Understanding Quality of Products from Customers’ Attitude Using Advanced Machine Learning Methods

Abstract

The trend of E-commerce and online shopping is increasing rapidly. However, it is difficult to know about the quality of items from pictures and videos available on the online stores. Therefore, online stores and independent products reviews sites share user reviews about the products for the ease of buyers to find out the best quality products. The proposed work is about measuring and detecting product quality based on consumers’ attitude in product reviews. Predicting the quality of a product from customers’ reviews is a challenging and novel research area. Natural Language Processing and machine learning methods are popularly employed to identify product quality from customer reviews. Most of the existing research for the product review system has been done using traditional sentiment analysis and opinion mining. Going beyond the constraints of opinion and sentiment, such as a deeper description of the input text, is made possible by utilizing appraisal categories. The main focus of this study is exploiting the quality subcategory of the appraisal framework in order to predict the quality of the product. This paper presents a quality of product-based classification model (named QLeBERT) by combining quality of product-related lexicon, N-grams, Bidirectional Encoder Representations from Transformers (BERT), and Bidirectional Long Short Term Memory (BiLSTM). In the proposed model, the quality of the product-related lexicon, N-grams, and BERT are employed to generate vectors of words from part of the customers’ reviews. The main contribution of this work is the preparation of the quality of product-related lexicon dictionary based on an appraisal framework and automatically labelling the data accordingly before using them as the training data in the BiLSTM model. The proposed model is evaluated on an Amazon product reviews dataset. The proposed QLeBERT outperforms the existing state-of-the-art models by achieving an F1macro score of 0.91 in binary classification.

1. Introduction

Online shopping and the Internet have dramatically changed traditional offline businesses into web-based ones using platforms like social media, search engines, and e-commerce sites like Amazon, Shopify, eBay, Etsy, WooCommerce, etc. Customers can effortlessly place their orders to purchase products and receive them at home. Online shopping has many advantages, e.g., fast access to various items, a better selection of items, detailed information about products, etc. The problem that arises in such shopping is that customers need to learn about the actual quality of the products. Business proprietors have created review or feedback systems for prior customers to cope with such issues. Feedback from customers is collected in different ways, e.g., through surveys on the social web, live chat, offering incentives to customers on their responses, creating a feedback system on the website, etc. The focus of this study is to identify product quality from customers’ attitude using advanced machine learning methods. Attitude is a somewhat durable organization of beliefs, thoughts, and behavior proclivities in connection with critical social elements, groups, occurrences, or symbols [1]. The appraisal framework examines the attitude of people in a written text. The attitude system in the appraisal framework has three main subdomains: appreciation, affect, and judgment. Affects are emotion-based; judgment deals with the evaluation of persons to extract their opinions, and appreciation deals with the extraction of opinions related to physical things, processes, etc. Appreciation is further divided into three main categories, i.e., reaction, composition, and valuation. Impact and quality are associated with the reaction. The impact is associated with opinions regarding the attractiveness and unattractiveness of things. Quality is related to the liking and disliking of things, e.g., beautiful, elegant, hideous, etc. Composition is further divided into two main parts, i.e., complexity and balance. Complexity is about opinions containing complications and the simplicity of things. Balance deals with ideas like balance and unbalanced things. Valuation, as its name implies, contains opinions related to concepts like the innovativeness, uselessness, cost, shoddiness, etc., of things [2]. Quality means the features of a particular product that satisfy customers’ needs. Quality assessment enables stakeholders to improve the standard of their products and make them more marketable, compete more effectively, gain market share, and generate sales revenue. Higher quality enables businesses to lower error rates, rework, field failures, warranty costs, customer unhappiness, and other costs. It also reduces the time to launch new products [3]. This study exploits an appraisal framework to prepare the quality of the product-based lexicon dictionary and then employs a BERT word embedding model and BiLSTM model to detect the quality of the product from customers’ attitudes in product reviews.

The main contributions of this study are the following:

• The development of annotation guidelines for a lexicon dictionary to identify product quality based on an appraisal framework.
• To the best of our knowledge, the proposed model is the first to identify the quality of a product from customers’ attitude based on an appraisal framework using a lexicon approach with N-grams and the utilization of a pre-trained BERT word embeddings model in combination with BiLSTM.

This paper is organized as follows: Section 2 concerns related work. Section 3 consists of the methodology. Section 4 is about the experiment. Section 5 comprises the results and discussion, while Section 6 contains the conclusions and future work.

2. Related Work

Electronic word-of-mouth is about the sharing of information by customers unceremoniously on the internet with respect to particular features of products and services, or about the vendor [4]. Customers, in their reviews, state their knowledge, approval, and opinions about goods or services on different web-based platforms, e.g., social webs, blogs, etc. [5]. Online reviews form an opinion and provide useful information that can be used to evaluate potential purchases [6]. The primary goal of natural language processing is the automatic processing of text to assess peoples’ attitudes, feelings, etc. Opinion mining involves the extraction of sentiments and the topics discussed within the text. Based on online reviews, the author in [7] employed text mining techniques to predict the positive and negative attitudes of consumers toward the hotel. Important features related to the positive and negative attitude of consumers were extracted. These critical features significantly assisted the marketers in planning keyword selection in their marketing policies. According to the author in [8], development of business is mainly based on recommendations which are the sole predictors.

In the age of e-commerce, online consumer reviews have become very helpful because they influence upcoming purchases. E-commerce sites frequently offer product review ranking services to assist customers in making decisions and solve the concerns about the quality of online goods. The quick emergence of social networks and the corresponding increase in data generation mean that sentiment analysis is no longer a static activity, and the methodologies that have emerged can be broadly divided into the lexicon-based approach [9] and machine learning approach [10]. The authors in [11] investigated the selection of online products by consumers using a web-based product recommendation system. They discovered that customers checked the recommendations from a web-based system for product recommendations twice as often as consumers who did not.

On the web, there is a wealth of information about consumer reviews. Consumer recommendations are also included in these reviews. These textual reviews provide information on the money-spinner and display a set of assessment variables regarding the presence or absence of recommendations. With the vast amount of information available on the internet, data mining techniques are used to extract hidden information about customer behavior [12,13]. The number of features that are included in a product that satisfy the customers’ needs and how those features are altered to conform to those expectations are considered a product’s quality. The authors presented machine learning techniques for extracting different product parameters that are available across various sources in an unstructured manner to improve product quality [14]. In [15], the authors proposed text mining to assess product reviews while taking into account the validity of each review in order to develop a trustworthy evidence-based strategy for online product evaluation. The authors in [16] used sentiment analysis to gather consumers’ reviews and predicted the product’s rating. In [17], the author suggested different aspects of a product, i.e., internal features, external features, industry standards, reliability, lifetime, services, customers’ response, exterior finish, and past performance of the product. Udeh et al. [18] employed a survey method and examined the hypothesis with the help of multiple regression analysis to check the influence of product quality on Pay TV customer happiness. Their research study concluded that customer satisfaction with Pay TV positively and significantly correlates with reception quality, content quality, and customer service.

The author in [19] used Twitter data to assess the consumers’ sentiments toward well-known brands. In [20,21], the authors obtained product features and envisioned market formation by using text mining techniques. Text mining techniques were also used by the authors in [22,23] to supplement numerical data in order to forecast product sales. The approach in [24] evaluated new quality value by collecting web-based customers’ comments about shopping and then implementing text classification techniques with a fuzzy comprehensive evaluation method. This evaluation procedure aided consumers in making better decisions when purchasing suitable products. The author in [25] prescribed innovative architecture for inventors by utilizing web-based reviews for product design based on reality-based online item reviews of a sample product in terms of precision, comparison, and rationality. In [26], the authors tried to get some critical information from asset reviews to improve the features of an asset and aid in improving customer service and knowledge. The authors used text mining algorithms to examine customer reviews of a product to identify issues that frequently cropped up and how they tended to develop over time. Cruz [27] investigated the relationship between the quality of products and the satisfaction of customers. Martinet et al. [28] explored social media contents with the help of an appraisal framework.

Social media is a significant source of information. To analyze its contents is a very laborious task through traditional data mining tools because their contents are large, noisy, unstructured as well as not being similar to each other. Conventional tools of data mining are slow, costly, depending on their size, and are biased as well [29]. Mining social media contents deals with people’s opinions, attitude, and emotion identification. Social media contents mining revolves around two main concepts, i.e., sentiment analysis and opinion mining. The word “sentiment analysis” was first introduced in [30], and the term “opinion mining” was initially used in [31]. Although there is major disagreement about the limitation between these two fields. Authors also believe that text mining techniques are used under opinion mining to discover exciting and intuitive correlations among opinions of authors in [32], while sentiment analysis is known as sentiment classification, which is related to the categorization of a text, or part of the text based on computing the amount of the individual’s opinion and the accurate information contained in the text and orientation [33].

The lexicon-based strategies make use of a vocabulary of words whose labels indicating the sentimental valences of those words [34]. These techniques break down a text into a collection of words, whose sentiment orientations are then summed up or combined to categorize the text. Although this method is straightforward, it mainly relies on manually tagging the text [35]. Baharudin et al. [36] proposed that sentence structure and context play an important role in the classification and orientation of sentiment. Each word in the sentence was given a sentiment score from the SentiWordNet lexicon in their work. The cumulative sum of the individual scores for each of the terms in the sentence determines the phrase’s overall classification. Although the method is intriguing, one of its drawbacks is that words with the same orientation but opposing meanings may be classified as having the incorrect lexicon labels in machine learning models.

Word-grams are employed in text classification to produce word co-occurrence patterns and vectors for machine learning classifiers [37]. Jain et al. [38] employed bi-grams and tri-grams in text representation to extract features from the text. Their research produced encouraging findings, demonstrating that N-grams can effectively represent text. They suggested an extensive, cognitive computing-inspired big data analytics framework for sentiment analysis and classification.

Techniques for word embeddings-based vector representation have recently gained importance in natural language processing [39]. Mudinas et al. [40] created a word sentiment label by combining a lexicon-based approach with a support vector machine classifier. Rezaeinia et al. in [41] assigned lexicon vectors to words in a text using a variety of lexicons called Lexicon2Vec, and to create a hybrid vector representation, they coupled their vector with Word2Vec and PoS2Vec.According to Mikolov et al. [42], the field of word-embedding feature selection research gathered steam in 2013. Word2Vec [43], Glove [44], and FastText [45] are the three primary word embedding algorithms that are used to turn words into vectors. The Bidirectional Encoder Representations from the Transformers (BERT) model has garnered a lot of interest because of bidirectional and attention processes [46]. In [47], the authors evaluated Word2Vec, Glove, FastText, and BERT in their study, which showed the significance of the BERT model in sentiment analysis. The capacity of BERT to read words in bidirectional form, in contrast to other word embedding models, would undoubtedly improve the contextual performance of the target text. The BERT’s disadvantage is that it reads the target text in its entirety. As a result, a BERT embedding-based model performs better than other models, resulting in an impressive performance in sentiment analysis tasks [47,48]. The usage of synonyms and the creation of vectors with lower dimensionality than the bag of words make word embeddings superior to the traditional bag of words representation [41,49]. Garg et al. [50] proved that Word2Vec embeddings outperformed alternative word embedding techniques. Currently, researchers are using pre-trained word embedding vectors to perform sentiment analysis because they are more precise and compatible with deep learning neural networks [51]. However, pre-trained word embeddings overlook the sentiment orientation of words and their semantics, which reduces the accuracy of sentiment categorization. Chen [52] examined the convolutional neural network’s performance using pre-trained Word2Vec vectors as inputs and by setting various hyper-parameters for the convolutional neural network model. For aspect-level sentiment analysis, the authors in [53] employed pre-trained Glove vectors as inputs in an attention-based LSTMs model. By adding domain information to the vector, the authors in [54] improved the pre-trained Word2Vec models for cross-domain classification. The authors in [55] categorized Konkani texts using FastText pre-trained word embeddings and neural networks. In [56], the authors compared various baseline models—Glove, Word2Vec, etc.—and proposed a LeBERT model by combining the sentiment lexicon and BERT via word N-grams to classify the sentiments using a convolutional network. According to related research, it is observed that most of the existing research for a product review system has been done using traditional sentiment analysis and opinion mining. It is also further examined in the existing literature that BERT is a superior model for word embedding and plays a significant part in advanced sentiment analysis. By delving further into conventional sentiment analysis and extending the work already being done, we propose the QLeBERT model in this study to predict product quality from consumers’ attitude in product reviews based on the quality subcategory of the attitude system in an appraisal framework. QLeBERT combines a lexicon approach based on the quality subcategory of appraisal framework with a BERT model via word N-grams and then, the BiLSTM model is used to predict the quality of the product.

3. Methodology

This section comprises text preprocessing, description of proposed model, proposed pseudo code for the generation of vectors, and the BiLSTM model.

3.1. Text Proprocessing

Preprocessing of text is concerned with eliminating unimportant features and extracting important features from the text so that a machine learning model can be built efficiently and accurately. Amazon customers’ product reviews datasets were preprocessed by the help of the following steps:

3.1.1. Tokenization

Tokens were created for each sentence in the customer review by using delimiters like a semicolon, colon, comma, dot, etc. Each token is independent of the others and has a separate meaning.

3.1.2. Text Conversion into Lower Case

To ensure that both upper- and lower-case text convey the same meaning, we converted all review text into lower case [57].

3.1.3. Removal of Punctuations

Reduction of punctuation is an important task when working with natural language processing activities because punctuation symbols do not have a lot of semantic meaning and may introduce noise into the data. Therefore, we omitted punctuation symbols from product reviews.

3.1.4. Removal of Stop Words

The stop words were eliminated since they did not significantly add meaning to the data. By eliminating stop words, the emphasis is transferred to the significant words, which decreases the text’s dimension and makes it simpler to identify patterns and semantics in the text.

3.2. Description of Proposed QLeBERT Model

The BERT model is a word embedding model that reads series of words of the input text in a bidirectional manner. It comprises an attention mechanism to allocate words to its vectors based on nearby words [58]. Even though BERT takes a word’s context into account when allocating the vector, it does so for every word in the input text, resulting in a high-dimensional vector. Second, word vectors produced by BERT lack semantic information, which is crucial for text classification. Conventional sentiment analysis uses a sentiment lexicon to identify sentimental terms in a text and then assign sentiment polarity to the words. However, this technique does not generate word vectors, which will lead to data sparsity. This research presents the QLeBERT model as shown in Figure 1. It combines a lexicon based on a quality subcategory of the attitude system in the appraisal framework, N-grams, and the BERT algorithm to improve conventional sentiment categorization. The main goal of QLeBERT is to segment the input text using N-grams methods by using lexicon-based methods to identify segments of the input text related to the quality subcategory words of the attitude system in the appraisal framework, subsequently converted into vectors by the BERT algorithm. A BiLSTM model is then run on the resulting vector to extract contextual features. The resultant features are analyzed by the Softmax function to predict the quality of the product.

In our proposed model, N-grams are used to incorporate the two most popular methods for creating text vectors for sentiment analysis: word embedding-based methods and a lexicon-based approach. N-grams are used to predict the subsequent word in a string of words, and they also generate a co-occurrence of words, which is an important element to identify sentiment in a text. Lexicons exhibiting the quality of a product are discovered by word N-grams comprising lexicon words based on the quality subcategory in an attitude system of the appraisal framework, then the BERT model is used to build vectors from N-grams words. In this research work, we employ N-grams to divide the customers’ review into different segments that represent the whole customers’ review. This is due to the fact that N-grams provide a text’s word co-occurrence in a more thorough manner than a bag of words.

3.3. Algorithm: QLeBERT Model

The pseudocodes of the proposed QLeBERT model are presented below as Algorithm 1:

Algorithm 1: Psecudocodes of the QLeBERT model

//This algorithm takes customer reviews CRi having words Wi and generates high //quality and low products h and l, respectively. Input: Customer Reviews CRi having words Wi Output: Customers’ Attitude toward Product as h or l Set N-gram value to N = 2 for all CRi ∈ Repi containing n tokens //Repository of customers’ reviews display Bigram Word apply QLeBERT model end for for all bigram_words if bigram_words exists in Quality Appraisal Category display bigram_words else discard bigram_words end if end for for each Wi read Wi from B_Wi   // B_Wi represents bag of words apply BERT model generate and update Wvec end for apply BiLSTM (Wvec)   // to classify the product as h or l exit

3.4. BiLSTM

A forward and a backward LSTM make up the BiLSTM model, and the LSTM in this model is a variant of a recurrent neural network (RNN). A gating unit in LSTM is used to solve the vanishing gradients problem. This gives the LSTM a stronger ability to counter issues like long-term dependencies and empowers the RNN to exploit the dependencies that exist in long-distance data in a better and more effective way. The authors in [59] proposed BiLSTM to obtain backward and forward features in a specific amount of time. Using the input sequence data, the BiLSTM model extracts contextual information. As shown in Figure 2, [60], $\overrightarrow{\overrightarrow{H\mathbf{0},}H\mathbf{1}},\overrightarrow{H\mathbf{2},}\overrightarrow{H\mathbf{3},}\dots \dots \dots \dots \dots \dots \dots \dots .\overrightarrow{Hn}$ show the forward hidden layer state of BiLSTM whereas $\overleftarrow{H\mathbf{0}}, \overleftarrow{H\mathbf{1},}\overleftarrow{H\mathbf{2},}\overleftarrow{H\mathbf{3},}\dots \dots \dots \dots \dots \dots \dots ..\overleftarrow{Hn}$ indicate the backward/reverse hidden layer state of BiLSTM. Equations (1) and (2) display the state equations for each moment of the model while Equation (3) jointly represents the final state of the BiLSTM model.

$$\overrightarrow{Hi}=LSTM(x_{\mathit{i}}, \overrightarrow{Hi}-\mathbf{1})$$

(1)

$$\overleftarrow{Hi}=LSTM(x_{\mathit{i}}, \overleftarrow{Hi}-\mathbf{1})$$

(2)

x_i shows input vector at time I, $\overrightarrow{Hi}-\mathbf{1}$ denotes the forward hidden layer vector at time i − 1, and $\overleftarrow{Hi}-\mathbf{1}$ denotes the reverse hidden layer vector at time i − 1, respectively. Output vector vi as shown in Equation (3) combines both forward and backward hidden layers vectors.

$$\mathbf{V}\mathbf{i}=\overrightarrow{Hi}+\overleftarrow{Hi}$$

(3)

3.5. Focal Loss Function

In machine learning, class imbalance occurs when one class has a disproportionately smaller number of samples compared to the other class [61,62,63]. Only 11% of the customer reviews in our problem are classified as low quality. Under-fitting and over-fitting problems are raised when we under-sample the majority class or over-sample the minority class, respectively, while resolving the issues of class imbalance. Over-sampling the minority class or under-sampling the majority class are the methods for determining the issue of class imbalance [64]. The disadvantage of these approaches is that they could result in either under-fitting (for the minority class) or over-fitting (for the majority class). In imbalance classes, training is ineffective because the majority of samples are simple instances with no relevant learning signal; the simple examples might overwhelm training and build defective models. The employment of a loss function is an additional tactic. Loss functions are mathematical formulas used to measure the degree to which predictions differ from the actual values, while smaller loss values suggest that the predictions are quite accurate, higher loss values imply that the model is making a substantial error. The Huber Loss function, Log-Cosh Loss function, Quantile Loss function, Cross-Entropy Loss function are the types of Loss function. A solution to the issue of class imbalance can be found in the Loss function. In particular, the Focal Loss function has been employed in object detection and image analysis [65]. Focal Loss function uses a method known as “Down Weighting” to address the issue of class imbalance. Down weighting is a method to add modulating factors to the Cross-entropy Loss function by giving emphasis on hard examples and by lessening the impact of easy examples.

$$FL\left(\mathit{p}\mathit{t}\right)=-\mathsf{\alpha }{\left(\mathbf{1}-\mathit{p}\mathit{t}\right)}^{\mathsf{\gamma }}log\left(\mathit{p}\mathit{t}\right)$$

(4)

$$p\left(t\right)=\{\begin{array}{c}p, if y=1\\ 1-p, Otherwise\end{array}$$

Let p be the logit probability for the classification of the quality of a product as high or low (y = 1 or 0). Alpha (α) and Gamma (γ) are hyper parameters to enhance binary cross entropy—$\mathbf{log}\left(\mathit{p}\mathit{t}\right).$ Factor ${\left(\mathbf{1}-\mathit{p}\mathit{t}\right)}^{\mathsf{\gamma }}$ is added by Focal Loss function to the standard cross entropy, as shown in Equation(4). Due to this method, the model is saved from overfitting and can differentiate between classes efficiently. The setting $\mathsf{\gamma }\ge \mathbf{0}$ reduces a relative loss for well-classified samples (pt > 0.5), putting more focus on hard samples which are difficult to classify.

4. Experiment

This section includes a description of datasets, evaluation metrics, experiment findings, and discussions. There is also a discussion of the methods and techniques used in the creation and assessment of models.

4.1. Datasets

In this study, the proposed model QLeBERT is implemented in the Jupyter Notebook using Datafiniti_Amazon_Consumer_Reviews_of_Amazon_Products taken from [66] comprising 28,332 customers’ reviews and nine benchmark Amazon product review datasets [67]: Linksys, Hitachi Routers, Creative Zen MicroMP3 Player, Canon Power Shot SD500, Canon S100, Diaper Champ, IPod, Nokia 6600, and Norton Antivirus, containing 2656 customer reviews, which were originally created for sentiment analysis.

4.2. Lexicon-Based Customers’ Attitude Analysis to Predict Product’ Quality Using Appraisal Framework

According to the language perspective, the appraisal framework presents the emotional state of the author toward a given event and the level of participation through linguistic terms [68].

Each word in the customer review is labeled as high quality and low quality based on the quality subcategory of the appraisal framework. We derived an overall quality of the product score by counting the numbers of high quality product-related words and low quality of product-related words, combining these values mathematically by using Equation (5):

$$\mathrm{QScore}=\frac{\mathrm{High} \mathrm{quality} \mathrm{of} \mathrm{the} \mathrm{product} \mathrm{related} \mathrm{words}-\mathrm{Low} \mathrm{quality} \mathrm{of} \mathrm{the} \mathrm{product} \mathrm{related} \mathrm{words}}{\begin{array}{c}\mathrm{Total} \mathrm{number} \mathrm{of} \mathrm{words}\end{array}}$$

(5)

The customers’ attitude toward product quality is that it is one of the low-quality products if the quality score is less than 0. Therefore, high quality is indicated by a score greater than 0. In the customer review, for example, “Software that came with MicroMP3 is awful”, the terms “awful” falls in the quality subcategory of the attitude system in the appraisal framework. The reviewers’ attitude is shown as negative toward the quality of the product, and the valence dictionary would label the same as low quality. Similarly, in a review sentence like “the picture quality of the camera is good in automatic mode”, the term “good” also belongs to the quality subcategory of the attitude system in an appraisal framework. It represents a positive attitude toward the picture quality of the camera and the valence dictionary would label it as high quality.

Product features—reliability, durability, function, aesthetics, and perception based on the quality subcategory of the attitude system in the appraisal framework—are extracted and the lexicon-based labeling is assigned to each feature, as given in

Table 1. Details about the features of products as high quality and low quality.

Dataset Name/Features→	Reliability		Durability		Function		Aesthetics		Perceived		Total High Quality Reviews	Total Low Quality Reviews
Quality-->	High	Low	High	Low	High	Low	High	Low	High	Low
Amazon Nine Products Customer Reviews	35	28	32	28	49	25	675	300	21	3	812	384
Datafiniti Amazon Consumer Reviews of Amazon Products	1152	463	3082	701	2949	500	2502	652	6498	755	16,183	3071
Total feature based quality related reviews	1187	491	3114	729	2998	525	3177	952	6519	758	16,995	3455

. Figure 3 depicts customers’ attitudes toward product features and are classified as high or low quality using an appraisal framework.

4.3. Experiment Setup

The reviews included in the dataset were preprocessed by doing tokenization, conversion of all the text of reviews into lower case for extraction of the same meaning for both upper and lower cases, and omission of punctuations and removal of stop words in the reviews. After preprocessing, the creation of a text vector through the BERT word embedding model and the design BiLSTM layers were done in Python language by using the Jupyter Notebook of Anaconda. The text vectors were divided into training and testing sets with a ratio of 80:20, respectively. The BiLSTM model was trained using 80% of the dataset, while 20% of the dataset was utilized to evaluate the model’s performance. Multiple sentences were included in each review; hence, the BERT embedding method used a pooled output.

The resultant vectors were then transmitted to the fully connected layer through the Relu function. In the end, the Softmax function was used to classify the product quality into low and high based on the attitude system of the appraisal framework.

4.4. Evaluation Metrics

The researcher classified sentences based on the appraisal theory to evaluate the quality of products. The combined dataset of Amazon product reviews is highly imbalanced, and high accuracy is predicted based on the majority class. [69] used different metrics to assess the performance regarding model prediction, involving F1_micro, F1_macro, F1_weighted, accuracy, and confusion matrix. These metrics are regularly used to handle problems related to unbalanced data.

The F1_micro refers to computing the global average F1 score. The researcher computed the overall True Positive, False Positive, and False Negative and then fed it into the F1 equation to achieve a micro F1 score.

$$\mathrm{F}1 \mathrm{micro}=\frac{True Positive}{True Positive+0.5\left(False Positive+False Negative\right)}$$

(6)

The F1_weighted score is computed by taking into account each class’s support with a mean of F1 scores of each class. The formula to calculate the F1_weighted score is shown in Equation (7).

F1 weighted = F1_label1 × W₁ + F1_label2 × W₂ + F1_label3 × W_{3……………………….}F1_labeln × W_n

(7)

The main theme of the precision is to minimize false positives, and the main focus of the recall is to minimize false negatives. The main goal of the imbalanced learning is to increase the recall score without hurting precision. F1-score combines the properties of recall and precision. F1_macro represents the arithmetic mean of per-class F1-scores. F1-macro is a commonly used metric for learning from imbalanced data and evaluating the performance of the classification model. The formula to compute the F1-macro score is shown as Equation (8).

F1-macro = Arithmetic Mean (F1-scores of n classes)

(8)

5. Results and Discussion

This section comprises the results of experiments. We test the proposed model on benchmark Amazon customers’ reviews datasets. In this research study, we employed the 100-dimensions Glove word embedding model which is pre-trained on the English Wikipedia Giga-word 5th Edition dataset and the 100-dimensions word2vec word embedding model which is pre-trained on Google news. We reduced 768 dimensions of the BERT model to 128 dimensions to reduce the storage requirement. In training, we employed tensor flow to implement and evaluate the model. Smaller numbers of reviews were used to validate our model. We first check the effect of the lexicon-based approach on the input data and vector as shown in

Table 2. Description of Dataset Before and After using the Lexicon-based approach.

Detail of Data Item	Without Implementing Lexicon-Based Approach	After Implementing Lexicon-Based Approach
Number of words	64,378	25,667

.

Table 2 shows that the lexicon-based method was applied to take out a portion of the input text, considerably reducing the size of the text as a whole and thereby decreasing the computational time for the model. We then carried out an experiment with BiLSTM to assess the performance of the QLeBERT model to predict the quality of a product.

The embedding shape and preprocessing for the BERT model are represented by the Keras layer. Due to the lack of computational resources, BERT small was employed to initialize the word embedding. The word embedding’s dimension was set to 128. The default settings for the baseline models, Glove and Word2Vec word embeddings with 100 dimensions each, were used. The Keras Layer is an input layer that comprises input vectors generated by word embedding models.

We first conducted an experiment to investigate the impact of the size of N-grams on the QLeBERT model using the BiLSTM model. The datasets about customers’ reviews of Amazon products were employed in the experiment.

Table 3. Performance measurement of QLeBERT- BiLSTM by using N-grams.

N-Grams	QLeBERT-BiLSTM
	F1-Micro	F1-Macro	F1-Weighted	Accuracy
N = 1	0.76	0.72	0.71	0.76
N = 2	0.95	0.91	0.95	0.95
N = 3	0.92	0.87	0.86	0.92
N = 4	0.89	0.86	0.88	0.89
Entire Words	0.83	0.80	0.82	0.83

displays the experimental outcomes for N = 1, 2, and all words.

For N = 1, it shows that the lexicon-based approach selected only one word to predict the quality of a product based on the quality subcategory of the attitude system in the appraisal framework. Only one word cannot represent the entire customer’ review. Therefore, a poor result was achieved. N = 2 demonstrated the best outcome by arriving the F1-macro score of 0.91. We checked the performance of the model up to N = 4. In the case of the entire text of the review, the lexicon-based approach was not applied, hence this return back to the BERT model.

Table 4. Performance measurement of Glove-BiLSTM by using N-grams.

N-Grams	F1 Micro	F1 Macro	F1 Weighted	Accuracy
N = 1	0.76	0.62	0.71	0.76
N = 2	0.76	0.71	0.76	0.76
N = 3	0.89	0.78	0.88	0.89
N = 4	0.81	0.77	0.80	0.81
Entire Words	0.72	0.70	0.72	0.72

demonstrates the performance results of the Glove word embedding approach with the BiLSTM model using N-grams. For N = 3, the Glove word embedding technique with the BiLSTM model achieved the best result, having an F1 macro score of 0.78 among different N-gram words with Glove word embedding and the BiLSTM model.

Table 5. Performance measurement of Word2Vec-BiLSTM by using N-grams.

N-Grams	F1 Micro	F1 Macro	F1 Weighted	Accuracy
N = 1	0.77	0.67	0.78	0.77
N = 2	0.83	0.80	0.82	0.83
N = 3	0.80	0.77	0.80	0.80
N = 4	0.81	0.76	0.79	0.81
Entire Words	0.75	0.69	0.70	0.75

depicts the performance measurement of Word2Vec model with the BiLSTM model using N-grams. For N = 2, the Word2Vec word embedding technique with the BiLSTM model achieved a better result, with an F1-macro score of 0.80 among different N-grams with the Word2Vec word embedding approach and the BiLSTM model.

5.1. Performance Measurement of QLeBERT Model in Comparison to Different Models

An experiment was conducted on the Amazon Customers’ Reviews Dataset to validate the performance of QLeBERT with Bigram (N = 2) compared to other pre-trained word embedding models: Glove, Word2Vec, and BERT with theBiLSTM model. The experiment was carried out with and without a lexicon-based approach, as shown in

Table 6. Comparative Analysis of QLeBERT with different models.

Model	F1 Micro	F1 Macro	F1 Weighted	Accuracy
Word2Vec	0.85	0.70	0.80	0.85
Glove	0.88	0.72	0.82	0.88
Lexicon-Glove	0.89	0.78	0.88	0.89
Lexicon-Word2Vec	0.83	0.80	0.82	0.83
BERT	0.83	0.81	0.78	0.83
QLeBERT(our model)	0.95	0.91	0.95	0.95

.

The proposed QLeBERT outclassed different models with and without a lexicon-based approach to predicting the quality of a product from the quality subcategory of the attitude system in the appraisal framework. F1-macro is an excellent performance assessment metric where classes are imbalanced.

5.2. Comparison of Proposed QLeBERT with Baseline Model

Finally, we compared our proposed QLeBERT model with the baseline model as reported by the authors in [56]. They used a sentiment lexicon and BERT via word N-grams to classify the sentiments using a convolutional network. In this research study, we used the lexicon approach based on the quality subcategory of the appraisal framework and BERT via word N-grams to classify the quality of the product using BiLSTM. The proposed QLeBERT, when tested on the Amazon product reviews dataset, achieved the highestF1-macro score of 0.91 as compared to the baseline approach to predicting the quality of a product from customers’ attitude using an appraisal framework.

6. Conclusions and Future Work

In this study, advance machine learning techniques are used to determine product quality from the customers’ attitude based on the quality subcategory of the attitude system in an appraisal framework. We proposed the QLeBERT model, comprising a lexicon-based approach, N-grams, BERT word embedding model, and BiLSTM. In the proposed model, a segment of a sentence where the quality of the product-related lexicon based on an appraisal framework using N-grams is selected and then the BERT word-embedding model was employed for vectorization. In the end, the BiLSTM classifier is used to classify the quality of the product as low or high. Glove, Word2vec, and BERT word-embedding models were considered as baseline models on a benchmark dataset of Amazon product reviews. As shown in the experiment results, QLeBERT had a high predictive F1-macro score of 0.91 as compared to baseline models. In future work, our QLeBERT model could be used with resource-poor languages, and evaluation of the model with other languages except for the English language will be a promising study.