Product review opinion based on sentiment analysis

Abstract

Sentiment analysis is a natural language processing (NLP) technique for determining emotional tone in a body of text. Using product reviews in sentiment analysis and opinion mining various methods have been developed previously. Although, existing product review analyzing techniques could not accurately detect the product aspect and non-aspect. Hence a novel Detach Frequency Assort is proposed to detect the product aspect term using TF-ISF (Term frequency-inverse sentence frequency) with Part of Speech (POS) tags for sentence segmentation and additionally using Feedback Neural Network to combine product aspect feedback loop. Furthermore, decision-making problem occurs during classification of sentiments. Hence, to solve this problem a novel technique named, Systemize Polarity Shift is proposed in which flow search based Support Vector Machine (SVM) with Bag of Words model classifies pre-trained review comments as positive, negative, and neutral sentiments. Moreover, the identification of specific products is not focused in sentiment analysis. Hence, a novel Revival Extraction is proposed in which a specific product is extracted based on thematic analysis method to obtain accurate data. Thus, the proposed Product Review Opinion framework gives effective optimized results in sentiment analysis with high accuracy, specificity, recall, sensitivity, F1-Score, and precision.

Keywords

Sentiment analysis opinion mining support vector machine thematic analysis

1 Introduction

Sentiment analysis is the practice of detecting the emotion of a statement using data, views, reviews, or utterances using natural language processing (NLP) or machine learning techniques [1]. A substantial proportion of brief text data containing user opinions in social media messages and user assessments of products and services has developed only as consequence of the greater use of internet apps. As a result, user sentiment analysis and user opinion mining have become attractive research topics. Consumer reviews on social media are critical for a range of economic purposes. The volume and relevance of user reviews posted on social media are increasing at an exponential rate, culminating in big data [2]. Regularly, a massive amount of data is created when people share their views or ideas while chatting with one another on social networking platforms like Twitter, Facebook, and Myspace [3].

Moreover, Social data is one of the big 5 data categories that is created through a variety of social channels and exhibits all three big data characteristics: velocity, heterogeneity, and massive volume. It also has a semantic component, which refers to the fact that it was created by humans and contains symbolic information with intrinsic subjective value. Sentiment analysis faces several challenges and opportunities as a result of this one-of-a-kind element of social big data. Opinion mining and sentiment analysis in social media occur with a slew of difficulties [4]. End-user authentication is one of the most challenging challenges to solve since there is a potential that noise will be introduced into the data [5]. Another major stumbling point is the inconsistency of social media data. The way people express their emotions and use language varies from one person to the next [6]. When shorthand notations are utilized, the classifier has a difficult time distinguishing between word attributes. People’s online thoughts and feelings have an impact on other people’s opinions on movies, regional conflicts, business, movies, goods, health care, and so on [7].

Generally, Customers make purchase decisions, watch movies, vote for political parties or candidates, and so on based on the ideas and feelings expressed by others on social media [8]. Sentiment analysis is a technique that employs natural language processing to categorize people’s emotions, opinions, and reviews expressed in text, emojis, and other forms as negative, positive, or neutral [9]. Sentiment analysis benefits businesses, legislators, human resources, educational institutions, governments, and other organizations. To gain a deeper understanding of people’s emotions, which could assist them to improve their products, sales, and services, as well as better understand their customers [10]. During natural disasters or tragedies, sentiment analysis can help governments and public service agents assist those who have been affected by the calamities, potentially saving many lives. Opinion mining is one of the hottest and fastest-growing research topics, making it difficult to keep up with everything that’s going on [11]. It tries to explore people’s feelings, opinions, and attitudes toward issues, events, circumstances, entities, people, and their traits as represented in text evaluations or comments via social media (e.g., social networking sites, forums, blogs, etc.) [12].

Where Amazon is an example of a world-renowned online retailer that allows customers to freely rate and review its products. Customers will be able to make decisions ranging from purchasing a product such as a camera, mobile phone, etc., to writing movie reviews and making investments, all of which will have a significant impact on daily life if these reviews are analyzed to determine whether they are positive or negative. By automatically evaluating consumer feedback, such as comments in survey responses and social media dialogues, brands can learn what makes customers pleased or dissatisfied. This allows them to customize products and services to meet the needs of their customers [13]. Understanding user opinion is important for a variety of commercial applications, including product evaluations, political campaigns, product feedback, marketing analysis, and public relations, among others [14]. It might also be used in more serious scenarios, such as security threats like monitoring for terrorism-related communications. Unstructured data, such as social media text, has become a key challenge for collecting information in the fourth industrial revolution (4IR), as it is freely available through individual user-created content on media platforms like Facebook or microblogs [15]. However, interpreting the text is a challenging problem for which neither researchers nor industry have provided a definitive answer [16].

According to customer reviews, existing techniques for detecting product features and non-aspects utilizing the frequent noun method were ineffective. Moreover, the objective of the research is to obtain an effective analysis of customer reviews in products using the sentiment analysis approach.

The main contribution of this paper is as follows:

Detach frequency assort is presented to detect aspect term of product from review comments and then sentences are formed using TS-ISF with POS tags.

Systemize Polarity Shift is introduced to detect sentiment polarity and classify negative, positive, and neutral sentiments using flow search SVM.

Product dataset is pre-trained with grouped reviews and the specific product is extracted based on thematic analysis in Revival Extraction.

Hence, product review opinions based on sentiment analysis effectively predict aspect terms, classify them based on polarity and extract specific products with increased accuracy and efficiency. The remainder of the paper has been organized as follows: Section 2 presents the recent literature; section 3 depicts the detailed description of the proposed methodology; section 4 discusses the implementation results; finally, section 5 concludes the paper.

2 Literature survey

Aziz et al., [17] Sentiment analysis is a technique for computationally identifying and categorizing opinions expressed in a piece of text, especially to determine whether the writer’s attitude toward a certain subject is negative, positive, or neutral. Many researchers have proposed novel sentiment categorization methods, especially those based on supervised machine learning (ML). However, there is a lack of research in Cross-Domain Sentiment Analysis that has shown positive results at the moment. The purpose of this study was to examine how well the ML approaches described performed in the Single Domain while replacing feature selection procedures and running tests. MNB, SVM-LK, and SGD were the most reliable ML techniques. Although it takes significantly less time to complete the analysis, TF-IDF is preferable to frequency terms for feature selection (reduce more than 60 percent) SVM-capabilities RBF’s produced strong results, particularly for less well-known domain knowledge.

Basiri et al., [18] Sentiment analysis is becoming the main research issue in the field of natural language processing and data mining over the last decade. Deep neural network (DNN) models have recently been utilized to do sentiment analysis tasks, and the results have been encouraging. We present an Attention-based Bidirectional CNN-RNN Deep Model to extract both past and future contexts (ABCDM). ABCDM outperformed the competition on both long and short tweet polarity categorizations when compared to six previously suggested DNNs for sentiment analysis. The experiments used five reviews and three Twitter datasets. Using two bidirectional LSTM and GRU networks, past and future contexts are extracted as semantic representations of the input text. This method works well for extracting past context, however, it does not work well for predicting future context with semantic analysis because of the GRU technique’s gradient issue.

Tian et al., [19] Aspect-based sentiment analysis (ABSA) investigations currently use neural graph-based models. We propose a method for using ABSA dependency types with type-aware graph convolutional networks (T-GCN) With state-of-the-art performance on six English benchmark datasets, the experimental findings support the validity and efficacy of our methodology. Show how T-GCN’s several layers help ABSA with quantitative and qualitative analysis. We introduce a T-GCN-based neural approach for ABSA in which the input graph is based on the dependency tree of the input sentence.

Tian et al., [19] The neural graph-based models are used in the current aspect-based sentiment analysis (ABSA) research. We propose a method for leveraging ABSA dependency types by employing type-aware graph convolutional networks (T-GCN) With state-of-the-art performance on six English benchmark datasets, the experimental findings support the validity and efficacy of our methodology. Show how T-GCN’s several layers help ABSA with quantitative and qualitative assessments. The input graph is built on the dependency tree of the input sentence, and we provide a neural approach for ABSA utilizing T-GCN. The majority of the work will involve empirical data analysis and inquiry, taking into account the different indicators related to user assessments. The relationship between the number of Helpful reviews and the price was demonstrated using a trend line. Both variables exhibit a negative relationship, as seen by the trend line. The expected values are different from the observed values, and the p-value is also small. As a result, our null hypothesis is untested.

Raviya et al., [21] Sentiment analysis on social media platforms such as Twitter and Facebook is now the most comprehensive method of acquiring insight into client opinions, according to the context of NLP, automated text classification can be a critical activity. There are two primary techniques for sentiment analysis: lexicon-based and machine learning-based methods. This paper’s combination of deep CNN and SVM will automatically discover and extract subjective sentence-level attributes to perform sentiment analysis of an online product review dataset with the highest accuracy and quickest computation time. The study of belief could provide useful knowledge. We employed the embedding layer Word2vec to get higher-quality subjective features of the sentences. Adding a smaller batch size to the CNN model reduces total losses. A dropout approach was implemented in the completely linked hidden layers, which resulted in an extra speed gain. While changing to smaller batches minimizes loss, it also increases sentiment polarization inaccuracy.

Cross-verification in large datasets was erroneous, resulting in ineffective opinion analysis [17]. Due to the illegal logical analysis aspect words from sentences could not be disclosed, extraction [18] of information from social media with an irregular framework, aspect of sentiment requires to give out exact reviews about the product [19]. With sentiment polarity identification, product reviews [20] from a single dataset could not support the null hypothesis [21]. As a result, a novel approach for analyzing product aspect terms, detecting sentiment polarity, and extracting the product from client product evaluations must be proposed. In the following section, the novel solution to the problem is detailed explained.

3 Product review opinion based on sentiment analysis

Sentiment analysis is one of the most rapidly growing research topics in computer science which identifies sentiments behind a body of text. A vast number of techniques have grown to analyze customer reviews. However, according to user reviews existing techniques are not able to distinguish between product aspects and non-aspects and results in low prediction accuracy. Hence, Detach Frequency Assort is proposed to detect the product aspect term from the review sentence in which TF-ISF term frequency-inverse sentence frequency is used with POS tags for sentence segmentation. Term frequency holds the counts of each word and inverse sentence frequency holds counts of words that occur in all sentences. Segmented terms are grouped with product aspects in which Feedback Neural Network is used to connect the product aspect feedback loop. Grouped reviews need to be classified so that opinion of the customer about the product could be identified. But existing techniques use opinion mining to classify the reviews based on the content but could not provide effective result since decision-making problem occurs while classifying them as positive, negative, and neutral. Hence, Systemize Polarity Shift is introduced in which sentiment polarity detection is processed by flow search in SVM with Bag of Words model and classified with pre-trained review as extra-ordinary, confident, adverse, and indigent. According to customer sentiment, the reviews are grouped into 4 terms, extraction of each product based on their opinion level could make the manufacturing level more efficient. The existing technique does not focus on sentiment analysis in specified products. Moreover, an effective analyzing technique or pattern is required to extract the required terms from different classifications. Hence, Revival Extraction is proposed in which products dataset is pre-trained with grouped reviews and extraction of a specific product is done based on thematic analysis method which is processed through regulation of same word, pattern or frequency to obtain accurate data. So that the specific product which has more usage and its efficiency could make manufacturing field yield higher. Hence our proposed Product Review Opinion framework gives effective optimized results in sentiment analysis for customer product reviews.

Figure 1 depicts the block diagram of the proposed product review opinion framework. Customer review from amazon dataset is given to product review opinion approach in which aspect term from review comments are determined and sentence is formed using TF-ISF with POS tags. Then, classify predicted sentences based on polarity (positive, negative, and neutral) using flow search SVM. Finally, Thematic analysis is performed to extract products.

Fig. 1

Block diagram of product review opinion.

3.1 Detach frequency assort

Sentiment analysis in analyzing customer reviews is the most rapidly growing field of research, and sustaining all the activity is difficult. Existing customer review analyzing techniques failed to accurately detect product aspects and non-aspects. Hence, a novel Detach Frequency Assort is proposed to detect the product aspect term from review comments based on TF-ISF (Term frequency-inverse sentence frequency) with POS tags. TF-ISF provides sentence segmentation and thereby forms sentences from review comments. Furthermore, term frequency keeps track of the number of times each word appears, whereas inverse sentence frequency keeps track of the number of times each word appears in all sentences. Also, in response to a query, sentence retrieval extracts relevant sentences from a comment base. Term Frequency-Inverse Sentence Frequency method (TF-ISF) is used in the Detach Frequency which performs sentence retrieval based on ranking function. The Equation (1) determines the sentence retrieval process in TF-ISF which is given by. $\begin{matrix} R (s / q) = \sum_{t \in q} log (tft, q + 1) \\ log (tft, s + 1) log (\frac{n + 1}{0.5 + sft}) \end{matrix}$ (1)

tft, q is the number of times the term t appears in the question q

tft, s is the number of times the term t appears in sentence s

sft is the number of sentences containing the phrase t

n is the collection’s total number of sentences

Equation (1) uses the previous and next sentences as the local context of each sentence in the same comment. Assume that a sentence’s significance is determined in part by its content and in part by the content of the two sentences immediately adjacent to it. The content of nearby sentences influences the importance of neighboring sentences. Based on these two assumptions, the recursive ranking function in TF-ISF for sentence retrieval is given in Equation (2): $\begin{matrix} R_{con} (s / q) = (1 - μ) . R (s / q) + μ . \\ [Rcon (Sprev (s) / q) + Rcon (Snext (s) / q)] \end{matrix}$ (2)

Where Sprev is the previous sentence of sentence S, Snext is the next sentence of S. Rcon (Sprev (s)/q) and Rcon (Snext (s)/q) is Relevance of the previous and next sentence. TF-ISF score is calculated by analyzing distribution of each word throughout the comments. It is used to determine the importance of a word in a comment based on its frequency inside a sentence and distribution over the sentence in a review comment’s sentences. The TF-ISF measure of a word w in a phrase s abbreviated as TF - ISF (w, s) , is calculated as follows: $TF - ISF (w, s) = TF (w, s) * ISF (w)$ (3)

In Equation (3), frequency is a term. The number of times the word w appears in sentence s is TF (w, s), and the inverse sentence frequency is ISF (w) . The formula for ISF (w) is as follows: $ISF (w) = \log (| S | / SF (w))$ (4)

In Equation (4), sentence frequency SF (w) is the number of sentences in which the word w occurs. In this phase, the top 20% of words with the highest TF-ISF scores are contenders for keywords.

Algorithm for Keywords Extraction in TF-ISF

Step 1: Take amazon review comments as input

Step 2: Remove stop words from input text

Step 3: Check the input words in noun morph and if necessary perform noun stemming.

Step 4: Calculate TF-ISF score of each remaining word using TF-ISF measure of a word w in a phrase

Step 5: Take high TF-ISF scores (top 20%) as candidates for keywords.

Step 6: Find the intersection of noun keywords with TF-ISF high-scoring keywords.

Step 7: Treat the noun words appearing in the title/headlines as keywords after removing stop words.

Step 8: The Union of keywords coming from step 5 and step 6 is final keywords.

The flowchart of Detach Frequency Assort is shown in Fig. 2 in which amazon review comments data is collected. Then, in preprocessing and data analysis step, the sentence is retrieved and segmented with the removal of stop words using TF-ISF transformation. Detach Frequency Assort method involved in performing POS tagging of unknown words and POS tagging of known words in word segmentation utilize information from words parts of speech tag. A part-of-speech tag (POS tag) is a unique label that is provided to each token (word) in a text corpus to denote the part of speech as well as other grammatical categories like tense, number (plural/singular), case, and so on. A tagset is a grouping of all the POS tags in a corpus, these tagset are quite distinct for unrelated languages and extremely similar for related languages. Tag sets can also be used to drill down into more detailed data. In POS tag sets, tags representing the most common components of speech may be constrained (N for noun, V for verb, A for adjective). In the data collection step, the raw data is collected from Datafiniti’s website, which has a total of 34,000 entries. Among these more than 25 columns are product id, category, sub-category, review title, review text, reviewer name, and so on. There were numerous null values in the input dataset, as well as improper data and pricing formats, outliers, and noise. Figure 3 shows a few columns from the original Amazon customer product review dataset. A few columns have null values and original product review dataset is displayed in Fig. 3.

Fig. 2

Flowchart of Detach Frequency Assort.

Fig. 3

Original Product review dataset.

Since the raw data included more than 20 columns, Preprocessing was used to reduce it to 10 columns. Also, utilized Excel’s built-in formatting tools to appropriately style a few data and pricing columns. Figure 4 depicts the preprocessed data that has been utilized to further analyze the data.

Fig. 4

Preprocessed Product review dataset.

The dataset of Amazon reviews was used in which preprocessing is performed by removing redundancy and then words are segmented. The user’s reviews and ratings of various products, as well as reviews about the user’s experience with the products, were taken into account in this dataset. Since text is the most unstructured form of all accessible data, it contains many sorts of noise and is difficult to analyze without pre-processing. Text preprocessing is done which is the full process of cleaning and standardizing text to make it noise-free and suitable for analysis. Feedback Neural Network is used to connect the product aspect feedback loop by grouping segmented phrases with product attributes. The architecture of the feedback neural network is shown in Fig. 5.

Fig. 5

Feedback Neural Network.

The sentence by-product aspect word is combined with the feedback neural network. The preprocessed data is taken as input to combine sentences with attributes in which aspect terms are connected in a feedback loop. The data can be returned into the input before being sent forward for additional processing and then finalizing output. This section detects the product aspect term from review comments and form grouped reviews. However, classification of reviews based on the content causes decision-making problems. Hence, the next section focus on the effective classification of customer sentiment reviews.

3.2 Systemize polarity shift

Systemize Polarity Shift is presented to provide effective results during the classification of sentiments based on polarity. The reviews are divided into four categories namely extra-ordinary, confident, adverse, and indigent depending on consumer sentiment; extracting each product based on their level of opinion might make the production process more efficient. Furthermore, this process shows the way to train a polarity detection model using preprocessing, feature extraction, filtering, and a sentiment polarity classifier. Flow search in SVM with a bag of word model has been used to classify the pre-trained reviews as positive, negative, and neutral sentiments.

Figure 6 depicts a block diagram of systemized polarity shift. Features are extracted from input grouped reviews. Feature extraction picks a relevant list of words as text features and removing a large number of words that do not add to the text’s emotion. Feature extraction helps to eliminate noise from text and get a more accurate emotion for a comment. Bag-of-words model extract features from review comments with the formation of feature vectors and assigns a score to each word in the grouped review. Then, SVM classification identifies an opinionated text and classifies it into a positive, negative, or neutral category. The neutral label is given to more objective material in which there is no explicit perspective in the text or a mix of positive and negative views. The review comment polarity classifier is developed using N-grams features (N = 1to2) and the SVM learning framework. As a repeated holdout approach, the dataset is partitioned into 10 parts for cross-validation. This method selects 90% of the dataset for the training set and 10% for the testing set, then repeats the process on ten different dataset areas. Finally, the result is averaged over all of the separated and rotating components. The classified positive sentiment is shown in Fig. 7.

Fig. 6

Block diagram of Systemize Polarity Shift.

Fig. 7

Positive sentiment.

Flow search in SVM effectively classifies positive sentiments. Hence, positive sentiments are extracted from the review comments with 99.2% confidence. Figure 8 portrays negative sentiment on a product that is evaluated correctly and confidently.

Fig. 8

Negative Sentiment.

The confidence level for predicting negative sentiment is boosted by training the proposed model with more negative words. Neutral sentiment conveys a sense of neutrality with accuracy and certainty. Training the model with whole amazon review comments containing more positive, negative, and neutral sentiments will improve the accuracy and confidence. Hence, systemize polarity shift effectively classifies sentiments into positive, negative, and neutral without decision-making difficulty. However, there is a need to focus on sentiment analysis in specified products. Hence, the next section focus on the extraction of a specific product to provide optimized sentiment analysis.

3.3 Revival extraction

Revival extraction has been utilized to extract specific products and to yield higher efficiency in sentiment analysis. The products dataset is pre-trained with grouped reviews, and thematic analysis is used to extract individual products. Thematic analysis regulates the use of the same word, pattern, or frequency.

Figure 9 depicts a flowchart of the revival extraction process in sentiment analysis. The data preparation in revival extraction is divided as follows due to the characteristics of grouped review comments on Amazon.

Fig. 9

Flowchart of Revival Extraction.

Step 1: From October 2015 to April 2019, customer posts on the review comments on Amazon were collected, which covered four different product types Health and Beauty, Electronics, Toys, Games. The messages for each brand product were divided into six periods, each lasting two months, and the data set comprised 300 vectors.

Step 2: Each vector is composed of several posted messages, which can be represented as period i = P1, P2, . . . , Pk, period A = P1, P2, . . . , Pk where Pj is the jth posted message and k is the number of posts.

Step 3: Period_i, message A’s posts were manually classified into four groups: Coproducer (CCO), Influencer (INF), Mutual Information (MUI), and Customer Value (CUV)

Step 4: The message post (Pj) was passed to two subtasks after step three: word segmentation and keyword identification.

Step 5: Created four-word sets to measure the factors in each stage of the brand (typesets). The sets were made up of the study variables’ corresponding activity concepts.

Step 6: compute the frequency of the research variables

Step 7: To reduce the bias caused by the varying length of the comments post, convert the data from step 6 to five-point scales (score).

Thematic analysis is a qualitative data analysis process that entails examining a data collection (such as transcripts from in-depth interviews or focus groups) and recognizing patterns in meaning throughout the data to extract themes. Thematic analysis focuses entirely or primarily on one level. Themes are detected within the explicit or surface meanings of the data using a semantic method, and the analyst is not searching for anything other than what a participant has been written in the review comments. Hence based upon the themes, the specific product can be extracted since the purpose of thematic analysis is to find themes, or relevant or intriguing patterns in data, and then utilize these themes to address the topic or make a point about a specific term.

Overall the proposed Product Review Opinion Based On Sentiment Analysis accurately predicts aspect term and group reviews using Detach Frequency Assort. Then, Systemize Polarity Shift classifies sentiments based on polarity without decision-making problems using SVM with bag of words model. Also, specific products are extracted using Revival Extraction based on thematic analysis. The next section explains the result obtained from the Product Review Opinion Based On Sentiment Analysis and discusses it in detail.

4 Result and discussion

This section provides a description of various implementation results and the performance analysis of the proposed model and also the comparison section to ensure enhancement of the proposed system.

4.1 Experimental setup

This work has been implemented and the simulation of the system was then done using python with the following system specification and the simulation results are discussed below,

Platform: Python

OS: Windows 8

Processor: Intel Core i5

RAM: 8GB RAM

4.1.1 Dataset description

This method consists of 34,000 samples of Amazon review comments are categorized into 10 categories. Also, the manufacturer, Manufacturer number, reviews date, reviews source, reviews text reviews username and test results of review comments are all kept in this file.

4.2 Simulation outputs

Due to different vendor listings, a single ASIN can have many names. There could also be a lot of names that are missing or have different titles (i.e. 8gb vs 8 gb, NAN for product names). We can observe that certain products have much more reviews than others based on the bar graph for ASINs, which could indicate a higher sale of those specific products. We can also see that the ASINs have a “right-tailed” distribution, which could indicate that particular products have larger sales, which could be associated with higher ASIN frequencies in the reviews. We also took the log of the ASINs to normalize the data so that we could obtain a more detailed picture of each ASIN, and we can see that the distribution is still right-tailed.

The first 19 ASINs show that consumers recommend the product, which is consistent with the reviews. rating / ASINs” analysis above, which indicates that the first 19 ASINs have good ratings between 4.0 and 5.0. Due to the smaller sample size, the remaining ASINs have varying outcomes, which should not be factored in.

The length of the vertical lines shows that for ASINs with lower frequencies as shown on the bar graph (top), their corresponding average review ratings on the point-plot graph (bottom) have significantly higher variance. As a result of the high variance, we believe that the average review ratings for ASINs with lower frequencies are not significant for our research. However, we believe that this is due to lower quality products, as seen by their lower frequencies for ASINs with lower frequencies. Furthermore, due to their significantly lower frequencies, the last four ASINs have no variance, and even though the review ratings are a perfect 5.0, we should not consider the significance of these review ratings due to their lower frequency.

4.3 Performance metrics of the proposed method

The performance of the proposed methodology and the obtained product review text data are detected by the following equation and the implementation results are obtained.

4.3.1 Accuracy

The accuracy of the clinical text data is calculated using, $Accuracy = [\frac{TP + TN}{TP + TN + FP + FN}] * 100$ (5)

TP- True Positive Value

TN- True Negative Value

FP- False Positive Value

FN- False Negative Value

Fig. 10

ASIN Frequency.

Fig. 11

ASIN Frequency with reviews. Rating.

Fig. 12

ASIN Frequency with reviews. doRecommend.

Figure 13 represents the overall accuracy of the proposed system; from the graph, it is clear that the proposed system gives high accuracy with 99.59% of review comments data. The accuracy of the proposed system is increased to 99.59% by using TF_ISF Approach since this approach extracts the data with statistical features.

Fig. 13

Overall accuracy.

4.3.2 Specificity

Specificity is derived from the equation $Specificity = \frac{True Negative}{True Negative + False Positive}$ (6)

The above-mentioned graph 14 clearly explains the specificity of the proposed model and the Specificity of the proposed model is about 98.6%. The highest of about 98.6% specificity is attained overall by the proposed methodology. The specificity of the proposed model is increased to 98.6% by using the TF-ISF approach since the quality of data is maintained by using this approach.

4.3.3 Sensitivity

Sensitivity is deduced using the formula $Sensitivity = \frac{TruePositive}{TruePositive + FalseNegative}$ (7)

Fig. 14

Overall Specificity

Figure 15 depicts the sensitivity of the proposed is determined as 98.80%. The sensitivity is overall between 97.50 to 98.80%. The sensitivity of the proposed system is determined by using the SVM approach since the recognition and division of data make the review text data more sensitive.

Fig. 15

Overall Sensitivity.

4.3.4 F1 score

F1 Score is defined as, $F 1 = \frac{2 \times (Precision * Recall)}{Precision + Recall}$ (8)

Where, $Precision = \frac{TP}{TP + FP}$ $Recall = \frac{TP}{TP + FN}$

The preceding graph 16 clearly shows the F1 Score of the suggested model, which is about 98.8 percent. As the number of samples increases, the specificity of the model also increases. Overall, the suggested technique achieves a high level of F1-Score of around 98.8 percent. The F1-Score of the proposed system is determined by using the Thematic analysis approach in which unique identification is required.

4.3.5 Precision

The closeness of two or more measurements to each other is known as precision. The formula is presented as, $Precision = \frac{TP}{TP + FP}$ (9)

Where,

TP - True Positive

FP - False Positive

Fig.16

Overall F1-Score.

Figure 17 represents the overall precision of the proposed system; from the graph, it is clear that the proposed system gives high precision with 98.60%. The precision of the proposed method is increased to 98.60% by using TF-ISF Technique.

Fig. 17

Overall Precision.

Fig. 18

Overall Recall.

4.3.6 Recall

A recall is defined as the ability of the model to accurately predict the output. The formula of recall is defined as, $Recall = \frac{TP}{TP + FN}$ (10)

Where,

TP - True Positive

FN - False Negative

From the above graph 18, it is observed that the recalls of the proposed system are about 98.0%. Hence the recalls increase with the increase in the number of samples. The recall of the proposed system is determined by using the Supply Vector Machine approach. Table 1 shows the accuracy, specificity, sensitivity, F1-Score, precision and recall of proposed system.

Table 1

Performance metrics of proposed approach

No. of products	Performance metrics
	Accuracy (%)	Specificity (%)	Sensitivity (%)	F1-Score (%)	Precision (%)	Recall (%)
5000	97.28	97.50	97.47	97	97.1	97.2
10000	97.69	97.54	97.54	97.5	97.2	97.2
15000	97.72	97.86	97.85	97.8	97.2	97.2
20000	97.91	97.99	98.00	98	97.3	97.3
25000	98.38	98.19	98.17	98	97.5	97.4
30000	98.92	98.30	98.30	98.4	97.6	97.5
35000	99.34	98.6	98.6	98.5	97.7	97.6
40000	99.54	98.75	98.5	98.5	98	97.8
45000	99.32	98.80	98.5	98.5	98.2	97.9
50000	99.39	98.91	98.80	98.0	98.6	98.0

This section describes the resultant performance of the proposed system. The next section describes a comparison of the various performance of the previous researches with the performance of the proposed method.

4.4 Comparison analysis

This section describes the various performance of the proposed method comparing with the results of previous methodologies and depicting their results based on various metrics. Table 2 depicts accuracy comparison of proposed system with existing approaches.

Table 2
Accuracy comparison

Methodology Accuracy (%)

RFC 82

MNB 83

SVM-RBF 83.54

SVM-LK 81

SGD 83.21

Proposed 99.59

Methodology	Accuracy (%)
RFC	82
MNB	83
SVM-RBF	83.54
SVM-LK	81
SGD	83.21
Proposed	99.59

The above graph 19 depicts the accuracy of product review text data is compared with the accuracy of the various previously proposed techniques. From the graph, it is clear that the stack accuracy of the proposed output achieves 99% which is 14% higher than the existing output when compared with baseline, RFC, MNB, SVM-RBF, and SGD. Table 3 depicts precision comparison of proposed system with existing approaches.

Fig. 19

Accuracy Comparison.

Table 3

Precision comparison

Methodology	Precision (%)
SS-LDA	80.78
Sentence-LDA	79.98
Biterm Topic Model	65.2
LDA	52.11
Proposed	98.6

The precision of Product review text data is compared with the precision of the various previously proposed techniques. From graph 20, it is clear that the stack precision of the proposed output achieves 99% which is 20% higher than the existing output when compared with baseline, SS-LDA, Sentence-LDA, Biterm topic model, and LDA. Table 4 depicts recall comparison of proposed system with existing approaches.

Fig. 20

Precision Comparison.

Table 4

Recall comparison

Methodology	Recall (%)
SS-LDA	84.76
Sentence-LDA	68.2
Biterm Topic Model	68
LDA	40
Proposed	98

Fig. 21

Recall Comparison.

The recalls of Product review text data are compared with the recalls of the various previously proposed techniques. From graph 21, it is clear that the stack recalls of the proposed output achieve 99% which is 15% higher than the existing output when compared with baseline, Sentence Segment LDA (SS-LDA), Sentence-LDA, Biterm topic model, and Latent Dirichlet Allocation (LDA). Table 5 depicts F1-Score comparison of proposed system with existing approaches.

Table 5

F1-Score comparison

Methodology	F1-Score (%)
UFAL/U	61.42
Basel./C	59.65
IIT-T./U	57
IIT-T./C	56.89
INSIG. /C	50
SS-LDA	63
Proposed	98.8

Fig. 22

F1-Score Comparison.

The F1-Score of Product review text data is compared with the F1-Score of the various previously proposed techniques. From graph 22, it is clear that the stack F1-Score of the proposed output achieves 99% which is 28% higher than the existing output when compared with baseline, UFAL/U, Basel. /C, IIT-T./U, IIT-T./C, INSIG. /C and SS-LDA. Table 6 depicts sensitivity comparison of proposed system with existing approaches.

Table 6

Sensitivity comparison

Methodology	Sensitivity (%)
RFC	82
MNB	83
SVM-RBF	84
SVM-LK	84
SGD	85
Proposed	98.8

Fig. 23

Sensitivity Comparison.

The Sensitivity of Product review text data is compared with the Sensitivity of the various previously proposed techniques. From graph 23, it is clear that the stack Sensitivity of the proposed output achieves 98.6% which is 14% higher than the existing output when compared with baseline, Request for Comments (RFC), MNB, Supply Vector Machine-Radial Basis Function (SVM-RBF), SVM-LK and Stochastic Gradient Descent (SGD). Table 7 depicts specificity comparison of proposed system with existing approaches.

Table 7

Specificity comparison

Methodology	Specificity (%)
RFC	82
MNB	83
SVM-RBF	84
SVM-LK	86
SGD	88
Proposed	98.6

Fig. 24

Specificity Comparison.

The Specificity of Product review text data is compared with the Specificity of the various previously proposed techniques. From graph 24, it is clear that the Specificity of the proposed output achieves 99% which is 16% higher than the existing output when compared with baseline, RFC, MNB, SVM-RBF, SVM-LK, and SGD.

Figure 25 illustrates the comparison in terms of accuracy Dataset such as Airline Twitter dataset, CD dataset, App dataset, and Movie dataset is compared with various previously proposed techniques. From graph 25, it is clear that the stack accuracy of the proposed output achieves 97.2%,97.1%,97.4%, and 97.5% which is 16% higher than the existing output when compared with baseline, SS-BED, HAN, ARC, Convolutional Recurrent Neural Network (CRNN), IWV, Attention-based Bidirectional Long Short Term Memory (AC-BILSTM) and ABCDM.

Fig. 25

Performance in terms of Accuracy.

Figure 26 depicts the comparison in terms of precision datasets such as Airline Twitter dataset, CD dataset, App dataset, Movie dataset is compared with various previously proposed techniques. From graph 26, it is clear that the stack accuracy of the proposed output achieves 97.7%,97.4%,97.2%, and 97.4% which is 12% higher than the existing output when compared with baseline, SS-BED, HAN, ARC, CRNN, IWV, AC-BILSTM and ABCDM.

Fig. 26

Performance in terms of Precision.

The above graph 27 depicts the comparison in terms of recall datasets such as Airline Twitter dataset, CD dataset, App dataset, Movie dataset is compared with various previously proposed techniques. From graph 27, it is clear that the stack accuracy of the proposed output achieves 97.6%,97.5%,97.5%, and 97.4% which is 20% higher than the existing output when compared with baseline, SS-BED, HAN, ARC, CRNN, IWV, AC-BILSTM and ABCDM.

Graph 28 illustrates the comparison in terms of the F1-Score dataset such as Airline Twitter dataset, CD dataset, App dataset, Movie dataset is compared with various previously proposed techniques. From graph 28, it is clear that the stack accuracy of the proposed output achieves 96.5%,97.6%,98%, and 97.8% which is 18% higher than the existing output when compared with baseline, SS-BED, HAN, ARC, CRNN, IWV, AC-BILSTM and ABCDM.

Fig. 27

Comparison based on Recall.

Fig. 28

Comparison based on F1-Score.

Figure 29 shows that the comparison in terms of sensitivity datasets such as the Airline Twitter dataset, CD dataset, App dataset, Movie dataset is compared with various previously proposed techniques. From the graph, it is clear that the stack accuracy of the proposed output achieves 97.5%,97.8%,98%, and 97.7% which is 16% higher than the existing output when compared with baseline, SS-BED, HAN, ARC, CRNN, IWV, AC-BILSTM and ABCDM.

Fig. 29

Comparison based on Sensitivity.

Fig. 30

Comparison based on Specificity.

Graph 30 illustrates the comparison in terms of the Specificity dataset such as Airline Twitter dataset, CD dataset, App dataset, Movie dataset is compared with various previously proposed techniques. From the graph, it is clear that the stack accuracy of the proposed output achieves 96.5%,96.7%,98%, and 97.5% which is 12% higher than the existing output when compared with baseline, SS-BED, HAN, ARC, CRNN, IWV, AC-BILSTM and ABCDM. Table 8 depicts the performance analysis of proposed system for various existing dataset.

Table 8

Comparison of various datasets with proposed in terms of performance metrics

Datasets	Performance metrics of proposed system
	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Sensitivity (%)	Specificity (%)
Airline Twitter	97.2	97.7	97.6	96.5	97.5	96.5
CD dataset	97.1	97.4	97.5	97.6	97.8	96.7
App dataset	97.4	97.2	97.5	98	98	98
Movie dataset	97.5	97.4	97.4	97.8	97.8	97.5

Overall, the proposed Product Review Opinion Based On Sentiment Analysis achieves maximum accuracy, specificity, sensitivity, F1-Score, precision, and recall of 99.59%, 98.6%, 98.8%, 98.8%, 98.6%, and 98% respectively.

5 Conclusion

These days, social media is getting a lot of attention. Public and private opinions on a variety of topics are constantly expressed and distributed on a variety of social media platforms. Amazon is one of the most prominent social networking platforms. Customer’s perspectives on the crucial to success in the marketplace are analyzed using Amazon sentiment analysis. The application also combines frequency and inverse sentence frequency techniques with a more accurate machine-based learning approach for sentiment analysis. As a result, the program’s sentiment will be divided into three categories: good, negative, and neutral. The product review text data is extracted with high accuracy, Sensitivity, F1-Score, Specificity, Recall, and Precision of 99.59 %, 98.80%, 98.80%, 98.6%, 98.0%, and 98.60%. The results show that the proposed method is compared with other existing techniques and the proposed outperforms all the other existing techniques.. In future, this work will be extended with enhanced techniques such as optimization algorithms, ensemble learning, deep learning, and neuro-fuzzy models to improve convergence rate and computational efficiency.

References

Tang

, Fu

, Yao

and Xu

, Aspect based fine grained sentiment analysis for online reviews, Information Sciences 488 (2019), 190–204. https://doi.org/10.1016/j.ins.2019.02.064.

Rezaeinia

S.M.

, Rahmani

, Ghodsi

and Veisi

, Sentiment analysis based on improved pre-trained word embeddings, Expert Systems with Applications 117 (2019), 137–149. https://doi.org/10.1016/j.eswa.2018.08.044.

Behera , Ranjan Kumar , Monalisa Jena , Santanu Kumar Rath , Sanjay Misra , Co-LSTM: Convolutional LSTM model for sentiment analysis in social big data, Information Processing & Management 58(1) (2021), 102435.

Mowlaei

M.E.

, Abadeh

M.S.

, Keshavarz

Aspect based sentiment analysis using adaptive aspect-based lexicons, Expert Systems with Applications (2020), 148. https:// doi.org/10.1016/j.eswa.2020.113234.

Bajpai

, Hazarika

, Singh

, Gorantla

, Cambria

, Zimmerman

Aspect-Sentiment Embeddings for Company Profiling and Employee Opinion Mining, arXiv preprint, arXiv, 1902.08342. 2019.

Ibrahim

N.F.

and Wang

, Decoding the sentiment dynamics of online retailing customers: time series analysis of social media, , Comput HumBehav 96 (2019), 32–45.

Kumari

, Tanwar

, Tyagi

, Kumar

, Parizi

R.M.

and Choo

K.K.R.

, Fog data analytics: a taxonomy and process model, , J Netw Comput Appl 128 (2019), 90–104.

Singh

, Garg

, Batra

and Kumar

, Probabilistic data structure-based community detection and storage scheme in online social networks, , Fut Gen Comput Syst 94 (2019), 173–184.

Reddy

G.T.

, Reddy

M.P.K.

, Lakshmanna

et al., Analysis of dimensionality reduction techniques on big data, , IEEE Access 8 (2020), 54776–54788.

10.

Ghani

, Bajwa

and Ashfaq

, A fuzzy logic based intelligent system for measuring customer loyalty and decision making, Symmetry 10 (12) (2018), 761.

11.

Diwakar

, Kumar

, Gour

, Khan

A.U.

Proposed machine learning classifier algorithm for sentiment analysis. In: Sixteenth International Conference on Wireless and Optical Communication Networks, IEEE Press (2019).

12.

Karthika

, Murugeswari

, Manoranjithem

, Sentiment analysis of social media network using random forest algorithm, In: International Conference on Intelligent Techniques in Control, Optimization and Signal Processing, IEEE Press (2019), 1–5.

13.

Shaheen

, Awan

S.M.

, Hussain

and Gondal

Z.A.

, Sentimentanalysis on mobile phone reviews usingsupervised learning techniques, , International Journal of ModernEducation and Computer Science 7 (2019), 32–43.

14.

Aziz

A.A.

and Starkey

, Predicting supervise machine learning performances for sentiment analysis using contextual-based approaches, , IEEE Access 8 (2020), 17722–17733.

15.

Khan

, et al., Sentiment classification of user reviews using supervised learning techniques with comparative opinion mining perspective. In: Arai K., Kapoor S. (eds) Advances in Computer Vision, Advances in Intelligent Systems and Computing, Springer, Cham. 944 (2020), 23–29.

16.

Rao

, Huang

, Feng

and Cong

, Lstm with sentence representations for document-level sentiment classification, , Neurocomputing 308 (2018), 49–57.

17.

Aziz , Azwa Abdul , Andrew Starkey , Marcus Campbell Bannerman , Evaluating cross domain sentiment analysis using supervised machine learning techniques, In 2017 Intelligent Systems Conference (IntelliSys), IEEE (2017), 689–696.

18.

Basiri , Mohammad Ehsan , Shahla Nemati , Moloud Abdar , Erik Cambria and Rajendra Acharya

, ABCDM: An attention-based bidirectional CNN-RNN deep model for sentiment analysis, , Future Generation Computer Systems 115 (2021), 279–294.

19.

Tian , Yuanhe , Guimin Chen , Yan Song , Aspect-based Sentiment Analysis with Type-aware Graph Convolutional Networks and Layer Ensemble, In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (2021), 2910–2922.

20.

Rashid , Aamir , Ching-yu Huang , Sentiment Analysis on Consumer Reviews of Amazon Products, International Journal of Computer Theory and Engineering 13(2) (2021)

21.

Raviya

, Mary Vennila

Deep CnnWith Svm-Hybrid Model For Sentence-Based Document Level Sentiment Analysis Using Subjectivity Detection.

22.

Agarwal

, Xie

, Vovsha

, Rambow

, Passonneau

R.J.

Passonneau, Sentiment analysis of twitter data, In Proceedings of the workshop on language in social media (LSM 2011) (2011), 30–38.

23.

Kouloumpis

, Wilson

, Moore

Twitter sentiment analysis: The good the bad and the omg, In Fifth International AAAI conference on weblogs and social media (2011).

24.

, Huang

and Bhayani

, Twitter sentiment analysis, , Entropy 17 (2009), 252.

25.

Pak

and Paroubek

, Twitter as a corpus for sentiment analysis and opinion mining, In , LREc 10 (2010), 1320–1326.

26.

Liu

, Sentiment analysis and opinion mining, Synthesis Lectures on Human Language Technologies 5(1) (2012), 1–167.

27.

Vishwanathan

S.V.M.

and Murty

M.N.

, SSVM: a simple SVM algorithm, In, IJCNN’02 (Cat. No. 02CH0) IEEE , Proceedings of the 2002 International Joint Conference on Neural Networks 3 (2002), 2393–2398.

28.

Joffe

, Thematic analysis. Qualitative research methods in mental health and psychotherapy: A guide for students and practitioners , 1 (2012), 210–223.

29.

Guest

, MacQueen

K.M.

and Namey

E.E.

, Introduction to applied thematic analysis, Applied Thematic Analysis 3(20) (2012), 1–21.

30.

Kaur

and Gupta

, A survey on sentiment analysis and opinion mining techniques, Journal of Emerging Technologies in Web Intelligence 5(4) (2013), 367–371.

31.

Medhat

, Hassan

and Korashy

, Sentiment analysis algorithms and applications: A survey, Ain Shams Engineering Journal 5(4) (2014), 1093–1113.

32.

Liu

, Sentiment analysis and subjectivity, , Handbook of Natural Language Processing 2 (2010), 627–666.

33.

Lin

, He

Joint sentiment/topic model for sentiment analysis, In Proceedings of the 18th ACM conference on Information and knowledge management (2009), 375–384.

34.

Hussein

DMEDM.

, A survey on sentiment analysis challenges, Journal of King Saud University-Engineering Sciences 30(4) (2018), 330–338.

35.

Liu