Sentiment aggregation of targeted features by capturing their dependencies: Making sense from customer reviews

Highlights

• Enables product designers to enhance the coverage of mapping review features to the most appropriate product components.

• Proposes an approach for sentiment aggregation of a product’s high-level and detailed features.

• Proposes a novel text mining approach for ascertaining the dependent features of a targeted feature.

Abstract

Ideation is an important phase in the new product development process at which product designers innovate and select novel ideas that can be added as features to an existing product. One way to find novel ideas is to transfer uncommon features of products of other domains and integrate them into the product to be improved. However, before incorporating such targeted features into the product, they need to be evaluated against the customers’ acceptance in social media using sentiment aggregation tools. Despite the many studies in sentiment analysis, mapping the customers’ opinions towards both high-level and technical features of a product extracted from social media to their best corresponding component in that product is still a challenge. Furthermore, none of the existing approaches ascertains the sentiment value of a targeted feature by capturing its dependencies on other features. In this paper, to address these drawbacks, we propose the sentiment aggregation framework for targeted features (SA-TF). SA-TF determines the sentiment of a targeted feature by assisting product designers in the tasks of mapping the features discussed in the reviews to the right product components, sentiment aggregation and considering feature dependencies to determine their polarity. The superiority of the different phases of SA-TF is demonstrated with experiments and comparing it with an existing approach.

Introduction

New product development (NPD) is a multi-step process that enables product designers to bring novel and innovative products to the market. Ideation is the first step of NPD in which product designers come up with ideas that introduce newness into a product (Cooper, 2008; Fojt, 1996; Koen et al., 2002; Koen, Bertels, & Kleinschmidt, 2014). One way for product designers to find ideas that introduce newness to a product is to search for the uncommon and unique features from products of other domains (woo Kang & Tucker, 2015; Moreno, Hernandez, Yang, & Otto, 2014; Tucker & Kang, 2012; Mirtalaie, Hussain, Chang, & Hussain, 2017a). In this paper, we term such novel ideas as targeted features. To diminish the risk of product failure, product designers, after determining the targeted features ascertain customers’ opinions toward them. This assists them to decide whether or not to consider a targeted feature any further (Geise, 2017; Grönlund, Sjödin, & Frishammar, 2010).

Due to the accelerated evolution in Web 2.0, customers’ opinions are easily accessible from Social Media (SM) portals. Feature-based sentiment analysis as a field of study is utilized to automate the extraction of people’s opinions and emotions from SM towards different features of a product (Hussein, 2018; Ireland & Liu, 2018; Jin, Ji, & Gu, 2016; Liu, 2012; Liu, Jiang, & Zhao, 2018; Rathan, Hulipalled, Venugopal, & Patnaik, 2018; Tuarob & Tucker, 2015; Tucker & Kim, 2011; Wang & Wang, 2014). After the features from customer reviews are extracted, they need to be aggregated for product designers to have a meaningful summary of what customers are saying towards a feature. Sentiment aggregation as a sub-field is used to address this task. This task also handles the mismatch cases when customers in their reviews refer to a feature with a different terminology as compared to how it is mentioned in a product’s standard terminology. Techniques such as topic-modelling (Lau, Li, & Liao, 2014; Suleman & Vechtomova, 2015; Ye, Li, Guo, Qian, & B, 2015) or ontology-based (Agarwal, Mittal, Bansal, & Garg, 2015; Mukherjee & Joshi, 2013, 2014; Tamilselvam, Nagar, Mishra, & Dey, 2017; Umamaheswari & Priya, 2016; woo Kang & Tucker, 2016) are used to reduce the sparseness between the features extracted from customer reviews and how they are mentioned according to the product’s standard terminology. This is done by mapping the extracted features from customer reviews to the high-level components of a product, available from a generic ontology or lexicon like ConceptNet and WordNet. While these approaches have no issues in matching the commonly known features of a product, they fail when customers in their reviews speak about a product’s technical features. This is because such technical features are not commonly available in the generic lexicon, and thus they remain unmatched (Ganesan, Zhai, & Viegas, 2012). Thus, an approach is needed to map the product’s technical features mentioned in SM reviews to a detailed product lexicon, so that they are utilized by the product designers (woo Kang & Tucker, 2016).

A product’s feature does not work in isolation. Instead, it not only collaborates with other features to achieve its purpose but also may influence different other features to deliver their output (Apel, Kolesnikov, Siegmund, Kästner, & Garvin, 2013). For example, a camera’s ‘advanced scene recognition system (ASRS)’ works collaboratively with features such as ‘flash’ and ‘exposure’ to accomplish its job. Furthermore, ASRS also influences other features, such as the camera’s ‘image’ in achieving its functionality. Such features are the dependent features of ASRS. Since such dependencies play a vital role in fulfilling a feature’s tasks (Ferber, Haag, & Savolainen, 2002), while computing the sentiment score of a targeted feature, we argue that the polarity of its dependent features too should also be considered. Although (Donaldson & Calder, 2012; Lienhard, Greevy, & Nierstrasz, 2007; Siegmund & Kolesnikov, 2012) detect features interactions/dependencies in various domains, to the best of our knowledge, they have not been considered while analysing the polarity of a targeted feature. Thus, the sentiment value which they determine for that targeted feature may not be accurate.

As (Geise, 2017; Grönlund et al., 2010) mention, product designers use customers’ opinions toward the targeted features to decide whether or not to consider them further. Thus, addressing the gaps mentioned above are important for product designers when they reach the stage of making informed decisions about which targeted features to consider beyond the ideation step of NPD. In this paper, we address these gaps, and the research question we solve is:

How can a product designer capture technical details about a feature mentioned in customer reviews and ascertain the polarity of a targeted feature by considering its dependencies?

We fill these gaps in this paper by proposing the Sentiment Aggregation framework for Targeted Features (SA-TF). SA-TF builds on the previous work of extracting product features from reviews along with their associated degree of sentiment (Mirtalaie et al., 2017) and contributes to the theory and practice of sentiment analysis from SM in three ways as follows:

• It contributes to the theory by developing an approach that maps the extracted product features (both high-level and technical) to the most appropriate component of the product. This is achieved by generating a product tree from the product’s specification document and developing a heuristic approach to map each extracted (review) feature to its best representative node in a tree.

• It contributes to the theory by developing an approach that utilises hierarchical relations of a feature (both high-level and technical ones) when determining its overall aggregated polarity. It also develops an approach that determines the sentiment of a targeted feature by capturing its dependencies with other features of the product.

• It contributes to practice by assisting product designers in taking the full potential of customers’ opinions from SM. Accurately determining customer opinion will lead product designers to innovate in their products better and incorporate ideas which are supported by the customers. This will diminish the risk of product failure and maintain a competitive advantage against their competitors.

The paper is organized as follows. Section 2 presents a review of the literature and shows how the gap addressed in the paper has theoretical groundings regarding the use of SM in the literature. Section 3 presents the adopted research methodology and the data analysis done by SA-TF. Section 4 presents the technical details on the process of feature matching in SA-TF. Sections 5 and 6 details on the process of sentiment aggregation and dependency consideration, respectively. Section 7 explains and compares the results of SA-TF with other approaches. Section 8 summarises the main findings along with discussing the theoretical and practical implications of the study. Section 9 concludes the paper with a discussion on future work.