentiment classification. The following contributions made by this paper have been summarized as follow.

The following sections in this paper present the following aspects: Section 2 describes the studies conducted on the topics discussed. Section 3 describes the research methodology. Section 4 discusses the results of this study. Lastly, Section 5 summarizes the results of this study.

Text processing and feature selection are two very important processes in sentiment classification systems. Text processing involves preparing a dataset with user reviews that were obtained from social sites, websites, forums, and more. This dataset of reviews usually contains a lot of noise, irrelevant, and meaningless fragments such as HTML tags, scripts, advertising, and other such things. Therefore, these reviews need to be cleaned and converted into a dataset format to prepare them to be processed for feature selection and sentiment classification through text processing. The feature selection process helps reduce the dimensions of the dataset generated from text processing. Typically, this user review dataset contains features of various sizes and is relatively difficult to be processed manually. In order to tackle this difficulty, it is processed through computing operations using feature selection algorithms. These feature selection algorithms produce small-dimensional datasets that can be used in the sentiment-classification process. Various feature selection algorithms have been developed, but most of them are recommended for studies in the field of engineering and other fields. Very few studies have been conducted on sentiment analysis.

Manimekalai et al. [21] described sentiments as opinion mining, which is the field of studying customer reviews and opinions, ideas, evaluations, human emotions, or attitudes towards different aspects such as services, organizations, products, events, individuals, and so on. Reference [22] used the example of a dataset of text documents with user reviews on a topic and cited that the main purpose of sentiment analysis is to identify and evaluate the features of the comments on an object within a document. Reference [23] explained that sentiment analysis works by extracting sentiments or opinions from textual documents. It is based on text-mining technology, natural language processing, and text classification. It involves the study of the relationship between text-shaped data through text processing. It serves to categorize text-shaped information such as consumer reviews of services, politics, products, and others into positive or negative sentiments [5,6]. Sentiment analysis can be divided into three techniques—lexicon-based sentiment analysis, machine learning-based sentiment analysis, and hybrid techniques, as shown in Fig. 1.

The machine learning technique is one of the most popular methods of classifying sentiments. This technique is combined with language methods to categorize sentiments within the text [1]. Three main processes in sentiment analysis were implemented in this study; namely data processing, feature selection, and sentiment classification. Each process plays an important function in improving the performance of sentiment classification and also helps to improve the accuracy of the results of the sentiment analysis. Before performing sentiment analysis, it is important to identify the data sources; whether they were obtained from social media, reports, documents, databases, or others. These data sources contain hidden information that needs to be analyzed to uncover information hidden in the information. Most data derived from social media contain noise and irrelevant data, as well as meaningless information [24]. These elements make the process of classifying sentiment on information obtained from social media very challenging. Accordingly, to enable the smooth processing of this information, it needs to undergo a transformation process, i.e., a process of purification and analysis called text processing [25]. The data that has gone through this process can then be used for the feature selection method, and then the sentiments with the information can be classified.

The rough set theory (RST) technique, which was introduced by [26], comes from simple model information. RST is considered a new mathematical tool for processing uncertain information after the uncertainty theory [27] and fuzzy set theory [28] for data analysis and data processing. RST has been identified as an effective mathematical method of minimizing data from information systems as well as a mathematical tool for solving uncertain problems [29]. RST has well-structured procedures, algorithms, and specialized tools to identify patterns and has the ability to find valid reductions, i.e., find the minimum set of features using the reduct function.

Teaching-learning-based optimization (TLBO) is similar to other algorithms inspired by natural situations in learning sessions within classrooms [20]. It is a population-based metaheuristic algorithm that uses population solutions to find global solutions. In TLBO, a population is a group of students or a full class. The process of TLBO can be divided into two sessions: the first session is the teaching phase, while the second session is the learning phase. The teaching session involves learning from the instructor, and the learning session is related to the learning process between students and students. This TLBO method is based on the effect of the teacher’s influence on student output; in this case, student outputs correspond to student results or performance. Instructors are generally considered to be highly educated individuals and will share their knowledge with their students. The quality of instructors will affect the quality of students’ performance. Thus, it can be said that quality teachers will produce quality students. TLBO is carried out to obtain a global solution without the need for burdensome computational efforts and high consistency [20].

There are three types of classification levels in sentiment analysis—document, sentence, and feature levels. At the document sentiment classification level, each document is classified according to positive and negative groups [23,30]. Classification of sentiments at the sentence level categorizes sentences into positive or negative groups. Finally, the classification of sentiments at the feature level helps identify sentiments behind each word based on the elements and features found in the review sentence, for example, as per the studies conducted by [31,32]. One of the challenges in identifying the relationship of sentiment words with a feature is that a sentiment word can have different meanings depending on the circumstances and fields in which it has been used. The classification at the feature level is very complicated as it entails clear identification of the relationship between the feature and the sentiment word in the review sentence. Despite these challenges, the result of this classification is more accurate and complete as compared to the classification at the document and sentence levels [33].

The research methodology consists of four phases starting with text processing phase, feature selection algorithm development phase, classification of sentiment phase; and testing, evaluation, and analysis phase, as shown in Fig. 2. The benchmark dataset used in this study was obtained from the Kaggle website. The dataset includes user reviews of six major U.S. airlines obtained from Twiter in February 2015. It contains 14,640 reviews, 2,363 of which have been labeled positive, 3,099 as neutral, and the remaining 9,178 as negative.

This section describes the results of this study on the selection of text processing methods and the selection of features using the RSTLBO algorithm.

This study shows that the spelling correction through natural language processing in the text processing method, combined with natural linguistic processing technique, has a more significant effect on the accuracy of sentiment classification as compared to other natural language processing methods. Moreover, this study also shows that now the RSTLBO algorithm can perform feature selection for classification sentiment better or on par with the existing feature selection algorithms with a classification accuracy of 80.2%, precision of 83.6%, and retrieval rate of 80.2%. Furthermore, this algorithm managed to reduce the feature size by 98.9%, which is 95 features from the total of the original 8,424 features.

Furthermore, this study’s results highlight that the rate of classification accuracy using a combination of text processing methods and RSTLBO feature selection algorithm at 80.2% is higher than that of the sentiment classification with just text processing method resulting in 76.9% accuracy. This is because RSTLBO can reduce the dimensions of the feature set and select a feature set that has a good quality to be used in sentiment classification. Statistical relevance of the experimental results is determined by the paired t-test based on ten-fold cross-validation and the results demonstrate the RSTLBO is statistically significant and outperformed other method.

This study also demonstrates that the text processing methodology has a very significant effect on sentiment analysis, as text processing immensely helps in eliminating noisy and irrelevant information from the data and, as a result, can generate relatively good classification results. If this noisy and irrelevant information is not eliminated, the feature selection process will be difficult, requiring high processing resources and indirectly negatively affecting the accuracy of sentiment classification. Thus, the feature selection algorithm RSTLBO helps to improve the accuracy of sentiment classification, based on the results of text processing, to a higher level.