1.1 The Role of Computational Analysis in the Social Sciences
The use of computers is nothing new in the social sciences. In fact, one could argue that some disciplines within the social sciences have even been early adopters of computational approaches. Take the gathering and analyzing of large-scale survey data, dating back to the use of the Hollerith Machine in the 1890 US census. Long before every scholar had a personal computer on their desk, social scientists were using punch cards and mainframe computers to deal with such data. If we think of the analysis of communication more specifically, we already see attempts to automate content analysis in the 1960ās (see, e.g. Scharkow, 2017).
However, something has profoundly changed in recent decades. The amount and type of data we can collect as well as the computational power we have access to have increased dramatically. In particular, digital traces that we leave when communicating online, from access logs to comments we place, have required new approaches (e.g., Trilling, 2017). At the same time, better computational facilities now allow us to ask questions we could not answer before.
GonzĆ”lez-BailĆ³n (2017), for instance, argued that the computational analysis of communication now allows us to test theories that were formulated a century ago, such as Tardeās theory of social imitation. Salganik (2019) tells an impressive methodological story of continuity in showing how new digital research methods build on and relate to established methods such as surveys and experiments, while offering new possibilities by observing behavior in new ways.
A frequent misunderstanding, then, about computational approaches is that they would somehow be a-theoretical. This is probably fueled by clichĆ©s coined during the āBig Dataā-hype in the 2010ās, such as the infamous saying that in the age of Big Data, correlation is enough (Mayer-Schƶnberger and Cukier, 2013); but one could not be more wrong: as the work of Kitchin (2014a, b) shows, computational approaches can be well situated within existing epistemologies. For the field to advance, computational and theoretical work should be symbiotic, with each informing the other and with neither claiming superiority Margolin, 2019). Thus, the computational scientistsā toolbox includes both more data-driven and more theory-driven techniques; some are more bottom-up and inductive, others are more top-down and deductive. What matters here, and what is often overlooked, is in which stage of the research process they are employed. In other words, both inductive and deductive approaches as they are distinguished in more traditional social-science textbooks (e.g., Bryman, 2012) have their equivalent in the computational social sciences.
Therefore, we suggest that the data collection and data analysis process is thought of as a pipeline. To test, for instance, a theoretically grounded hypothesis about personalization in the news, we could imagine a pipeline that starts with scraping online news, proceeds with some natural-language processing techniques such as Named Entity Recognition, and finally tests whether the mentioning of persons has an influence on the placement of the stories. We can distinguish here between parts of the pipeline that are just necessary but not inherently interesting to us, and parts of the pipeline that answer a genuinely interesting question. In this example, the inner workings of the Named Entity Recognition step are not genuinely interesting for us ā we just need to do it to answer our question. We do care about how well it works and especially which biases it may have that could affect our substantive outcomes, but we are not really evaluating any theory on Named Entity Recognition here. We are, however, answering a theoretically interesting question when we look at the pipeline as a whole, that is, when we apply the tools in order to tackle a social scientific problem. Of course, what is genuinely interesting depends on oneās discipline: For a computational linguist, the inner workings of the named entity recognition may actually be the interesting part, and our research question just one possible ādownstream taskā.
This distinction is also sometimes referred to as ābuilding a better mousetrapā versus āunderstandingā. For instance, Breiman (2001) remarked: āMy attitude toward new and/or complicated methods is pragmatic. Prove that youāve got a better mousetrap and Iāll buy it. But the proof had better be concrete and convincing.ā (p. 230). In contrast, many social scientists are using statistical models to test theories and to understand social processes: they want to specifically understand how x relates to y, even if y may be better predicted by another (theoretically uninteresting) variable.
This book is to some extent about both building mousetraps and understanding. When you are building a supervised machine learning classifier to determine the topic of each text in a large collection of news articles or parliamentary speeches, you are building a (better) mousetrap. But as a social scientist, your work does not stop there. You need to use the mousetrap to answer some theoretically interesting question.
Actually, we expect that the contents of this book will provide a background that helps you to face the current research challenges in both academia and industry. On the one hand, the emerging field of Computational Social Science has become one of the most promising areas of knowledge and many universities and research institutes are looking for scholars with this profile. On the other hand, it is widely known that nowadays the computational skills will increase your job opportunities in private companies, public organizations or NGOs, given the growing interest in data-driven solutions.
When planning this book, we needed to make a couple of tough choices. We aimed to at least give an introduction to all techniques that students and scholars who want to computationally analyze communication will probably be confronted with. Of course, specific ā technical ā literature on techniques such as, for instance, machine learning can cover the subject in more depth, and the interested student may indeed want to dive into one or several of the techniques we cover more deeply. Our goal here is to offer enough working knowledge to apply these techniques and to know what to look for. While trying to cover the breadth of the field without sacrificing too much depth when covering each technique, we still needed to draw some boundaries. One technique that some readers may miss is agent-based modeling (ABM). Arguably, such simulation techniques are an important technique in the computational social sciences more broadly (Cioffi-Revilla, 2014), and they have recently been applied to the analysis of communication as well (Waldherr, 2014, Wettstein, 2020). Nevertheless, when reviewing the curricula of current courses teaching the computational analysis of communication, we found that simulation approaches do not seem to be at the core of such analyses (yet). Instead, when looking at the use of computational techniques in fields such as journalism studies (e.g., Boumans and Trilling, 2016), media studies (e.g., Rieder, 2017), or the text-as-data movement (Grimmer and Stewart, 2013), we see a core of techniques that are used over and over again, and that we have therefore included in our book. In particular, besides general data analysis and visualization techniques, these are techniques for gathering data such as web scraping or the use of APIās; techniques for dealing with text such as natural language processing and different ways to turn text into numbers; supervised and unsupervised machine learning techniques; and network analysis.
