Introduction

E-commerce refers to the purchase and sale of products and services through the internet. It contains a large number of data, processes, and tools for customers and sellers, such as smart device shopping, cash on delivery, and online payment encryption [1]. According to a research report, due to the covid-19 pandemic, online sales have increased [16]. Customers are accustomed to submitting reviews or comments after receiving deliveries to share their opinions about the quality of the products and services. However, in the instance of online buying, e-commerce service providers go to great lengths to assist users in selecting trustworthy products. Various businesses, particularly e-commerce, heavily rely on sentiment analysis to boost product quality and make the right business decisions in today's competitive business world. E-commerce companies want to know what their customers think about their sellers and products that help them to maintain their online reputation [13]. Shoppers can easily get ideas about which product will be best for them compared to other products using this active feedback information [14]. Sentiment analysis is a procedure of gathering customer feedback about a product and categorizing it into distinct polarities using contextual data [15]. The task of classifying text is known as polarity classification. Texts have evolved into a treasure trove of important information and people's perspectives on a variety of topics [11]. 

Consumers value other people's opinions and experiences and reading a review on a product is the sole method to learn what other customers think about it. Opinions derived from consumers’ experiences with certain products have a direct impact on future customer purchases [1]. Negative and slightly negative ratings, on the contrary, frequently result in sales loss. The target for entrepreneurs who want to be successful in business is to understand their clients' input and polarize correctly over a large amount of data [9]. Some research has been completed only in positive, negative, and neutral classes. Research deals with some human sentiment that is not related to e-commerce business [11]. To establish a better e-commerce system it is necessary to work on all types of human sentiment based on consumers. There is no research completed on five (Negative, Positive, Neutral, Slightly Negative, and Slightly Positive) types of human sentiment that can be established a better electronic commerce business system. This core inspires us to work on it. 

The fundamental purpose of this research is to improve the limitations of datasets and existing preprocessing and feature extraction techniques on product reviews in Bangla, English, and Romanized Bangla. Different types of customer reviews are present in the e-commerce review section but working with only a selected language and labeling data according to very fewer human sentiments can never perfectly contribute to developing a better e-commerce system. So our main goal is to polarize all types of reviews in the review section. Opinion mining of reviewers is the procedure of gathering information from the internet to determine customers’ thoughts, hence we labeled our dataset into five classes [13]. Our research presents an ML-based model for sentiment identification in e-commerce product evaluations for three languages. Working with a new domain of three different languages, five types of human sentiment is the main contribution of our work and also some previous research in the same area has been improved by the proposed technique. 

Related Works

Sentiment analysis is now an active field of research. Many researchers are performing research on sentiment analysis. Some researchers performed sentiment analysis on product review data of different e-commerce websites in various languages. Some of the previous work related to our research are discussed. In this section, we have summarized past findings and have shown the limitations and how these limitations are handled by our proposed technique. 

In [1], authors used amazon reviews data to perform research only in English language and applied six machine learning algorithms, where Linear SVM achieved the highest accuracy of 94.02%. By observing the amazon review section, we noticed that customers express a variety of sentiments in their reviews. According to positive and negative sentiments, they categorized their data into only positive and negative. In this research, we made an addition of product review sentiment analysis in three different languages and have categorized our datasets into negative, positive, neutral, slightly negative, and slightly positive. 

Shafin et al. [2] used 1020 reviews data and applied five machine learning algorithms. They got the highest of 88.81% accuracy in SVM. They used only Bangla positive and negative data in their research whether in our research we have used three different product reviews. Aspect-based sentiment analysis is performed by Satuluri Vanaja and Meena Belwal [3]. They used amazon reviews and got the highest 90.423% accuracy in the Naive Bayes algorithm. They labeled their data into three types of sentiments and did classification using only Naive Bayes and SVM but in our research, we have applied six machine learning algorithms that have got better accuracy in SVM and Random Forest. 

In [4], the authors used IMDb, blogs, and social media data, to research sentiment analysis whereby applying machine learning algorithms and got the highest of 83% accuracy which is not better than our research's best-obtained accuracy. They labeled their dataset into only two classes. Mst. Tuhin Akter, Manoara Begum, and Rashed Mustafa performed research on Bangla product reviews where they collected data from ‘Daraz’ [5]. They applied five machine learning algorithms and got the highest of 96.25% accuracy in KNN. They collected only Bangla reviews from daraz and labeled their dataset into three types of sentiments which are less than our research. 

A feature-based opinion mining was performed where they collected their data from kaggle.com [6]. Their proposed approaches were analyzed by SVM and Random Forest, where Random Forest gave the best accuracy of 97%. Their used dataset had only negative, positive, and neutral classes. There are lots of fake reviews given on any e-commerce platform. In [7], their proposed approach covered three basic phases to get a model for fake review detection. They applied five machine learning algorithms and achieved 87.87% accuracy in Logistic Regression which is not greater than the best-obtained accuracy of our research. Comparative sentiment analysis of sentence embedding was performed by Poornima A. and K. Sathiya Priya [8]. They used tweeter data and applied MNB, SVM, and Logistic regression. They got the highest of 86.23% accuracy in Logistic regression which is 7.77% less than our best-obtained accuracy. In [9], they completed research on product reviews using machine learning approaches, where they used only 900 daraz reviews data which is a very small dataset. By applying various machine learning algorithms they got the highest 78% of accuracy in Ridge Algorithm and the lowest accuracy is 66% in Random forest which can be improved more. Character-level supervised RNN approach was proposed where they developed a deep learning model [10]. They collected data from social media platforms and applied recurrent neural networks (LSTM, GRU) for training their model and got the highest of 80% accuracy. Though the above-mentioned works are compatible, there are still some lacking in some perspectives according to the sentiment analysis of the e-commerce review. The best of our investigation, there is currently no work presented in the e-commerce review from Romanized Bangla texts, it is one of the novelties of our research. Our research contains three different datasets. We have labeled our datasets into five types of human sentiments (negative, positive, neutral, slightly negative, and slightly positive) which is another novelty of our research. 

D. Oversampling

An imbalanced training data set will certainly affect prediction accuracy in sentiment analysis [5]. The term "class imbalance" refers to a significant disparity in class sizes. In this circumstance, any classifier becomes biased in favor of the dominant class, ignoring the minority class, resulting in incorrect predictions. To overcome this limitation, we applied random oversampling techniques and made all minority classes equal to the majority classes. When the data amount is insufficient, oversampling is a strategy for resampling the data to address imbalance dataset difficulties. By creating additional samples, it enlarges the minority class dataset. After applying the random oversampling technique dataset has a total of 15630 review data. There are 8785 Bangla Data, 4440 English, and 2405 Romanized Bangla Data. Fig.2 demonstrated the datasets before oversampling, and Fig. 3 demonstrated the datasets after oversampling.

E. Feature Extraction

To increase classification efficiency and accuracy, the technique of projecting greater feature space into lesser feature space is known as feature extraction. [5]. It is basically a process to remove unnecessary data elements that may impact the model's accuracy. To decrease the dimensionality of the classification algorithm, it is necessary to limit the number of resources used to describe a large set of data by developing new features from a combination of the genuine set [18]. It has been established that the key to good model creation is optimal feature extraction. To solve the challenges, we have used the TF-IDF feature extraction technique. We applied the feature extraction technique TF-IDF in machine learning algorithms. TF-IDF converts strings to numbers in order to provide numerical data to ML models[12]. 

F. Splitting the dataset into train and test

We split all three datasets into 80% for training and 20% for testing, respectively. After preprocessing and oversampling all three datasets and performing feature extraction, Bangla Dataset had 7028 texts for training and 1757 for testing after it was separated. English Dataset has 3552 texts for training and 888 texts for testing. 1924 texts were placed to use in training and 481 for testing in the Romanized Bangla Dataset. The models are trained with training sets and then tested with testing sets. 

G. Machine Learning Algorithms

To anticipate categorical labels, machine learning-based classification algorithms are widely utilized [12]. We have used supervised learning because in additionally it is a classification model problem. Although there are various machine learning classification methods, we have employed six in our study: Multinomial Naive Bayes, Support Vector Machine(SVM), Logistic Regression, Random Forest, k-Nearest Neighbors(KNN), and Decision Tree. 

H. Performance Analysis

Our models were evaluated using the testing datasets, and the parameters were assessed after they had been trained. The performance was assessed using the confusion matrix, accuracy, precision, and recall, f1-score, and ROC area. The confusion matrix is a set of four expected and actual values used to assess the effectiveness of machine learning classification techniques. [12]. In Fig. 4. True Positive, False Positive, False Negative and True Negative are denoted by the letters TP, FP, FN, and TN, respectively. The confusion matrix is critical for determining accuracy, precision, recall, and, most significantly, the ROC area. The accuracy, precision, recall, and f1-score are calculated using (1), (2), (3), and (4) [17]. 

Fig. 5, Fig. 7, Fig. 9 show that for Bangla Dataset, Support Vector Machine, and for English and Romanized Bangla Dataset, Random Forest was the most accurate algorithm and classified maximum negative, positive, neutral, slightly negative, and slightly positive labeled testing data correctly. Furthermore, in Fig. 6. Support Vector Machine occupied more ROC area for every class in Bangla Dataset, in Fig. 8. Random Forest occupied more ROC area for English Dataset, and in Fig. 10. again Random Forest occupied the more ROC area in Romanized Bangla Dataset. Finally, more detailed results of all machine learning algorithms respectively shown on all datasets in TABLE II, TABLE III, and TABLE IV.