UR: SMART–A tool for analyzing social media content

Social media analysis is a vibrant research area and multiple benefits for digital marketing are discussed in literature (e.g., discovery of brand fans, most important social mentions, etc.) (Micu et al. 2018; Perakakis et al. 2019; Alalwan et al. 2017). Accordingly, a huge variety of social media analysis approaches have been developed in recent years (e.g., Vashishtha and Susan 2019; Stieglitz et al. 2018; Yue et al. 2019). To narrow the scope, we focus on sentiment analysis, classification and clustering in particular, because their effectiveness has been proven in practice (e.g., Alalwan et al. 2017) and they can be purposefully combined to come to a mixed method approach (e.g., Stieglitz et al. 2018) that enables social media data analysis from different angles, which is the research objective of this study.

The literature recognizes the potential of combining different kinds of analysis methods for social media data (Stieglitz et al. 2018). Generally, the combination of analysis methods is widely used in research due to the fact that mixing methods results in a more accurate and complete depiction of the phenomenon under investigation (Johnson 1995; Johnson and Christensen 2000; Patton 1990; Tashakkori and Teddlie 1998). Correspondingly, in different research fields, combinations of different analysis methods were successfully applied, e.g., for IS research methodologies (cf. Gable 1994), process-oriented quality management (cf. Stracke 2006) or Method Engineering (cf. Brinkkemper 1996; Ralyté et al. 2003; Tolvanen et al. 1996). The use of complementary methods is generally thought – as confirmed by all the above-mentioned examples – to lead the investigation to more valid results either by extending the investigation’s perspectives or by uncovering new or deeper dimensions. This rests on the premise that the weaknesses in each single method will be compensated by the counter-balancing strength of another (Jick 1979). Since combining methods is time-consuming and compensation relations between methods are not always obvious, a multi method approach is not without some shortcomings and may not be suitable for all research purposes (Jick 1979). The selection of the appropriate methods for combination needs to be carefully justified and made explicit in terms of the definite research aim.

In the field of social media analysis, Chen et al. (2014) for instance investigate the major concerns and worries of students in their academic and private lives by looking at Twitter feeds. Therefore, they use content analysis (and the naïve bayes algorithm) to classify the tweets and analyze the categories with the help of statistical frequency measures (total/relative number of tweets for each category, etc.) (cf. Chen et al. 2014). Similarly, Tinati et al. (2014) propose to analyze Twitter data via content analysis and quantitative measures (e.g., number of followers, interaction with others, retweets, etc.) to provide a richer and more contextualized view of the data. Borgmann et al. (2016) analyze the Twitter posts at a medical conference via a combination of descriptive statistics and content analysis to uncover those topics which the participants discussed most often. Further, Brown et al. (2019) show how the combination of various text analyses approaches on Instagram posts helps to recognize young people experiencing serious psychological crisis. Additionally, the research of Tseng et al. (2019) integrates various kinds of social media data for building a hierarchical structure with sustainable supply chain capabilities in the context of the textile industry. Lastly, an example from the food industry is provided by He et al. (2013), who introduce a text mining process for social media content using different kinds of analyses, to enable competitive comparisons for the pizza industry. Even though research on the combination of different kinds of analysis methods has a long tradition, existing research primarily focuses on domain-specific problems that are solved with a predefined set of analysis methods. A general approach that proposes a multi-faceted combination of analysis methods – to be applied independent of a certain domain – is missing in the social media analysis literature to the best of our knowledge. With this research, we aim to close this gap.

As already mentioned in Sect. 1, there is a vast number of available analytics tools in the context of social media on the market. We took an up-close look at the most popular tools (e.g., Falcon.io, Facelift, Buffer, Keyhole, Brandwatch) and analyzed their functionality in terms of textual data processing (see Table 1). For this, we installed demo versions of the tools as well as interviewed sales representatives from these companies on the specific features.

Considering Sect. 2.1, sentiment analysis, classification and clustering have been introduced as established social media analysis approaches that are promising for the development of a mixed method approach. Furthermore, the literature in Sect. 2.2 highlights the benefits of using quantitative measures (e.g., number of followers, retweets, etc.) to complement findings from the content or text analysis for instance. Accordingly, we compare existing social media tools with the help of the above-mentioned features – i.e., support of (1) “multi-language sentiment analysis”, (2) “classification”, (3) “clustering” and (4) “quantitative analysis” – and further add the criterion of whether a (5) “mixed method approach” is supported or not, which was derived from our research question (see Table 1).

None of the tools provides a mixed method approach, considering both a quantitative (e.g., number of likes) and qualitative (e.g., text classification) analysis as focused in this research. Additionally, multi-language sentiment analysis is not supported by some tools. This is especially important for this study since the analysis of German social media posts needs to be considered for our context. Further, it should be mentioned that the functionality of some tools extends the focus of the analysis at large. Sales representatives from “Falcon.io” stated that the tool’s focus is the publishing functionality for social media posts. Other sales representatives, for instance from “Facelift”, stated that their tool is considered a social media management tool and therefore offers further domain-specific analysis functionalities. To sum up, it can be stated that the available tools provide a vast number of highly beneficial functionalities for digital marketing, which, however, are not focused on the type of social media analysis that is pursued in our research (e.g., no mixed method approach). Hence, the comparison of the tools unveils the need for a dedicated mixed method social media analysis tool in case deeper insights of the data analysis are required by a company.

The design science (DS) paradigm has its roots in engineering and the science of the artificial (Simon 1996) and is fundamentally a problem-solving paradigm (Hevner and Chatterjee 2010). Research projects that follow the DS paradigm are concerned with the design, development, implementation, use, and evaluation of socio-technical systems in organizational contexts. Design scientists produce and apply knowledge of tasks or situations to create effective artifacts (March and Smith 1995). These artifacts are delineated in different structured forms such as software, formal logic, and rigorous mathematics to informal natural language descriptions (Hevner et al. 2004).

An important step in DS research is to prove the utility, quality, and efficacy of the artifact via well-executed evaluation methods. Since the artifact’s performance is related to the environment in which it is used, an incomplete understanding of the environment can induce inappropriately designed artifacts (March and Smith 1995). Therefore, Hevner’s “design cycle” (Hevner 2007) substantiates the importance of constructing and evaluating the artifact, and suggests balancing the efforts spent for both activities, which must additionally be convincingly based in relevance and rigor (Hevner 2007).

However, DS research aims not only to increase the technical knowledge base through the designed artifact but also contributes to the scientific knowledge base by performing the research process rigorously (e.g., by reflecting the construction and/or evaluation of the artifact). Especially in recent sources, the focus on the contribution to the scientific knowledge base (or scientific theories) is strongly discussed (Hevner 2007; Gregor and Hevner 2013; Baskerville et al. 2018) and possible contributions are vaguely outlined. It is noted as the key differentiator between the professional practice of building IT artifacts and DS research (Hevner 2007; Gregor and Hevner 2013) and represents Hevner’s “rigor cycle” (Hevner 2007).

Last but not least, the use and evaluation of the artifact in the application domain lead to several practical contributions, which constitute Hevner’s “relevance cycle” (Hevner 2007) and which can be seen as self-evident objectives of a DS research project.

To establish design requirements for UR:SMART, we consulted our network of practice partners from various branches such as technology, consumer products and mechanical engineering, which operate in both the B2C and B2B sector. In total, eleven companies agreed to take part in our study. All these companies openly stated their commitment to social media analysis as part of their digital marketing strategy or planned to profoundly invest in corresponding data analysis efforts in the near future. Therefore, they were considered suitable candidates for our investigation.

For the derivation of requirements, we used (1) semi-structured interviews and (2) literature about social media analysis tools (e.g., Aggarwal and Zhai 2012; Batrinca and Treleaven 2015; Guesalaga and Kapelianis 2016; Fan and Gordon 2014; Kohli et al. 2018; Maynard et al. 2012; Stieglitz et al. 2018). Semi-structured interviews put us in a position to adapt the questions to a company’s specific context and ask for more in-depth information on demand, which helped to uncover specific challenges of social media analyses in practice as well as desired key functionalities of a possible solution (Corbin and Strauss 2015; Meuser and Nagel 2009). Our interview partners were mainly social media managers with profound knowledge in this field. Moreover, in some cases we were able to interview C-level executives.

First, we asked whether the companies were using automated tools to analyze their social media data. Second, the specific aim and field of use of each type of social media analysis performed was studied (e.g., on products, brand building, reputation or recruiting). Another object of investigation was the analysis method, describing the way in which social media content was being analyzed (either manually or in an automated way). Furthermore, limitations of the tools and desired enhancements of their analysis functionality were considered.

For the systematic analysis of the data we transcribed the interview results and drew upon a twofold procedure for the visualization of the results as shown in Table 2. Following Mayring (2000), we used a deductive content analysis on the one hand for the questions regarding the use of a social media tool and the analysis method used. These questions were asked in a binary manner (e.g., yes/no, automated/manual). On the other hand, we used an inductive content analysis for the questions regarding the focus and desire of the companies to adequately consider their individual character and goals. However, we need to mention that the derived results should not be seen as exhaustive since it only reflects the situations of the interview partners.

By doing so, we were able to discover several problems as well as limitations faced by these companies concerning social media analysis. Overall, it became evident that the efforts required for a manual data analysis are still a big challenge throughout all sectors independent of the company size. Most companies rely on manual efforts for data analyses, as only 36% of the companies interviewed use a commercial social media analysis tool, whereas the rest purely rely on the integrated analytics functionalities of the individual social media platforms (e.g., Facebook Analytics), which are generic and very limited. Regarding the field of use of the social media activities, most companies put an emphasis on brand and reputation building as well as improving their products based on the “voice of the customer” (Pande et al. 2014). In this regard, the desire to gain more detailed insights into the customers’ opinions, which extend the findings retrieved from a manual analysis, became evident during the interviews. Several questions such as, “How can product criticism be automatically detected for specific product improvements?” and “How can a shitstorm be detected, so that it does not result in a loss of reputation?” arose, which can only be purposefully answered with the help of an automated analysis approach.

Based on the interviews and findings from literature, we came up with the core design requirements as shown in Table 3, which guided the upcoming design phase of our research. Hence, the requirements either affected the functionalities “data import/extraction” (DR 1 and 2), “language and user adaptation” (DR 3–5) or “data analysis” (DR 6–8). Whereas the benefits of sentiment analysis and classification of social media posts are widely acknowledged (Batrinca and Treleaven 2015; Maynard et al. 2012; Mittal and Patidar 2019), the potential of a mixed method approach (see DR 8) – as expected by our interviewees – is outlined in more depth in the next section.

To specify the mixed method analysis approach (see Table 3 – DR 8), we derived context scenarios from our interviews (Sect. 4.1) which capture the future usage situation of the tool in regards to a combination of social media analysis methods (Schilling 2016). The goal was to describe the purposeful integration of social media analysis methods in form of a “story” to uncover inconsistencies and precisely define the value of a mixed method approach for employees’ daily routines (cf. Schilling 2016). In this respect, user personas were also deduced that depict the motivation, practices and preferences of users to integrate social media analysis approaches in more depth (cf. Schilling 2016).

As a result, considering the combination of analyses that focus on structured social media data or the posts’ semantics – and thus a more sophisticated social media analysis, two context scenarios were stated to be particularly important (see also Table 3), namely “Product Commendation/Criticism” and “Topic Identification”:

First, a literature review regarding sentiment analysis was conducted to identify relevant approaches and algorithms suitable to analyze the sentiment of customer posts (cf. vom Brocke et al. 2009). An investigation of 196 relevant publications led to a total of 17 potentially suitable approaches for the sentiment analysis. Due to the characteristics of social media posts (e.g., shortness, emojis, company specific language), dictionary-based approaches represent a generally accepted approach for the automated sentiment analysis of such textual content (Feldman 2013). Depending on the sentiment of each single token (e.g., word, emoticon), an aggregated sentiment-value is calculated with the value indicating a positive (> 0), neutral (0) or negative (< 0) post (Feldman 2013).

Considering the classification of data, 130 relevant publications were examined during a second literature review, leading to nine potentially suitable approaches. To provide the capability of adapting the classification to individual or quickly changing contexts (e.g., upcoming campaigns or quickly changing trends), UR:SMART focuses on the assembly of data for predefined classes (Feldman and Sanger 2007; Heyer et al. 2006; Read et al. 2012). Therefore, a set of generally valid main categories (e.g., service, product or campaigns), independent of company or branch specifics, were worked out in cooperation with our practice partners (see Sect. 4). Additionally, to handle the individual topics and needs of each company, subcategories for each main category can be acquired. These subcategories are highly specialized and tailored to the companies’ specifics as well as the aims of their social media channels.

Furthermore, it was necessary to conceive an overall data model for the underlying database to ensure consistent data formats, the free combination of various analysis methods as well as suitable data treatment for analysis combinations. The data model consists of five classes (see Fig. 3) and each class represents a set of objects:

All classes include various attributes which describe properties of the objects (e.g., id, postid, timestamp, message) or the class (e.g., numberOfPosts, numberOfLikes) and operations the class can execute (e.g., getTotalSentisScore, getMainCategory).

Furthermore, all classes have relations to other classes. Two of them can be defined more precisely as compositions and one as a generalization or specialization, respectively. A company post is part of a social media account and each comment is clearly assigned to a company post. Comments cannot exist as separate parts detached from a specific company post and company posts cannot exist separately from a certain social media account. Each company post and each comment is classified by a main- and a subcategory. The class “subcategory” is considered a specialized form of the superclass “category”.

A further important step focused on the design of the GUI of UR:SMART. For that purpose, we used wireframes to arrive at a first shot of the GUI. The intention was to provide intuitive navigation with only a limited set of buttons and elements. One version of a wireframe that was further specified throughout the development process is shown in Fig. 4. Hence, in the upper area of the screen (see Fig. 4), a design proposal for the selection feature for the social media channels along with an indication of the desired timeframe is shown (see Table 3 – DR No. 1). Further, it was planned to show the analysis results immediately with separate graphics being used for the sentiment analysis as well as the classification of posts (see lower screen of Fig. 4) (DR No. 6 & 7). Hence, a pie chart was chosen to visualize the distribution of posts across the sentiments (e.g., positive, negative, etc.) and bar charts were considered suitable for visualizing the classification results, with the bars indicating the absolute number of posts assigned to a particular category (e.g., service, product, etc.). Such bar charts should be available for all sentiments alike. Further, a rough sketch of the user menu is shown on the left-hand side. Depending on whether the “extraction” or “results” item is selected, only the corresponding elements (selection functionality or results visualization) should be visible on the main screen. Finally, the “logo” was defined as a placeholder for an individual company’s emblem.

Generally, we used JAVA for developing UR:SMART to ensure high performance and to provide standardized libraries and interfaces. The functionality of creating graphical representations from the social media analysis was implemented as a platform-independent web application. To realize the data model, we used an H2-mysql-database, which guaranteed data compatibility through predefined and documented data interfaces. Accordingly, it is ensured that the input as well as output of an analysis are standardized and it is possible to purposefully reduce the data size, which leads to a significantly faster analysis time.

To enable the analyses, the data extraction and preprocessing functionality needed to be implemented and the relevant steps required for data preprocessing considered (cf. Aggarwal and Zhai 2012). Thereby, tokenization decomposes all textual data into smaller parts, for example single words, and removes unneeded symbols and special characters (Carstensen et al. 2009). Additionally, stop word reduction eliminates words that do not carry opinions by using publicly available stop word lists (Angulakshmi and ManickaChezian 2014). Subsequently, a stemming process eliminates prefixes and suffixes, reducing all words to their stem or basic form (Akaichi et al. 2013). Finally, a normalization algorithm completes the step of Data Preprocessing and transforms all remaining text into lower case characters (Angulakshmi and ManickaChezian 2014).

For the implementation of the sentiment analysis, we used the widely accepted implementation of a dictionary-based approach “SentiWordNet 3.0”. SentiWordNet 3.0 represents a lexical resource for an automated sentiment classification (Baccianella et al. 2010). However, SentiWordNet 3.0 only provides a lexical resource for English. To support German social media posts as well, we used SentiWS, a German language resource for analyzing the sentiment of German texts (Remus et al. 2010). As SentiWS did not match the structural requirements of the SenitwordNet 3.0 approach, we adapted SentiWS by converting the structure of the German dictionary to fit the one of SentiWordNet 3.0 (Remus et al. 2010). Both resources contain lists of words carrying a positive or negative opinion, respectively. Despite the mentioned techniques, most approaches for sentiment analysis cannot handle some special content immediately (e.g., emoticons). Therefore, feature extraction functionality considering the definition of feature types and the selection of specific features (e.g., emoticons, parts of speech, sentiment-carrying expressions) is necessary (Selvam and Abirami 2013). Due to the frequent occurrence of these features in our datasets, we integrated specific dictionaries to meet certain characteristics (e.g., dialect or emojis) of textual data. To establish a proper feature resource, we examined our dataset and extracted the most common features.

To identify irony and slang, the dictionary was extended with expressions pointing to special events (e.g., product launch) of the branches considered as well as irony patterns (cf. Trevisan et al. 2014). To receive irony patterns, we analyzed sample sets of posts of our cooperating partners, which were then added to the dictionary afterwards. A commonly encountered structure for irony patterns in the sample was a composition of an extremely negative and positive expression in a post (e.g., “The web shop is down again! Wow, this is great!!”).

Moreover, we classified the occurring emoticons as positive or negative to identify the emotions expressed within the posts. The sentiment of each word (as well as each special text component) is expressed by the variable “sentiScore”, a number within a predefined range of [-2; + 2], with a high number (near + 2) representing a very positive and a low number (towards -2) a rather negative sentiment (Feldman 2013). In this respect, posts that included words with highly positive and extremely negative sentiments at the same time were marked as potential candidates for irony. Consequently, the overall sentiment of textual data is reportable and ascribed to the categories “strong positive”, “positive”, “neutral”, “negative” and “strong negative”. The gathered data is stored in a database and can be graphically displayed in pie charts and an ECG-like representation featuring a temporal scale of sentiment progression.

To realize the classification functionality in UR:SMART, we combined multinomial naïve bayes (MNB) with a dictionary-based seed word library to identify the category of textual data. This algorithm computes the posterior probability of a class, based on the distribution of the words in the social media post. The position of the words in the text is not considered since the MNB works with a “bag of words” assumption. However, in contrast to other naïve bayes variants, the MNB takes the word frequencies into account and thus, also allows duplicates. For details on MNB we refer to Aggarwal and Zhai (2012).

This library includes specific seed words for all acquired main- and subcategories and therefore allows an assignment of posts and comments to these (Zagibalov and Carroll 2008). The main category “product”, for example, is extended by the integration of several subcategories, including company-specific product lists, parts lists as well as product accessories. Starting from the preprocessed data, all words are analyzed regarding these seed words, enabling a strong customization of the classification. Additionally, by identifying similar words surrounding existing seed words, the seed word library is constantly enhanced by company-specific expressions (Isoaho et al. 2019; Liu 2012). As a result, topics that are currently popular among customers, e.g., within a social media channel can be identified and graphically displayed. The results of the sentiment analysis and the assignment to the defined classes are then brought together by an overall view, featuring the most represented categories and subcategories within each sentiment section. Additionally, all underlying textual data are obtainable with the help of various sort and filter algorithms.

Screenshots of “UR:SMART” are shown in Fig. 5:

The research brings up several contributions for practice. First, the efforts and resources required for performing social media analysis that is introduced by more and more companies in the course of their digitalization efforts can be significantly reduced by our solution through an automated and target-oriented analysis. Thereby, specific issues came up (e.g., “Product Commendation/Criticism” and “Topic Identification”) for which a combined mixed method approach truly contributes to a better understanding of the social media data, a topic many companies are still struggling with these days (cf. Ried 2019). In this respect, our approach facilitates the identification of the “voice of the customer” (Pande et al. 2014) based on social media data, because access to customers’ current moods and expectations is enabled. Based on the findings, process improvement efforts may be triggered, or countermeasures taken to avoid customer dissatisfaction among others.

Second, applying a mixed method approach generates beneficial insights that are superior to those findings that will be received from performing a singular analysis approach that either prioritizes structured or unstructured data (e.g., Chen et al. 2014; Tinati et al. 2014). As an example, by the target-oriented combination of analysis approaches, it is not only possible to cluster user suggestions on how to improve the product or service portfolio by topic and sentiment, but also to easily deduce a ranking of these propositions, e.g., based on “likes” and “shares” immediately. That way, employees receive a more profound and individual foundation for decision-making based on the information captured in social media posts.

Third, customers’ reactions to social media marketing or advertising campaigns become evident (cf. Castronovo and Huang 2012; Shareef et al. 2019). This is highly beneficial for companies, because knowledge about the influence of social media campaigns on the success of introducing new products is still limited (Baum et al. 2019). Based on the topics captured in customers’ reactions to corresponding advertisement postings – which are recognized by our tool – practitioners directly receive feedback from customers. This helps to better assess the type of information about a new product or service (e.g., availability, functionality, price, etc.), which is truly perceived as relevant by the target group. Further, companies may learn how to present this information via social media channels in an appealing way that is really appreciated by consumers, raises their attention and elicits a reaction. All this will help to design future marketing campaigns more purposefully.

Fourth, the SUMI usability study brought up encouraging results concerning the design of our tool. Hence the “global scale” rating indicated a high “perceived quality of use” from the side of the users. Thereby, the tool was judged to be easy to handle, self-explanatory, visually appealing and helpful to analyse social media posts (see Sect. 7). The proposed design may thus serve as a reference for constructing social media analysis tools. Hence, the arrangement of the elements and menu fields of the GUI, enabling easy extraction of social media posts as well as the graphical presentation of the results in the form of pie charts, histograms and time diagrams put users in the position to easily navigate the tool and derive insights from the data analysis. Additional findings are indicated by the key figures as provided by the quantitative analysis. Furthermore, the functionality captured by the design along with the data model provided those features by which users felt ideally supported when working on a social media analysis case study in our experiment. However, the SUMI study also led to suggestions for further development of the tool. Hence, it was proposed to add, for instance, export functionality for the result presentations to common MS Office packages. Moreover, suggestions to further decompose certain subcategories came up. Finally, users wished for pop-up windows with instructions for use to further improve the handling. This feedback will be considered in future steps to improve “UR:SMART”.

In addition to practice, our approach also proposes several benefits for academia. First, the combination of data analysis approaches as well as the integration of methods in general has proven useful for various research disciplines (Stieglitz et al. 2014). This concerns the development of scientific procedures such as triangulation and inter-method mixing (cf. Jick 1979; Johnson and Turner 2003) but also topics such as for instance Method Engineering and process-oriented quality management among others (cf. Brinkkemper 1996; Johannsen 2011; Ralyté et al. 2003; Tolvanen et al. 1996). In this paper, we combine different data analysis methods for social media (e.g., Stieglitz et al. 2014). Concretely, we propose to integrate different perspectives on social media data. We could show that this allows the retrieval of additional information from the data (e.g., the reasons for customers’ negative attitude towards certain offerings) complementing those findings gained by a pure application of a certain analysis approach in isolation. This fosters a company’s “understanding” of the massive amount of data extracted from social media channels, which is a topic firms do have to actively approach according to a current study by CapGemini (2020). Generally, companies strive to harness the large quantity of data to strengthen customer relationships these days (Wedel and Kannan 2016) and therefore increasingly hire data scientists to build up competences in data-driven decision-making (Provost and Fawcett 2013). It is not fully clear yet in academic literature, however, “which types of analytics work for which types of problems and data” and “what new methods are needed for analyzing new types of data” (Wedel and Kannan 2016, p. 97). However, these issues are of vital importance for companies to prevail on the market (Wedel and Kannan 2016). In this respect, social networks such as Facebook and Instagram largely contribute to the fast growth of data volumes within companies (Felt 2016) as they play a decisive role in content marketing and product advertisement (Baltes 2015; Statista 2018b) but also for customer service purposes (Statista 2017). At that point, we contribute to one of the central questions of social media analysis research, namely how data analysis methods can be combined to purposefully derive insights from social media data (cf. Stieglitz et al. 2014, 2018).

Second, the automated analysis of social media posts still holds various challenges when it comes to its practical application (Stieglitz et al. 2018). In our research, we built on a dictionary-based approach for the sentiment analysis. While freely available lexical resources exist for that purpose (e.g., SentiWordNet), these are not customized for particular industries or company types. Hence, to raise the accuracy level of our approach, the lexical resource (SentiWordNet 3.0) had to be adapted to match the particular needs of our collaborating partners, who mainly came from southern Germany. So, researchers are called to enhance freely available dictionaries by branch-, regional- and company-specific language peculiarities to help raise the accuracy levels of such resources that can be used by researchers straight away.

Third, while most companies are nowadays able to process and store a large amount of data (e.g., by help of Apache Hadoop), they still do not know how to extract valuable insights from the data sets (CapGemini 2019; Ried 2019). This can also be seen as one reason for the lack of success of digital transformation initiatives (CapGemini 2019). Based on a series of interviews with eleven cooperation partners, we deduced highly relevant application scenarios for analyzing social media content that help companies to purposefully trigger improvement efforts based on the analyses for instance. While the technical side of social media analysis is a livelily discussed topic (Feldman and Sanger 2007; Liu 2012; Medhat et al. 2014), the question of how to use and combine the existing analysis approaches – to truly derive beneficial insights for firms – still lacks a theoretical foundation (cf. Ried 2019; Stieglitz et al. 2014, 2018). Hence, the research brings up valuable indications for promising application scenarios in social media analysis.

Fourth, our tool represents a helpful instrument to identify the “voice of the customer” (Pande et al. 2014) based on freely available social media data. This is particularly helpful for business process management (BPM) and quality management (QM) initiatives, as their success largely depends on the proper analysis of consumers’ expectations and needs (cf. Pande et al. 2014). Accordingly, process weaknesses can be eliminated, and the process design aligned with customer requirements more purposefully. While BPM research often focuses on the question of how to utilize employees’ process knowledge to mitigate process weaknesses (cf. Seethamraju and Marjanovic 2009) by means of systematic procedure models (e.g., Adesola and Baines 2005; Coskun et al. 2008; Harrington 1991; Zellner 2011), of how to use und develop “patterns” to support the “act of improvement” (Bergener et al. 2015; Forster 2006; Höhenberger and Delfmann 2015; Lang et al. 2015) or of how to apply process mining techniques to acknowledge deviations between an as-is and a should-be process (van der Aalst 2012), future work may focus on the integration of BPM approaches and social media analysis in more depth.

In view of the ongoing debate in DS about the need for a theoretical contribution to research that goes beyond the technical contribution (i.e., the artifact) as outcomes of a DS research project (cf. Baskerville et al. 2018) we can summarize the following contributions to theories: