Using machine learning and natural language processing to measure consumer reviews for product attribute insights

Researchers from Western University, SUNY Buffalo State College, University of Cincinnati, and City University of Hong Kong published a caller insubstantial successful the Journal of Marketing that presents a methodological model for managers to extract and show accusation related to products and their attributes from user reviews.

Understanding however factual merchandise attributes signifier higher-level benefits for consumers tin payment assorted firm teams. Concrete, oregon "engineered attributes" notation to method specifications and merchandise features. For example, successful the discourse of tablet computers, specified attributes see RAM, CPU, weight, and surface resolution. Understanding however combinations of these lower-level attributes signifier higher-level benefits, oregon "meta-attributes," for consumers, specified arsenic Hardware and Connectivity, tin supply managers with actionable insights. Sales teams request to recognize the higher-level merchandise benefits that drive consumer buying behavior. Product plan teams indispensable pass with engineering and manufacturing to recognize the relationships betwixt the product's method specifications and its perceived benefits. Engineering teams request to beryllium capable to estimation the trade-offs of method subcomponents to physique the merchandise exemplary that fulfills the much abstract benefits associated with the product's meta-attributes.

The accepted method of surveys tin beryllium time-consuming and whitethorn output inconsistent results crossed antithetic sampling periods. Thus, determination remains a important spread successful mentation and practice: How tin the nexus betwixt engineered attributes and meta-attributes beryllium uncovered straight from user input to pass managerial decisions? 

To capable this gap, the probe squad devised a methodological model based on machine learning and natural language processing to get an embedded practice of merchandise attributes. Specifically, embedded practice describes (represents) textual information specified arsenic idiosyncratic merchandise attributes utilizing the words that situation specified textual information (i.e., the contextual information) successful user reviews. The practice is quantified utilizing neural networks that alteration mathematically measurement of the degrees of similarity betwixt assorted merchandise attributes based connected however they are described by consumers themselves (i.e., the contextual information), frankincense revealing similarities and differences successful the attributes' usage by consumers.

From this embedded representation, the exemplary past identifies multi-level clusters of merchandise attributes that bespeak the levels of abstract merchandise benefits. "In different words," says Wang, "this caller method algorithmically extracts consumers' ain words successful the reviews they constitute to quantify circumstantial contexts that are expressed successful narration to idiosyncratic merchandise attributes.

This past enables grouping the merchandise attributes unneurotic based connected their contextual similarities to uncover higher-level benefits that tin power user restitution oregon dissatisfaction with a product." The sentiments associated with these meta-attributes are utilized to measure objects of managerial interest, specified arsenic a merchandise oregon brand, and past tin spell deeper to analyse which engineered attributes chiefly thrust user sentiments successful narration to the meta-attributes.  

The probe makes 3 main contributions. First, it provides a methodological model for managers to extract and show accusation related to products and their attributes from user reviews. As He explains, "Because our model exploits the contexts surrounding merchandise attributes expressed successful user reviews, managers tin usage it to straight monitor how meta-attributes germinate wrong brands and to comparison brands wrong a merchandise class to pass their product-related decisions. We supply validations that our hierarchical operation of meta-attributes adequately approximates consumers' underlying review-writing behaviors." Second, the probe extends sentiment investigation of user reviews by demonstrating hierarchical sentiment analysis, which aggregates sentiment scores associated with idiosyncratic attributes based connected an property hierarchy.

Starting astatine the reappraisal level, sentiment scores can beryllium aggregated upwards to output insights for assorted units of analysis, specified arsenic SKU, merchandise series, and brands. "Using hierarchical sentiment analysis, managers tin spell beyond relying connected reappraisal ratings, which lone picture products arsenic a full and cannot beryllium accredited to circumstantial merchandise attributes. We show that this flexible attack to sentiment investigation tin make tailored dashboards and perceptual maps from user reviews that tin pass managerial decisions," says Curry.  

Third, the survey uses user reviews of tablets to supply a applicable objection of the method. In particular, it analyzes user sentiments astir Hewlett-Packard and Toshiba to research imaginable reasons wherefore these brands yet discontinued their tablet merchandise lines. Ryoo explains that "Using our property hierarchy, we measure their meta-attributes and past drill down to the level of engineered attributes to find that the constricted fig of apps disposable for HP's tablets and the thickness and value of Toshiba's tablets were the main drivers of consumers' antagonistic sentiments astir the products.

We past analyse the meta-attributes of market-leading brands Samsung and Apple to research imaginable drivers of their successes." Berger et al. enactment that "for information to beryllium useful, researchers indispensable beryllium capable to extract underlying insight—to measure, track, understand, and construe the causes and consequences of marketplace behavior." In this sense, this method is highly utile for processing selling strategies due to the fact that it provides invaluable insights into the relationships betwixt merchandise attributes and user valuations.

---

---