Discovering market insights from online product reviews through sentiment analysis

Abstract

Online shopping is increasingly preferred, products and services are compared, and customer reviews are shared on numerous digital platforms thanks to the successive iterations of the web and the widespread of the Internet. This type of customer feedback is of great value for marketing activities. As the content created on digital platforms is easily accessible and the sentiment of text data can be analyzed using various learning algorithms, marketers can conveniently produce market intelligence about product and service perceptions and customer behaviors. Due to the continuously expanding volume of customer evaluations, approaches based on manual analysis are insufficient to provide a broad perspective of consumer opinions for a product. The need to automate the process has led to the development of various analysis tools. As a result, it has become necessary to build data analysis systems that can automatically evaluate customer reviews and summarize their behavior and opinions. Using natural language processing tools and deep learning algorithms to extract insights from digital platforms, sentiment analysis can identify the emotional value of communications for businesses and help marketers understand their customers better. More effective marketing strategies can be designed using unstructured data. Accordingly, analyzing unstructured data to identify aspects, sentiments, emotions, keywords, and purposes can help to improve the market research of a business. Furthermore, marketers may also use data analysis to find customer pain points, determine their bottlenecks, and track their competitors. This study essentially presents a systematic review of text and market analytics. It also delivers detailed findings and theoretical backgrounds of Turkish NLP applications, transformers, and market research methods. Firstly, the rapid growth of digital platforms and e-commerce under technological improvements and pandemic conditions is acknowledged. During the lockdown, how e-commerce, which had already experiened a significant growth trend, accelerated and grew is explained. Secondly, how to determine the metrics that measure consumer habits and behaviors in digital marketplaces are discussed. Macro-level indicators related to e-commerce volume are given for both Turkey and the world. A framework for which factors have affected e-commerce recently and which trends it can follow in the future is presented. Thirdly, research on how customer satisfaction and loyalty can be measured considering a customer's experience on digital platforms is conducted. Issues that affect and shape customer relations in digital channels are defined, and which metrics to follow for their measurement is discussed. Finally, the first chapter investigates how customer satisfaction and loyalty can improve customer experience in e-commerce platforms and their impact on profitability. The literature related to text analytics and market research is thoroughly reviewed. In the text analytics section, the classification of texts is primarily discussed. Sentiment analysis involves the task of assigning text data to a sentiment class. The methods used in text classification are mentioned, and the ways to evaluate classification performance are explained. Subsequently, how to represent text data that is to be introduced as an input to an NLP model is explained. The word embeddings used to describe text data are mentioned, and their positive and negative aspects are explained. An example discussing how two words with the same meaning can make a difference in word embeddings is provided. Natural language processing problems are sought to be answered by deep learning. The fundamentals of deep learning are explained in detail, and how it can work with text data is presented. The difficulties of working with text data are defined, and the methods developed for deep learning algorithms to overcome these difficulties are described. Using the knowledge and experience gained from similar natural language processing problems instead of learning from scratch is examined under the heading of transferred learning. Finally, the theoretical background of attention mechanisms, which are frequently used in the solution of NLP problems, is explained. The solution proposed by the transformers is examined. Sentiment analysis helps businesses monitor their customers' perceptions of their products and provides market intelligence about the ecosystem. Online product reviews may be a significant data source for gaining market insights and analyzing the competitors. It also allows for learning what customers in the target segment think about the competitors that are operating in the same market. These deep insights help fill gaps in products and services and help sellers stay ahead of the competition. In addition, understanding customers' experience on digital platforms provides information about customers' satisfaction levels and enables to monitor their interactions with the digital channels. In addition, sentiment analysis helps one evaluate how a customer group perceives a product or service. Diversifying digital channels used by online platforms shape the marketing analytics processes. Market strategies are reshaped with insights gained from the market using analytical tools. These tools can be used effectively for businesses to achieve their operational and financial goals with the help of marketing data. The first metric examined in marketing analytics within this context is the market share, which is one of the most critical metrics used by businesses to determine their current situation and the market targets to be followed. In addition, the need share and the wallet share metrics, which focus on customers and show the share of businesses in customer expenditures, are explained. Two different approaches to the wallet share, where customers are examined individually and evaluated cumulatively, are also discussed. The literature on Turkish sentiment analysis has generally employed traditional machine learning methods; however, models using deep learning and transformative architecture have recently become more popular. Sentiment analysis has been studied in various fields, using movie reviews, hotel reviews, social media shares, and online marketplaces reviews. The size of data sets, the way data annotation is performed (manually vs. automatically), and the data preprocessing methods are the main differentiators of these studies. Consequently, the end-to-end sentiment classification task is examined step by step. Firstly, how tokens represent text data for the BERT model, used as a sentiment classifier, and how the word embedding method used means the semantic relationship in the text is described. BERT's solution proposals, especially for text classification tasks, are shown in detail. Then, how the pre-training and fine-tuning tasks perform the learning process using attention mechanisms in the modeling phase, and performance monitoring metrics are explained. The level of detail of the sentiment analysis is given, and difficulties encountered in the natural language processing problems in an agglutinative language, Turkish, are explained. The wallet allocation rule and how the sellers can use it to determine a better market strategy is explained. Furthermore, how the wallet share can affect the financials and how the enterprises can turn this information into an advantage knowing this effect is discussed. Experiences are gained using customer reviews on an e-commerce marketplace for online sellers within the scope of this study. Data analysis, how product categories are determined, data annotation, adjustments to the data set, data preprocessing, identifying testing and training data sets, tokenization of text data, modeling, generating predictions, and performance monitoring processes are described. The model's outputs, estimation results, are introduced as inputs to calculate the wallet share. Finally, the wallet shares that the customers have allocated to the sellers are counted. Different multi-class classification model alternatives (5-class, 3-class, and 2-class) are evaluated, and the 3-class classification is eventually selected. The 3-class classification weighted-averaged f1 scores are 0.8631 for the category of case, 0.8626 for men's perfume, and 0.9016 for mobile phones. Since the wallet allocation rule calculates the customer-based wallet share, the reviews are randomly assigned to the customers. The models are developed using different customer sizes (100, 300, and 500) to see the effects on estimating the wallet share. The model developed with the biggest customer size gives the lowest absolute error. The assumption can be summarized that the wallet shares calculated with the online reviews - collected from three categories, grouped into three sentiment classes, and received from 500 customers. Needs for improvement are specified, and a brief guideline is identified for topics that can be considered in future work. The learning model's automatic retraining structure is proposed, and aspect-based sentiment analysis is described.