1

Automated Text Analysis for Consumer Research

ASHLEE HUMPHREYS

REBECCA JEN-HUI WANG

Forthcoming, Journal of Consumer Research

© The Author 2017. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights
reserved. For permissions, please e-mail: [email protected]
 2

Ashlee Humphreys, Associate Professor, Integrated Marketing Communications, Medill

School of Journalism, Media, and Integrated Marketing Communications, Northwestern University.

Correspondence concerning this article should be addressed to Ashlee Humphreys, IMC, Medill

School of Journalism, Northwestern University, MTC 3-109, 1870 Campus Drive, Evanston, IL 60208.

Electronic mail may be sent to [email protected]. Rebecca Jen-Hui Wang is an assistant

professor at Lehigh University. She can be reached at [email protected]. The authors would like to

thank David Dubois, Alistair Gill, Jonathan Berman, Ann Kronrod, Joseph T Yun, Jonah Berger, and

Kent Grayson for their feedback and encouragement on the manuscript, and Andrew Wang for his help

with data collection for the web appendix. Supplementary materials are included in the web

appendix accompanying the online version of this article.

Eileen Fischer served as editor and Linda Price served as associate editor for this article.
 3

ABSTRACT

The amount of digital text available for analysis by consumer researchers has risen

dramatically. Consumer discussions on the internet, product reviews, and digital archives of news

articles and press releases are just a few potential sources for insights about consumer attitudes,

interaction, and culture. Drawing from linguistic theory and methods, this article presents an overview

of automated text analysis, providing integration of linguistic theory with constructs commonly used in

consumer research, guidance for choosing amongst methods, and advice for resolving sampling and

statistical issues unique to text analysis. We argue that although automated text analysis cannot be used

to study all phenomena, it is a useful tool for examining patterns in text that neither researchers nor

consumers can detect unaided. Text analysis can be used to examine psychological and sociological

constructs in consumer-produced digital text by enabling discovery or by providing ecological validity.

Keywords: automated text analysis; computer-assisted text analysis; automated content

analysis; computational linguistics

 4

Over the last two decades, researchers have seen an explosion of text data generated by

consumers in the form of text messages, reviews, tweets, emails, posts and blogs. Some part of this

rise is attributed to an increase in sites like Amazon.com, CNET.com, and thousands of other product

websites that offer forums for consumer comment. Another part of this growth comes from consumer-

generated content including discussions of products, hobbies, or brands on feeds, message boards, and

social networking sites. Researchers, consumers, and marketers swim in a sea of language, and

more and more of that language is recorded in the form of text. Yet within all of this information

lies knowledge about consumer decision-making, psychology, and culture that may be useful to

scholars in consumer research. Blogs can be used to study opinion leadership; message boards can tell

us about the development of consumer communities; feeds like Twitter can help us unpack social

media firestorms; and social commerce sites like Amazon can be mined for details about word-of-

mouth communication.

Correspondingly, ways of doing social science are also changing. Because data has

become more readily available and the tools and resources for analysis are cheaper and more

accessible, researchers in the material sciences, humanities, and social sciences are developing

new methods of data-driven discovery to deal with what some call the “data deluge” or “big

data” (Bell, Hey, and Szalay 2009; Borgman 2015). Just as methods for creating, circulating, and

storing online discussion have grown more sophisticated, so too have tools for analyzing

language, aggregating insight, and distilling knowledge from this overwhelming amount of data.

Yet despite the potential importance of this shift, consumer research is only beginning to incorporate

methods for collecting and systematically measuring textual data to support theoretical propositions

and make discoveries.

In light of the recent influx of available data and the lack of an overarching framework for

doing consumer research using text, the goal of this article is to provide a guide for research designs
 5

that incorporate text and to help researchers assess when and why text analysis is useful for answering

consumer research questions. We provide an overview of both deductive top-down, dictionary-based

approaches as well as inductive and abductive bottom-up approaches such as supervised and

unsupervised learning to incorporate discovery-oriented as well as theoretically-guided methods. These

designs help make discoveries and expand theory by allowing computers to detect and display patterns

that humans cannot and by providing new ways of “seeing” data through aggregation, comparison, and

correlation. We further offer guidance for choosing amongst different methods and address common

issues unique to text analysis such as sampling internet data, developing wordlists to represent a

construct, and analyzing sparse, non-normally distributed data. We also address validity, reliability,

generalizability, and ethical issues for research using textual data.

Although there are many ways to incorporate automated text analysis into consumer research,

there is not much agreement on the standard set of methods, reporting procedures, steps of data

inclusion, exclusion, and sampling, and, where applicable, dictionary development and validation. Nor

has there been an integration of the linguistic theory on which these methods are based into consumer

research, which can enlighten us to the multiple dimensions of language that can be used to measure

consumer thought, interaction, and culture. While fields like psychology provide some guidance for

dictionary-based methods (e.g. Tausczik and Pennebaker 2010) and for analysis of certain types of

social media data (Kern et al. 2016), they don’t provide grounding in linguistics, cover the breadth of

methods available for studying text, or provide criteria for deciding amongst approaches. In short,

most of the existing literature examines only a handful of aspects of discourse that pertain to the

research questions of interest, does not address why one method is chosen over others, and does

not discuss the unique methodological issues consumer researchers face when dealing with text.

This paper therefore offers three contributions to consumer research. First, we detail how

linguistic theory can inform theoretical areas common in consumer research such as attention,
 6

processing, interpersonal interaction, group dynamics, and cultural characteristics. Second, we outline a

practical roadmap for researchers who want to use textual data, particularly unstructured text obtained

from real-world settings such as tweets, newspaper articles, or online reviews. Lastly, we examine what

can and cannot be done with text analysis and provide guidance for validating results and interpreting

findings in non-experimental contexts.

The rest of the paper is organized around the roadmap in Figure 1. This chart presents a

series of decisions a researcher faces when analyzing text. We outline six stages: 1) developing a

research question, 2) identifying the constructs, 3) collecting data, 4) operationalizing the constructs,

5) interpreting the results, and 6) validating the results. Although text analysis need not necessarily

unfold in this order (for instance, construct definition will sometimes occur after data collection),

researchers have generally followed this progression (see e.g. Lee and Bradlow 2011).

Roadmap for Automated Text Analysis

Methods of automated text analysis come from the field of computational linguistics

(Kranz 1970; Stone 1966). The relationship between computational linguistics and text analysis

is analogous to that of biology to medicine or of physics to engineering. That is, computational

linguistics, at its core, emphasizes advancing linguistics theory and often focuses on the accuracy

of prediction as an end in itself (Hausser 1999, p.8). Computer-assisted or automated text

analysis, on the other hand, refers to a set of techniques that use computing power to answer

questions related to psychology (e.g. Chung and Pennebaker 2013; Tausczik and Pennebaker

2010), political science (e.g. Grimmer and Stewart 2013), sociology (Mohr 1998; Shor et al.

2015), and other social sciences (Carley 1997; Weber 2005). In these fields, language represents

some focal construct of interest, and computers are used to measure those constructs, provide

systematic comparisons and sometimes find patterns that neither human researchers nor subjects
 7

of the research can detect. In other words, while computational linguistics is a field that is primarily

concerned with language in the text, for consumer researchers, text analysis is merely a lens through

which to view consumer thought, behavior, and culture. Analyzing texts, in many contexts, is not the

ultimate goal of consumer researchers, but is instead a precursor for testing the relationship between

or amongst the constructs or variables of interest.

As such, we use the term automated text analysis or computer-assisted text analysis over

computational linguistics (Brier and Hopp 2011). Although we follow convention by using the

term “automated,” this should not imply that human intervention is absent. In fact, many of the

computer-enabled tasks such as dictionary construction, validation, and cluster labeling are

iterative processes that require human design, modification, and interpretation. Some prefer the

term computer-assisted text analysis (Alexa 1997) to explicitly encompass a broad set of

methods that take advantage of computation in varying amounts ranging from a completely

automated process using machine learning to researcher-guided approaches that include manual

coding and wordlist development. In the following sections, we discuss the design and execution of

automated text analysis in detail, beginning with selection of a research question and connecting

linguistic aspects to important constructs in consumer research.

----Insert Figure 1 about here----

Stage 1: Developing a Research Question

As with any research, the first step is developing a research question. To understand the

implementation of automated text analysis, one should start by first considering if the research question

lends itself to text analysis. Contemplating whether text analysis is suitable for the research context

is perhaps the most important decision to consider, and there are at least three purposes for which text

analysis would be inappropriate.

 8

First, much real-world textual content is observational data that occurs without the controlled

conditions of an experiment or even a field test. Depending on the context and research question,

automated text analysis alone would not be the best method for inferring causation when studying a

psychological mechanism.1 If the researcher needs precise control to compare groups, introduce

manipulations, or rule out alternative hypotheses through random assignment protocols (Cook,

Campbell, and Day 1979), textual analysis would be of limited use.2

Secondly, if the research question concerns data at the behavioral or unarticulated level (e.g.

response time, skin conductance, consumer practices, etc.), text analysis would not be appropriate.

Neural mechanisms that govern perception or attention, for example, would be ill-suited for the

method. Equally, if one needs a behavioral dependent variable, text analysis would not be appropriate

to measure it. For example, when studying self-regulation, it is clearly important to include behavioral

measures to examine not just behavioral intention—what people say they will do—but action itself.

This restriction applies to sociologically-oriented research as well. For example, with practice theory

(Allen 2002; Schatzki 1996) or ethnography (Belk, Sherry, and Wallendorf 1988; Schouten and

McAlexander 1995), observation of consumer practices is vital because consumer behavior may

diverge markedly from discourse (Wallendorf and Arnould 1989). Studying text is simply no substitute

for studying behavior. Not all constructs lend themselves to examination through text, and these

constructs tend to be behaviorally-oriented constructs.

Lastly, there are many contexts in which some form of text analysis would be valuable, but

automated text analysis would be insufficient. Identifying finer shades of meaning such as sarcasm and

differentiating amongst complex concepts, rhetorical strategies, or complex arguments are often not

possible via automated processes. Additionally, studies that employ text analysis often sample data

1
For alternative perspectives of studying causation with historical case data and macro-level data, see Mahoney and
Rueschemeyer (2003) and Jepperson and Meyer (2011).
2
However, text analysis has been used to code thought protocols in experimental settings (e.g. Hsu et al. 2014).
 9

from public discourse in the form of tweets, message boards, or posts, and there is a wide range of

expression that consumers may not pursue in these media because of stigma or social desirability.

There is a rich tradition of text analysis in consumer research such as discourse analysis (Holt and

Thompson 2004; Thompson and Hirschman 1995), hermeneutic analysis (Arnold and Fischer 1994;

Thompson, Locander, and Pollio 1989), and human content analysis (Kassarjian 1977) for uncovering

rich, deep, and sometimes personal meaning of consumer life in the context in which it is lived.

Although automated text analysis could be a companion to these methods, it cannot be a standalone

approach for understanding this kind of richer, deeper, and culturally-laden meaning.

So, when is automated text analysis appropriate? In general, it is good for analyzing data in a

context where humans may be limited or partial. Computers can sometimes see patterns in language

that humans cannot detect, and they are impartial in the sense that they measure textual data evenly and

precisely over time or in comparisons between groups without preconception. Further, by quantifying

constructs in text, computers provide new ways of aggregating and displaying information to uncover

patterns that may not be obvious at the granular level. There are at least four types of problems where

these advantages can be leveraged.

First, automated text analysis can lead to discoveries of systematic relationships in text

and hence amongst constructs that may be overlooked by researchers or consumers themselves.

Patterns in correlation, notable absences, and relationships amongst three or more textual

elements are all things that are simply hard for a human reader to see. For example, in medical

research, Swanson (1988) finds a previously unrecognized relationship between migraine

headaches and magnesium levels through the text analysis of other, seemingly unrelated

research. Automated text analysis may also provide alternative ways of “reading” the text to

make new discoveries (Kirschenbaum 2007). For instance, Jurafsky et al. (2014) find expected

patterns in negative restaurant reviews such as negative emotion words, but they also discover words
 10

like “after”, “would”, and “should” in these reviews, which are used to construct narratives of

interpersonal trauma primarily based on norm-violations. Positive restaurant reviews, on the other

hand, contain stories of addiction rather than simple positive descriptions of food or service. These

discoveries, then, theoretically inform researchers’ understanding of negative and positive sentiment,

particularly with consumer experiences.

Using text analysis, researchers have also discovered important differences between expert and

consumer discourse when evaluating products (Lee and Bradlow 2011; Netzer et al 2012). In the case

of cameras, for example, systematic linguistic comparison of expert reviews to consumer reviews

reveals that there is a significant disconnect between what each of these groups consider important. For

example, in their reviews, consumers value observable attributes like camera size and design, while

experts stress less visible issues like flash range and image compression (Lee and Bradlow 2011). In

the case of prescription drugs, the differences between consumers and experts take on heightened

meaning, as textual comparison of patient feedback on drugs to WebMD shows that consumers report

side-effects missing from the official medical literature (Netzer et al. 2012). In this way, text analysis

can reveal discoveries that would be hard to detect on a more granular level, and the scope and

systematicity of the analysis can grant more validity and perhaps power to consumers’ point of view.

Secondly, researchers can use computers to execute rules impartially in order to measure

changes in language over time, compare between groups, or aggregate large amounts of text.

These tasks are more than mere improvements in efficiency in that they present an alternative

way of “seeing” the text through conceptual maps (Martin, Pfeffer, and Carley 2013), timelines

(Humphreys and Latour 2013), or networks (Arvidsson and Caliandro 2016), and provide

information about rate and decay. For example, using features like geolocation and time stamps

along with textual data from Twitter, Snefjella and Kuperman (2015) develop new knowledge about
 11

construal level such as its rate of change given a speaker’s physical, temporal, social, or topical

proximity.

By providing an explicit rule set and having a computer execute the rules over the entire

dataset, researchers reduce the possibility that their texts will be analyzed unevenly or

incompletely. When making statistical inferences about changes in concepts over time, this is

especially important because the researcher needs to ensure that measurement is consistent

throughout the dataset. For example, by aggregating and placing counts of hashtags on a

timeline, Arvidsson and Caliandro (2016) demonstrate how networks of concepts used to discuss

Louis Vuitton handbags peak at particular times and in accordance with external events,

highlighting attention for a particular public. If a researcher wants to study a concept like brand

meaning, text analysis can help to create conceptual or positioning maps that represent an

aggregated picture of consumer perceptions that can then be used to highlight potential gaps or

tensions in meaning amongst different constituencies or even for one individual (Lee and

Bradlow 2011; Netzer et al. 2012).

Thirdly, text analysis can be a valuable companion to experimental research designs by

adding ecological validity to lab results. For example, Mogilner et al. (2011) find robust support for

changes in the frame of happiness that correspond with age by looking at a large dataset of personal

blogs, patterns they also find in a survey and laboratory experiment. In a study of when and why

consumers explain choices as a matter of taste versus quality, Spiller and Belogolova (2016) use text

analysis first to code a dependent variable in experimental analysis, but then add robustness to their

results by demonstrating the effect in the context of online movie reviews. In this way, text analysis is

valuable beyond its more traditional uses for coding thought protocols, but also useful for finding and

measuring psychological and sociological constructs in naturally-occurring consumer discourse.

 12

Lastly, there are some relationships for which observational data is the most natural way

to study the phenomenon. Interpersonal relationships and group interaction can be hard to study

in the lab, but they can be examined through text analysis of online interaction or transcripts of

recorded conversation (e.g. Jurafsky, Ranganath, and McFarland 2009). For example, Barasch

and Berger (2014) combine laboratory studies with dictionary-based text analysis of consumer

discussions to show that consumers share different information depending on the size of their audience.

Given these considerations, once deciding that text analysis is appropriate for some part

of the research design, the next question is what role it will play. Text could be used to represent

the independent variable (IV), dependent variable (DV) or both. For example, Tirunillai and

Tellis (2012) operationalize “chatter” using quantity and valence of product reviews to represent

the IV, which predicts firm performance in the financial stock market. Conversely, Hsu et al.

(2014) experimentally manipulate distraction, the IV, and measure thoughts as the DV using text

analysis. Other studies use text as both the IV and the DV. For example, Humphreys (2010)

examines how terms related to casino gambling and entertainment in newspaper articles

converged over time along with a network of other concepts such as luxury and money, while

references to illegitimate frames like crime fell. As these cases illustrate, text analysis is a

distinct component of the research design, to be executed and then incorporated into the overall

design.

More generally, text analysis can occupy different places in the scientific process,

depending on the interests and orientation of the researchers. It is compatible with both theory-

testing and discovery-oriented designs. For some, text analysis is a way of first discovering

patterns that are later verified using laboratory experiments (Barasch and Berger 2014; Berger

and Milkman 2012; Packard and Berger 2016). Others use text analysis to enrich findings after

investigating a psychological or social mechanism (Mogilner et al. 2011; Spiller and Belogolova
 13

2016). In the same way, sociological work has used text analysis to illustrate findings after an

initial discovery phase through qualitative analysis (Arsel and Bean 2013; Humphreys 2010) or

to set the stage by presenting socio-cultural discourses prior to individual or group-level analysis

(Arsel and Thompson 2011).

Stage 2: Construct Identification

After deciding that text analysis might be appropriate for the research question, the next step

is to identify the construct. Doing so, however, entails recognizing that text is ultimately based on

language. To build sound hypotheses and make valid conclusions from text, one must first

understand the underpinnings of language.

Language indelibly shapes how humans view the world (Piaget 1959; Quine 1970; Vico

1725/1984; Whorf 1944). It can be both representative of thought and instrumental in shaping

thought (Kay and Kempton 1984; Lucy and Shweder 1979; Sapir 1929; Schmitt and Zhang 1998;

see also Graham 1981; Whorf 1944). For example, studies have shown that languages with

gendered nouns like Spanish and French are more likely to make speakers think of physical

objects as having a gender (Boroditsky, Schmidt, and Phillips 2003; Sera, Berge, and del Castillo

Pintado 1994). Languages like Mandarin that speak of time vertically rather than horizontally

shape native speakers’ perceptions of time (Boroditsky 2001), and languages like Korean that

emphasize social hierarchy reflect this value in the culture (McBrian 1978). These effects

underscore the fact that by studying language, consumer researchers are studying thought and

that language is conversely important because it shapes thought.

As a sign system, language has three aspects—semantic, pragmatic, and syntactic (Mick

1986; Morris 1994)—and each aspect of language provides a unique window into a slightly different

part of consumer thought, interaction, or culture. Semantics concerns word meaning that is explicit in
 14

linguistic content (Frege 1892/1948) while pragmatics addresses the interaction between linguistic

content and extra-linguistic factors like context or the relationship between speaker and hearer (Grice

1970). Syntax focuses on grammar, the order in which linguistic elements are presented (Chomsky

1957/2002). By understanding and appreciating these linguistic underpinnings, researchers can

develop sounder operationalizations of the constructs and more insightful, novel hypotheses. We will

discuss semantics, pragmatics, and syntax in turn as they are relevant to constructs in consumer

research. Extensive treatment of these properties can be found in Mick (1986), although previous use

of semiotics in consumer research has been focused primarily on objects and images as signs

(Grayson and Shulman 2000; McQuarrie and Mick 1996; Mick 1986; Sherry, McGrath, and Levy

1993) rather than on language itself. In discussing construct identification, we link linguistic theory

to topics of interest in consumer research—attention, processing, social influence, and group

properties.

To more fully understand what kinds of problems might be fruitfully studied through text

analysis, we detail four theoretical areas of consumer research that link with linguistic

dimensions of semantics, pragmatics, and syntax. Specifically, attention can be examined

through semantics, processing through syntax, interpersonal dynamics through pragmatics, and

group level characteristics through semantics and higher order combinations of these dimensions.

Attention

The first area where text analysis is potentially valuable to consumer research is in the

study of attention. Consumer attention is important in the evaluation of products and

experiences, self-awareness, attitude formation, and attribution, to name only a few domains.

Language represents attention in two ways. When consumers are thinking of or attending to an

issue, they tend to express it in words. Conversely, when consumers are exposed to a word, they
 15

are more likely to attend to it. In this way, researchers can measure what concepts constitute

attention in a given context, study how attention changes over time, and evaluate how concepts

are related to others in a semantic network. Through semantics, researchers can measure

temporal, spatial, and self-focus, and in contrast to self-reports, text analysis can reveal patterns

of attention or focus of which the speaker may not be conscious (Mehl 2006).

Semantics, the study of word meaning, links language with attention. From the

perspective of semantics, a word carries meaning over multiple, different contexts, and humans

store that information in memory. Word frequency, measuring how frequently a word occurs in text,

is one way of measuring attention and then further mapping a semantic network. For example, based

on the idea that people discuss the attributes that are top-of-mind when thinking of a particular

car, Netzer et al. (2012) produce a positioning map of car attributes from internet message board

data using supervised learning.

Researchers can infer the meaning of the word, what linguists and philosophers call the sense

(Frege 1892/1948), through its repeated and systematic co-occurrence with a system of other words

based on the linguistic principle of holism (Quine 1970). For example, if the word “Honda” is

continually and repeatedly associated with “safety,” one can infer that these concepts are related in

consumers’ minds such that Honda means safety to a significant number of consumers. In this way, one

can determine the sense through the context of words around it (Frege 1892/1948; Quine 1970), and

this holism is a critical property from a methodological perspective because it implies that the meaning

of a word can be derived by studying its collocation with surrounding words (Neuman, Turney, and

Cohen 2012; Pollach 2012). Due to the inherent holism of language, semantic analysis is a natural

fit with spreading activation models of memory and association (Collins and Loftus 1975).

Text analysis can also measure implicit rather than explicit attention through semantics.

The focus of consumer attention on the self as opposed to others (Spiller and Belogolova 2016)
 16

and temporal focus such as psychological distance and construal (Snefjella and Kuperman 2015)

are patterns that may not be recognized by consumers themselves, but can be made manifest

through text analysis (Mehl 2006). For example, a well-known manipulation of self-construal is

the “I” versus “we” sentence completion task (Gardner, Gabriel, and Lee 1999). Conversely, text

analysis can help detect differences in self-construal using measures for these words.

Language represents the focus of consumer attention, but it can also direct consumer

attention through semantic framing (Lakoff 2014; Lakoff and Ferguson 2015). For example,

when Oil of Olay claims to “reverse the signs of aging” in the United States, but the same

product claims to “reduce the signs of aging” in France, the frame activates different meaning

systems, “reversing” being more associated with agency, and “reduction” being a more passive

framing. As ample research in framing and memory has shown, consumers’ associative networks

can be activated when they see a particular word, which in turn may affect attitudes (Humphreys

and Latour 2013; Lee and Labroo 2004; Valentino 1999), goal pursuit (Chartrand and Bargh

1996; Chartrand et al. 2008) and regulatory focus (Labroo and Lee 2006; Lee and Aaker 2004).

Language represents not only the cognitive components of attention, but also reflects the

emotion consumers may feel in a particular context. Researchers have used automated text analysis to

study the role of emotional language in the spread of viral content (Berger and Milkman 2012),

response to national tragedies (Doré et al. 2015), and well-being (Settanni and Marengo 2015). As we

will later discuss, researchers use a broad range of sentiment dictionaries to measure emotion and

evaluate how consumer attitudes may change over time (Hopkins and King 2010), in certain

contexts (Doré et al. 2015), or due to certain interpersonal groupings. Building on these

approaches, researchers studying narrative have used the flow of emotional language (e.g. from

more to less emotion words) to code different story arcs such as comedy (positive to negative to

positive) versus tragedy (negative to positive to negative) (Van Laer et al. 2017).
 17

Processing

The structure of language, or syntax, can provide evidence of different kinds of

processing for senders and can prompt different kinds of responses from readers. Syntax refers to

the structure of phrases and sentences in text (Morris 1938). In any language, there are many

ways to say something without loss of meaning, and these differences in grammatical

construction can indicate differences in footing (Goffman 1979), complexity (Gibson 1998), or

assertiveness (Kronrod, Grinstein, and Wathieu 2012). For example, saying “I bought the soap”

rather than “The soap was bought” has different implications for attribution of agency, which

could have consequences for satisfaction and attribution in the case of product success or failure.

Passive versus active voice is one key difference that can be measured through syntax,

indicated by word order or by use of certain phrases or verbs. Active versus passive voice, for

instance, affects persuasiveness of the message. Specifically, consumer-to-consumer word-of-

mouth that is expressed in a passive voice may be more persuasive than active voice, particularly

when language contains negative sentiment or requires extensive cognitive processing (Bradley

and Meeds 2002; Carpenter and Henningsen 2011; see also Kronrod et al. 2012). When speakers

use passive sentences, they shift the attention from the self to the task or event at hand (Senay,

Usak, and Prokop 2015), and passive voice may therefore further signify lower power or a desire

to elude responsibility. For instance, literature in accounting suggests that companies tend to

report poor financial performance with passive voice (e.g., Clatworthy and Jones 2006).

Syntactic complexity (Gibson 1998; Wong, Ormiston, and Haselhuhn 2011) can

influence the ease of processing. Exclusion words like “but” and “without” and conjunctions like

“and” and “with” are used in more complex reasoning processes, and the frequency of these

words can therefore be used to represent the depth of processing in consumer explanations,
 18

reviews, or thought listings. Sentence structures also lead to differences in recall and

memorability. For example, Danescu-Niculescu-Mizil et al. (2012a) examine the syntactic

characteristics using quotations from the movie IMDB website, finding that memorable

quotations tend to have less common word sequence but common syntax.

Similarly, exclusions (without, but, or) are used to make distinctions (Tausczik and

Pennebaker 2010), while conjunctions (and, also) are often used to tell a cohesive story (Graesser

et al. 2004), and syntax can be further used to identify narrative versus non-narrative language

(Jurafsky et al. 2009; Van Laer et al. 2017), which could be used to study transportation, a factor

that has been shown to affect consumers’ processing of advertising and media (Green and Brock

2002; Wang and Calder 2006). Categories like exclusion and conjunctive words also potentially

provide clues as to decision strategy—those using exclusion (e.g. “or”) might be using a

disjunctive strategy, while those using words such as “and” may be using a conjunctive strategy.

Certainty can be measured by tentative language, passive voice, and hedging phrases such as

counting occurrences of terms like “I think” or “perhaps”.

In sum, theories of semantics help consumer researchers link language with thought such

that they can use language to study different aspects of attention and emotion. Using theories of

syntax, on the other hand, sheds light on the complexity of thought, as it looks for markers of

structure that indicate the nature—complexity, order, or extent of—thinking, which has

implications for processing and persuasion (Petty, Cacioppo, and Schumann 1983). Here, text

analysis can be used to test predictions or hypotheses about attention and processing in real

world data, even if it cannot necessarily determine cognitive mechanism underlying the process.

Interpersonal dynamics
 19

The study of interpersonal dynamics—including the role of status, power, and social

influence in consumer life—can be meaningfully informed by linguistic theory and text analysis.

Social interaction and influence are key parts of consumer life, but can be difficult to study in the

lab. Consumers represent a lot about their relationships through language they use, and we can

use this knowledge to understand more about consumer relationships on both the dyadic and

group level.

The theory for linking language with social relationships comes from the field of

pragmatics, which studies the interactions between extra-linguistic factors and language.

Goffman (1959) and linguists following in the field of pragmatics such as Grice (1970) argue

that people use linguistic and non-linguistic signs to both signal and govern social relationships

indicating status, formality, and agreement. By understanding when, how, and why people tend

to use these markers, we can understand social distance (e.g. McTavish et al. 1995), power (e.g.

(Danescu-Niculescu-Mizil et al. 2012b), and influence (Gruenfeld and Wyer 1992). Pragmatics

is used to study how these subtle, yet pervasive cues structure human relationships and represent

the dynamics of social interaction in turn. In fact, about 40% of language is composed of these

functional markers (Zipf 1932).

One way to capture pragmatic elements is through the analyses of pronouns (e.g., “I”,

“me,” “they”) and demonstratives (i.e., “this”, “these”, “that”, and “those”), words that are the

same over multiple contexts, but whose meaning is indexical or context dependent (Nunberg

1993). Pronoun use can be guided by different sets of contextual factors such as intimacy,

authority or self-consciousness, and pragmatic analyses can be usefully applied to research that

pertains to theories of self and interpersonal interaction, particularly through the measurement of

pronouns (Packard, Moore, and McFerran 2014; Pennebaker 2011). Pronouns can detect the

degree to which a speaker is lying (Newman et al. 2003), feeling negative emotion (Rude,
 20

Gortner, and Pennebaker 2004), and collaborating in a social group (Gonzales, Hancock, and

Pennebaker 2010). Similarly, linguistic theories suggest that demonstratives (“this” or “that”)

mark solidarity and affective functions and can therefore be effective in “achieving camaraderie”

and “establishing emotional closeness between speaker and addressee” (Lakoff 1974, p. 351;

Potts and Schwarz 2010). Demonstratives have social effects, as shown in both qualitative and

quantitative analyses of politicians’ speeches (Acton and Potts 2014), and can be used for

emphasis. For example, product and hotel reviews with demonstratives (e.g. “that” or “this”

hotel) have more polarized ratings (Potts and Schwarz 2010).

Through pragmatics, speakers also signify differences in status and power. For example,

people with high status use more first person plural (“we”) and ask fewer questions (Sexton and

Helmreich 2000), while those with low status use more first person singular like “I” (Kacewitz et

al. 2011; Hancock et al. 2010). Language also varies systematically according to gender, and

many argue this is due to socialization into differences in power reflected in tentative language,

self-referencing, and the use of adverbs and other qualifiers (Herring 2000, 2003; Lakoff 1973).

Because norms are important in language, dyadic components prove to be an important

part of the analysis when studying interpersonal interaction. For example, by incorporating

dyadic interaction versus analyzing senders and receivers in isolation, Jurafsky et al. (2009)

improve accuracy in their identification of flirtation, awkwardness, and friendliness from a range

of 51 to 72% to 60 to 75%, with the prediction for women being the most improved when

accounting for dyadic relationships.

Language is social, and pragmatics illustrate that not all words are meant to carry

meaning. Phatic expressions, for example, are phrases in which the speaker’s intention (i.e. what

is meant) is not informative, but rather, social or representational (Jakobson 1960; Malinowski

1972). For example, an expression like “How about those Cubs?” is an invitation to talk about
 21

the baseball team, not a sincere question. A tweet like “I can’t believe Mariah Carey’s album

comes out on Monday!” is not intended to communicate information about a personal belief or

even the release date, but is an exclamation of excitement (Marwick and boyd 2011). The phatic

function can be informative in text analysis when one is interested simply in a word’s ability to

represent a category or concept to make it accessible or to form bond with others, irrespective of

semantic content. Here, the mention of the name is not used as a measure of semantics or

meaning but rather of presence versus absence, and hence mere accessibility, and, more broadly,

cultural awareness (Arvidsson and Caliandro 2016).

Group and Cultural Level Characteristics

Lastly, and of particular interest to scholars of sociology and culture, language can be

used to represent constructs at the group, cultural, and corpus level. At this level, group attention,

differences amongst groups, the collective structure of meaning or agreement shared by groups,

and changes in cultural products over time can be measured. Further, the ability of text analysis

to span levels of analysis from individuals to dyadic, small group, and subcultural interaction is

particularly apt for a multi-disciplinary field like consumer research.

In socio-cultural research, semantics is again key because words can represent patterns of

cultural or group attention (Gamson and Modigliani 1989; McCombs and Shaw 1972; Schudson

1989). For example, Shor et al.’s (2015) study of gender representation in the news measures the

frequency of female names in national newspapers to understand changes in the prominence of

women in public discourse over time (Shor et al. 2015), and van de Rijt et al. (2013) similarly

use name mentions to measure the length of fame. These are matters of attention, but this time

public, collective attention rather than individual attention. Historical trends in books (Twenge,

Campbell, and Gentile 2012) and song lyrics have also been discovered through text analysis.
 22

For example, in a study of all text uploaded to Google books (4% of what has been published),

Michael et al. (2011) find a shift from first person plural pronouns (we) to first person singular

(I, me), and interpret this as reflecting a shift from collectivism to individualism (see also

DeWall et al. 2011). Merging these approaches with an extra-linguistic DV, researchers can

sometimes predict book sales, movie success (Mestyán, Yasseri, and Kertész 2013), and even

stock market price using textual data from social media (Bollen, Mao, and Zeng 2011; De

Choudhury et al. 2008; Gruhl et al. 2005).

Studies of framing and agenda setting naturally use semantic properties to study the

social shaping of public opinion (Benford and Snow 2000; Gamson and Modigliani 1989;

McCombs and Shaw 1972). For example, measuring the diffusion of terms such as “illegal

immigrants” versus “undocumented workers” helps sociologists and socio-linguists understand

the role of social movements in setting the agenda and shaping public discourse (Lakoff and

Ferguson 2015). Humphreys and Thompson (2014), for example, use text analysis to understand

how news narratives culturally resolve anxiety felt by consumers in the wake of a crisis such as

an oil spill.

However, some caution is warranted when using cultural products to represent the

attitudes and emotions of a social group. Sociologists and critical theorists acknowledge a gap

between cultural representation and social reality (Holt 2004; Jameson 2013). That is, the

presence of a concept in public discourse does not mean that it directly reflects attitudes of all

individuals in the group. In fact, many cultural products often depict fantasy or idealized

representations that are necessarily far from reality (Jameson 2013). For this reason, as we will

discuss, sampling and an awareness of the source’s place in the larger media system are

particularly important when conducting this kind of socio-cultural analysis.

 23

In addition to using words to measure patterns of individual and collective attention,

researchers can put together textual elements to code patterns such as narrative (Van Laer et al.

2017), style matching (Ludwig et al. 2013; Ludwig et al. 2016), and linguistic cohesiveness

(Chung and Pennebaker 2013). For example, people tend to use the same proportion of function

words in cities where there is more even income distribution (Chung and Pennebaker 2013). In

studies of consumption this could be used to study agreement in co-creation (Schau, Muniz, and

Arnould 2009) and subcultures of consumption (Schouten and McAlexander 1995), and perhaps

even to predict fissioning of a group (Parmentier and Fischer 2015). One might speculate that

homophilous groups will display more linguistic cohesiveness, and this may even affect other

factors like strength of group identity, participation in the group, and satisfaction with group

outcomes. Words associated with assent like “yes” and “I agree” can be used to measure group

agreements (Tausczik and Pennebaker 2010). In this way, text analysis can be used for studying

group interactional processes to predict quality of products and satisfaction with participation in

peer to peer production (c.f. Mathwick, Wiertz, and De Ruyter 2008).

These four domains—attention, processing, interpersonal interaction, and group and

cultural level properties—provide rich fodder for posing research questions that can be answered

by studying language and by defining constructs through language. By linking linguistic theory

pertaining to semantics, pragmatics, and syntax, researchers can formulate more novel,

interesting, and theoretically rich research questions, and they can develop new angles on

constructs key to understanding consumer thought, behavior, and culture. Per our roadmap in

Figure 1, we now proceed with data collection.

Stage 3: Data Collection

 24

Once a researcher identifies the research question and related constructs, the next step is

to collect the data. There are four steps to data collection: identifying, unitizing, preparing, and

storing the data.

Identifying Data Sources

One virtue—and perhaps also a curse—of text analysis is that many data sources are available.

Query through experiment, survey or interview, web scraping of internet content (Newsprosoft 2012;

Pagescrape 2006; Velocityscape 2006; Wilson 2009), archival databases (ProQuest; Factiva), digital

conversion of printed or spoken text, product websites like Amazon.com, expert user groups like

Usenet, and internet subcultures or brand communities are all potential sources of consumer data. In

addition to data collection through scraping, some platforms like Twitter offer access to a 10% random

sample of the full “firehose” or APIs for collecting content being posted by users according to

keyword, hashtag, or user types.

Sampling is likely the most important consideration in the data collection stage. In the abstract,

any dataset will consist of some sample from the population, and the sample can be biased in various

but important ways. For example, Twitter users are younger and more urban than a representative

sample of the US population (Duggan 2013; Mislove et al. 2011). Generally, only public Facebook

posts are available to study, and these users may have different characteristics than those who restrict

posts to private. In principle, these concerns are no different from those present in traditional content

analysis (see Krippendorff 2004 for a discussion of sampling procedures). However, sampling from

internet data presents two unique issues.

First, filtering and promotion on websites can push some posts to the top of a list. On the one

hand, this makes them more visible—and perhaps more influential—but on the other hand, they may

be systematically different from typical posts. These selection biases are known to media scholars and
 25

political scientists, and researchers have a variety of ways for dealing with these kinds of enduring, but

inevitable biases. For example, Earl et al. (2004) draw from methods previously used in survey

research to measure non-response bias through imputation (see e.g. Rubin 1987). Researchers can

sample evenly or randomly from categories on the site to obviate the problem of filtering on the

website itself (see e.g. Moore 2015).

The second issue is that keyword search can also introduce systematic bias. Reflecting on the

issues introduced by semantic framing, researchers may miss important data because they have the

wrong phrasing or keyword. For instance, Martin, Pfeffer, and Carley (2013) find that although the

conceptual map for interviews and the text of newspaper articles about a given topic largely overlaps,

article keywords—which are selected by the authors and tend to be more abstract—do not. There are at

least two ways to correct for this. First, researchers can skip using keyword search altogether and

sample using site architecture or time (Krippendorff 2004). Second, if keywords are necessary,

researchers can first search multiple keywords and provide information about search numbers and

criteria for selection, providing, if required, analyses of alternative keyword samples.

In addition to considering these two unique issues, a careful and defensible sampling strategy

that is congruent with the research question should be employed. If, for example, categories such as

different groups or time periods are under study, a random stratified sampling procedure should be

considered. If the website offers categories, the researcher may want to stratify by these existing

categories. A related issue is the importance of sound sampling when using cultural products or

discourse to represent a group or culture of interest. For instance, DeWall et al. (2011) use top

10 songs and Humphreys (2010) uses newspapers with the largest circulation based on the

inference that they will be the most widely-shared cultural artifacts. As a general guide, being

aware of the place and context of the discourse—including its senders, medium, and receivers

(Hall 1980)—can resolve many issues (see also Golder 2000).

 26

Additionally, controls may be available. For example, although conducting traditional content

analysis, Moore (2015) compares non-fiction to fiction books when evaluating differences between

utilitarian verses hedonic products rather than sampling from two different product categories. At the

data collections stage, metadata can also be collected and later used to test alternative hypotheses

introduced by selection bias (e.g. age).

Researchers need to account for sample size, and in the case of text analysis, there are

two size issues to consider—the number of documents available, and the amount of content (e.g.

number of words and sentences) in each document. Depending on the context as well as the

desired statistical power, the requirements will differ. One method to avoid overfitting or biases

due to small samples is the Laplace correction, which starts to stabilize a binary categorical

probabilistic estimate when a sample size reaches thirty (Provost and Fawcett 2013). As a

starting rule-of-thumb, having at least thirty units is usually needed to make statistical inferences,

especially since text data is non-normally distributed. However, this is an informal guideline, and

depending on the detectable effect size, a power analysis would be required to determine the

appropriate sample size (Corder and Foreman 2014). One should be mindful of the virtues of

having a tightly controlled sample for making better inferences (Pauwels 2014). Big data is not

always better data (Borgman 2015).

Regarding the number of words per unit, if using a measure that accounts for length of

the unit (e.g. words as a percent of the total words), data can be noisy if units are short, as in

tweets. Tirunillai and Tellis (2012), for example, discard online reviews that have fewer than ten

words. However, the number of words required per unit is largely dependent on the base

frequency of dictionary or concept-related keywords, as we will later discuss. For psychometric

properties like personality, Kern et al (2016) suggest having at least 1,000 words per person and
 27

a sufficiently large dataset of users to cover variation in the construct. In their case, extraversion

could be reliably predicted using a set of 4,000 Facebook users with only 500 words per person.

Preparing Data

After the data is identified and stored as a basic text document, it needs to be cleaned and

segmented into units that will be analyzed. Spell-checking is often a necessary step because text

analysis assumes correct, or at least consistent, spelling (Mehl and Gill 2008). Problematic

characters such as wing dings, emoticons, and asterisks should be eliminated or replaced with

characters that can be counted by the program (e.g. “smile”). On the other hand, if the research

question pertains to fluency (e.g. Jurafsky et al. 2009) or users’ linguistic habits, spelling

mistakes and special characters should be kept and analyzed by custom programming. Data

cleaning—which includes looking through the text for irrelevant text or markers—is important, as false

inferences can be made if there is extraneous text in the document. For example, in a study of post-9/11

text, Back, Küfner, and Egloff (2010) initially reported a falsely inflated measurement of anger because

they did not clean automatically generated messages (“critical” error) in their data (Back, Küfner, and

Egloff 2011; Pury 2011).

Languages other than English can also pose unique challenges. Most natural language

processing tools and methodologies that exist today focus English, a language from a low-context

culture composed of individual words that, for the most part, have distinct semantics, specific

grammatical functions, and clear markers for discrete units of meaning based on punctuation. However,

grammar and other linguistic aspects (e.g. declination) can meaningfully affect unitization decisions.

For example, analysis of character-based languages like Chinese requires first segmenting characters

into word units and then dividing sentences into meaningful sequences before a researcher can do part-

of-speech tagging or further analyses (Fullwood 2015; Sun et al. 2015).

 28

Unitizing and Storing the Data

After cleaning, a clearly organized file structure should be created. One straightforward

way to achieve this organization is to use one text file for each unit of analysis or “document.” If,

for example, the unit is one message board post, a text file can be created for each post. Data

should be segregated into the smallest units of comparison because the output can always be

aggregated upward. If, for example, the researcher is conducting semantic analysis of book

reviews, a text file can be created for each review, and then aggregated to months or years to

assess historical trends or star groupings to compare valence.

Two technological solutions are available to automate the process of unitizing. First,

separating files can be automated using a custom program such as a Word macro, to cut text

between identifiable character strings, paste it to a new document, and then save that document

as a new file. Secondly, many text analysis programs are able to segregate data within the file by

sentence, paragraph, or a common, unique string of text or this can be done through code. If units

are separated by one of these markers, separate text files will not be required. If, for example,

each message board post is uniquely separated by a hard return, the researcher can use one file

containing all of the text and then instruct the program to segment the data by paragraph. If the

research is comparing by groups, say by website, separate files should be maintained for each

site and then segregated within the file by a hard return between each post.

Researchers might also want to consider using a database management system (DBMS)

in addition to a file-based structure in order to support reproducible or future research, especially

for big datasets and for information other than discourse such as speakers’ attributes (e.g. age,

gender and location). If the text document that stores all the raw data starts to exceed the

processing machine’s random access memory RAM available (some 32-bit text software caps the
 29

file size at 2GB), it may be challenging to work with directly, and depending on the way the

content is structured in the text file, writing into a database may be necessary. For those

experienced with coding, a variety of tools exist for extracting data from text files into a database

(e.g., packages “base,” “quanteda” and “tm” in R or Python’s natural language processing

toolkit, NLTK).

Stage 4a: Choose an Operationalization Approach

Once data has been collected, prepared, and stored, the next decision is choosing the

appropriate research approach for operationalizing the constructs. Next to determining if text

analysis is appropriate, this is the most important impasse in the research. We discuss the pros

and cons of different approaches and provide guidance as to when and how to choose amongst

methods (Figure 1). The Web Appendix presents a demonstration of dictionary-based and

classification approaches, applied to understand word-of-mouth following a product launch.

If the construct is relatively clear (e.g. positive affect), one can use dictionary or rule set

to measure the construct. Standard, psychometrically-tested dictionaries are available for

measuring a variety of constructs (table 1). Researchers may want to consult this list or one like

it first to determine if the construct can be measured through an existing wordlist.

If the operationalization of the construct in words is not yet clear or the researcher wants

to make a posteriori discoveries about operationalization, one should use a classification

approach in which the researcher first identifies two or more categories of text and then analyzes

recurring patterns of language within these sets. For example, if the researcher wants to study

brand attachment by examining the texts produced by brand loyalists versus non-loyalists, but

does not know exactly how they differ or wants to be open to discovery, classification would be

appropriate. Here, the researcher preprocesses the documents and uses the resulting word
 30

frequency matrix as the independent variable (IV), with loyalty as the already-existing dependent

variable (DV). This leaves one open to surprises about which words may reflect loyalty, for

example.

At the extreme, if the researcher does not know the categories at play but has some

interesting text, she could use unsupervised learning to have the computer first detect groups

within the text and then further characterize the differences in those groups through language

patterns, somewhat like multi-dimensional scaling or factor analysis (see e.g. Lee and Bradlow

2011). As shown in Figure 1, selecting an approach depends on whether the constructs can be

clearly defined a priori, and we discuss these decisions in detail.

----Insert table 1 about here----

Top-down Approaches

Top-down approaches involve analyzing occurrences of words based on a dictionary or a

set of rules. If the construct is relatively clear—or can be made clear through human analysis of

the text (Corbin and Strauss 2008)—it makes sense to use a top-down approach. We discuss two

types of top-down approaches: dictionary-based and rule-based approaches. In principle, a

dictionary-based approach can be considered a type of rule-based approach; it is a set of rules for

counting concepts based on the presence or absence of a particular word. We will treat the two

approaches separately here, but thereafter will focus on dictionary-based methods, as they are

most common. Methodologically, after operationalization, the results can be analyzed in the

same way, although interpretation of the results may differ.

Dictionary-based Approach. Although certainly some of the most basic methods

available, dictionary-based approaches have remained one of the most enduring methods of text

analysis and are still used as a common tool in the text analysis toolkit to produce new
 31

knowledge (e.g. Boyd and Pennebaker 2015; Eichstaedt et al. 2015; Shor et al. 2015; Snefjella

and Kuperman 2015; see appendix).

Dictionary-based approaches have three advantages for research in consumer behavior

that draws from psychological or sociological theories. First, they are easy to implement and

comprehend, especially for researchers that have limited programming or coding experience.

Second, combined with the fundamentals of linguistics, they allow intuitive operationalization of

constructs and theories directly from sociology or psychology. Finally, the validation process of

dictionary-based approaches is relatively straightforward for non-specialists, and findings are

relatively transparent to reviewers and readers.

For a dictionary-based analysis, researchers define and then calculate measurements that

summarize the textual characteristics that represent the construct. For example, positive emotion

can be captured by the frequency of words such as “happy,” “excited,” “thrilled,” etc. The

approach is best suited for semantic and pragmatic markers, and attention, interaction, and group

properties have all been studied using this approach (see appendix).

One of the simplest measurements used in dictionary-based analysis is frequency, which

is calculated based on the assumption that word order does not matter. These methods, also

called the “bag of words” approach, assume that the meaning of a text depends only on word

occurrence, as if the words are drawn randomly from a bag. While these methods are based on

the strong assumption that word order is irrelevant, they can be powerful in many circumstances

for marking patterns of attentional focus and mapping semantic networks. For pragmatics and

syntax, counting frequency of markers in text can produce measurement of linguistic style or

complexity in the document overall. Note that when using a dictionary-based approach, tests will

be conservative. That is, by predetermining a wordlist, one may not pick up all instances, but if

meaningful patterns emerge, one can argue that there is an effect, despite the omissions.
 32

A variety of computer programs can be used to conduct top-down automated text analysis and

as auxiliaries for cleaning and analyzing the data. A word processing program is used to prepare the

text files, an analysis program is needed to count the words, and a statistical package is often necessary

to analyze the output. WordStat (Peladeau 2016), Linguistic Inquiry and Word Count (LIWC;

Pennebaker, Francis, and Booth 2007), Diction (North, Iagerstrom, and Mitchell 1999), Yoshikoder

(Lowe 2006), and Lexicoder (Daku, Young, and Soroka 2011) are all commonly used programs for

dictionary-based analysis although it is also possible with more advanced packages such as R and

Python.

Rule-based Approach. Rule-based approaches are based on a set of criteria that indicate a

particular operationalization. By defining and coding a priori rules according to keywords,

sentence structures, punctuations, styles, readability and other predetermined linguistic elements,

a researcher can quantify unstructured texts. For example, if a researcher is interested in

examining passive voice, she can write a program that, after tagging the part-of-speech (POS) of

the text, counts the number of instances of a subject followed by an auxiliary and a past

participle (e.g. “are used”). Van Laer et al. (2017) use a rule-based approach to classify sentences

in terms of genre, using patterns in emotion words to assign a categorical variable that classifies

a sentence as having rising, declining, comedic, or tragic action. Rule-based approaches are also

often used in readability measures (e.g., Bailey and Hahn 2001; Li 2008; Ghose, Ipeirotis, and Li

2012; Tan, Gabrilovich, and Pang 2012) to operationalize fluency of a message.

Bottom-up Approaches

In contrast to top-down approaches, bottom-up approaches involve examining patterns in

text first, and then proposing or interpreting more complex theoretical explanations and patterns.

Bottom-up approaches are used in contexts where the explanatory construct or the
 33

operationalization of constructs is unclear. In some cases where word order is important (for

example in syntactic analyses), bottom-up approaches via unsupervised learning may also be

helpful (Chambers and Jurafsky 2009). We discuss two common approaches used in text

analysis: classification and topic discovery.

Classification. In contrast to dictionary-based or rule-based approaches, where the

researcher explicitly identifies the words or characteristics that represent the construct,

classification approaches are used when dealing with constructs that may be more latent in the

text, meaning that the operationalization of a construct in text cannot be hypothesized a priori.

Instead of manually classifying every document of interest, supervised classification allows

researchers to group texts into pre-defined categories based on a subset or “training” set of the

data. For example, Eliashberg, Hui, and Zhang (2007) classify movies based on their return on

investment and then, using the movie script, determine the most important factors in predicting a

film’s return-on-investment such as action genre, clear and early statement of the setting, and clear

premise. After discovering these patterns, they theorize as to why they occur.

There are two advantages to using classification. First, it reduces the amount of human

coding required, yet produces clear distinctions between texts. While dictionary-based

approaches provide information related to magnitude, classification approaches provide

information about type and likelihood of being of a type, and researchers can go a step further by

understanding what words or patterns lead to being classified as a type. Second, the classification

model itself can reveal insights or test hypotheses that may be otherwise buried in a large amount

of data. Because classification methods do not define a wordlist a priori, latent elements, such as

surprising combinations of words or patterns that may have been excluded in a top-down

analysis, may be revealed.

 34

Researchers use classification when they want to know where one text stands in respect

to an existing set or when they want to uncover meaningful, yet previously unknown patterns in

the texts. In digital humanities research, for example, Plaisant et al. (2006) use a multinomial

naïve Bayes classifier to study word associations commonly associated with spirituality in the

letters of Emily Dickenson. They find that, not surprisingly, words such as “Father and Son” are

correlated with religious metaphors, but they also uncover the word “little” as a predictor, a

pattern previously unrecognized by experienced Dickenson scholars. This discovery then leads to

further hypothesizing about the meaning of “little” and its relationship to spirituality in

Dickenson’s poems. In consumer research, in studying loyalists vs. non-loyalists, researchers

might find similarly surprising words such as hope, future, and improvement, and these insights

might provoke further investigation into self-brand attachment and goal orientation.

Topic Discovery. If a researcher wants to examine the text data without a priori

restrictions on words, rules, or categories, a topic discovery model is appropriate. Discovery

models such as Latent Dirichlet Allocation (LDA) are analyses that recognize patterns within the

data without predefined categories. In the context of text analysis, discovery models are used to

identify whether certain words tend to occur together within a document, and such patterns or

groupings are referred to as “topics.” Given its original purpose, topic discovery is used

primarily to examine semantics. Topic discovery models typically take a word frequency matrix

and output groupings that identify co-occurrences of words, which can then predict the topic of a

given text. They can be helpful when researchers want to have an overview of the text beyond

simple categorization or to identify patterns.

Topic discovery models are especially useful in situations where annotating even a subset

of the texts has a high cost due to complexity, time or resource constraints, or a lack of distinct, a

priori groupings. In these cases, a researcher might want a systematic, computational approach
 35

that can automatically discover groups of words that tend to occur together. For example,

Mankad et al. (2016) use unsupervised learning and find that hotel reviews mainly consist of five

topics, which, according to the groups of words for each topic, they label as “amenities,”

“location,” “transactions,” “value,” and “experience.” Once topics have been identified, one can

go on to study their relationship with each other and with other variables such as rating.

Stage 4b: Execute Operationalization

After choosing an approach, the next step is to make some analytical choices within the

approach pertaining to either dictionary or algorithm type. These again depend on the clarity of

the construct, the existing methods for measuring it, and the researcher’s propensity for

theoretically-driven versus data-driven results. Within top-down approaches, decisions entail

choosing one or more standardized dictionaries versus creating a custom dictionary or ruleset.

Within bottom-up methods of classification and topic modeling, analytic decisions entail

choosing a technique that fits suitable assumptions and the clarity of output one seeks (e.g.

mutually-exclusive versus fuzzy or overlapping categories).

Dictionary- and Rule-based Approaches

Standardized Dictionary. If one chooses a dictionary-based approach, the next question is

whether to use a standardized dictionary or to create one. Dictionaries exist for a wide range of

constructs in psychology, and less so, sociology (Table 1). Sentiment, for example, has been

measured using many dictionaries: Linguistic Inquiry Word Count (LIWC), ANEW (Affective

Norms for English Words), the General Inquirer (GI), SentiWordNet, WordNet-Affect, and

VADER (Valence Aware Dictionary for Sentiment Reasoning). While some dictionaries like

LIWC are based on existing psychometrically tested scales such as PANAS, others such as
 36

ANEW (Bradley and Lang 1999) have been created based on previous classification applied to

offline and/or online texts and human scoring of sentences (Nielsen 2011). VADER (Hutto and

Gilbert 2014) includes the word banks of established tools like LIWC, ANEW, and GI, as well

as special characters such as emoticons and cultural acronyms (e.g. LOL), which makes it

advantageous for social media jargon. Additionally, VADER’s model incorporates syntax and

punctuation rules, and is validated with human coding, making its sentence prediction 55%-96%

accurate, which is on par with Stanford Sentiment Treebank, a method that incorporates a more

complex computational algorithm (Hutto and Gilbert 2014). However, a dictionary like LIWC

bases affect measurement on underlying psychological scales, which may provide tighter

construct validity. If measuring a construct such as sentiment that has multiple standard

dictionaries, it is advisable to test the results using two or more measures, as one might employ

multiple operationalizations.

In addition to standardized dictionaries for measuring sentiment, there are a range of

psychometrically-tested dictionaries for concepts like construal level (Snefjella and Kuperman

2015), cognitive processes, tense, and social processes (Linguistic Inquiry Word Count;

Pennebaker, Francis, and Booth 2001), pleasure, pain, arousal, motivation (Harvard IV

Psychological Dictionary; Dunphy, Bullard, and Crossing 1974) primary versus secondary

cognitive processes (Regressive Imagery Dictionary; Martindale 1975) and power (Lasswell’s

Value Dictionary; Lasswell and Leites 1949; Namenwirth and Weber 1987; table 1). These

dictionaries have been validated with a large and varied number of text corpora, and because

operationalization does not change, standard dictionaries enable comparison across research,

enhancing concurrent validity amongst studies.

For this reason, if a standard dictionary exists, researchers should use it if at all possible

to enhance the replicability of their study. If they wish to create a new dictionary for an existing
 37

construct, researchers should run and compare the new dictionary to any existing dictionary for

the construct, just as one would with a newly developed scale (Churchill 1979).

Dictionary Creation. In some cases, a standard dictionary may not be available to

measure the construct, or semantic analyses may require greater precision to measure culturally

or socially specific categories. For example, Eritimur and Coskuner-Balli (2015) use a custom

dictionary to measure the presence of different institutional logics in the market emergence of

yoga in the United States.

To create a dictionary, researchers first develop a word list, but here there are several

potential approaches (Figure 1). For theoretical dictionary development, one can develop the

wordlist from previous operationalization of the construct, scales, and by querying experts. For

example, Pennebaker, Francis, and Booth (2007) use the Positive and Negative Affect Schedule

or PANAS (Watson, Clark, and Tellegen 1988), to develop dictionaries for anger, anxiety, and

sadness. To ensure construct validity, however, it is crucial to examine how these constructs are

expressed in the text during post-measurement validation.

If empirically guided, a dictionary is created from reading and coding the text. The

researcher selects a random subsample from the corpus in order to create categories using the

inductive method (Katz 2001). If the data is skewed (i.e. if there are naturally more entries from

one category than others), a stratified random sampling should be used to ensure that categories

will evenly apply to the corpus. Generally sampling 10 to 20% of the entire corpus for qualitative

dictionary development is sufficient (Humphreys 2010). Alternatively, the size of the subsample

can be determined as the dictionary is developed using a saturation procedure (Weber 2005). To

do this, code 10 entries at a time until a new set of 10 entries yields no new information. Corbin

and Strauss (1990) discuss methods of grounded theory development that can be applied here for

dictionary creation.
 38

If the approach to dictionary development is purely inductive, researchers can build the

wordlist from a concordance of all words in the text, listed according to frequency (Chung and

Pennebaker 2013). In this way, the researcher acts as a sorter, grouping words into common

categories, a task that would be performed by the computer in bottom-up analysis. One

advantage of this approach is that it ensures that researchers do not miss words that occur in the

text that might be associated with the construct.

After dictionary categories are developed, the researcher should expand the category lists

to include relevant synonyms, word stems, and tenses. The dictionary should avoid homonyms

(e.g. river “bank” vs. money “bank”) and other words where reference is unclear (see Rothwell

2007 for a guide). Weber (2005) suggests using the semiotic square to check for completeness of

concepts included. For example, if “wealth” is included in the dictionary, perhaps “poverty”

should also be included. Because measurement is taken from words, one must attend to and

remove words that are too general and thus produce false positives. For example, “pretty” can be

used as a positive adjective (e.g. “pretty shirt”) or for emphasis (e.g. “that was pretty awful”).

Alternatively, a rule-based approach can be used to work around critical words that cause false

positives in a dictionary. It is then important that rules for inclusion and exclusion be reported in

the final analysis.

Languages other than English can produce challenges in dictionary creation. If one is

developing a dictionary in a language or vernacular where there are several spellings or terms for

one concept, for example, researchers should include those in the dictionary. Arabic, for

example, is characterized by three different vernaculars within the same language—Classical

Arabic in religious texts, Modern Standard Arabic, and a regional dialect (Farghaly and Shaalan

2009). Accordingly, researchers should be mindful of these multi-valences, particularly when

 39

developing dictionaries in other languages or even other vernaculars within English (e.g. internet

discourse).

Dictionary Validation. After developing a preliminary dictionary, its construct validity

should be assessed. Does each word accurately represent the construct? Researchers have used a

variety of validation techniques. One method of dictionary validation is to use human coders to

check and refine the dictionary (Pennebaker et al. 2007). To do this, the dictionary is circulated

to three research assistants who vote to either include or exclude a word from the category and

note words they believe should be included in the category. Words are included or excluded

based on the following criteria: (1) if two of the three coders vote to include it, the word is

included, (2) if two of the three coders vote to exclude it, the word is excluded, (3) if two of the

three coders offer a word that should be included, it is added to the dictionary.

A second option for dictionary validation is to have participants play a more involved role in

validating the dictionary through survey-based instruments. Kovács et al. (2013), for example, develop

a dictionary by first generating a list of potential synonyms and antonyms to their focal construct,

authenticity, and then conducting a survey in which they have participants choose the word closest to

authenticity. They then use this data to rank words from most to least synonymous, assigning each a

score from 0 to 1. This allows dictionary words to be weighted as more or less part of the construct

rather than either-or indicators. Another option for creating and validating a weighted dictionary is to

regress textual elements on a dependent variable like star rating to get predictors of, say, sentiment.

This approach would be similar to the bottom-up approach of classification (see e.g. Tirunillai and

Tellis 2012).

Post-Measurement Validation. After finalizing the dictionary and conducting a preliminary

analysis, the results should be examined to ensure that operationalization of the construct in words

occurred as expected, and this can be an iterative process with dictionary creation. The first method of
 40

post-measurement validation uses comparison with a human coder. To do this, select a subsample of

the data, usually about 20 entries per concept, and compare the computer coding with ratings by a

human coder. Calculate Krippendorff’s alpha to assess agreement between the human coder and the

computer coder (Krippendorff 2010; Krippendorff 2007). Traditional criteria for reliability apply;

Krippendorff’s alpha for each category should be no lower than 70%, and the researcher should

calculate Krippendorff’s alpha for each category and as an average for all categories (Weber 2005).

Packard and Berger (2016) conduct this type of validation, finding 94% agreement between computer

and human coded reviews.

The advantages of using a human coder for post-measurement validation are that results can be

compared to other traditional content analyses and that this method separates validation from the

researcher. However, there are several disadvantages. First, it is highly variable because it depends on

the expertise and attentiveness of one or more human coders. Secondly, traditional measures of inter-

coder reliability such as Krippendorf’s alpha were intended to address the criteria of replicability

(Hughes and Garrett 1990; Krippendorff 2004), the chance of getting the same results if the analysis

were to be repeated. Because replicability is not an issue with automated text analysis—the use of a

specific word list entails that repeated analyses will have exactly the same results—measures of inter-

coder agreement are largely irrelevant. While it is important to check the output for construct validity,

the transparency of the analysis means that traditional measures of agreement are not always required

or helpful. Lastly, and perhaps most importantly, the human coder will likely be more sensitive to

subtleties in the text, and may therefore over-code categories or may miscode due to unintentional

mistakes or biases. After all, one reason the researcher selects automated text analysis is to capture

aspects humans cannot detect.

The second alternative for validation is to perform a check oneself or to have an expert perform

a check on categories using a saturation procedure. Preliminarily run the dictionary on the text and
 41

examine 10 instances at a time, checking for agreement with the construct or theme of interest and

noting omissions and false positives (Weber 2005). The dictionary can then be iteratively revised to

reduce false positives and include observed omissions. A hit rate, the percent of accurately coded

categories, and a false hit rate, the percent of inaccurately coded categories, can be calculated and

reported. Thresholds for acceptability using this method of validation are a hit rate of at least 80% and a

false hit rate of less than 10% (Wade, Porac, and Pollock 1997; Weber 2005).

Like any quantitative research technique, there will always be some level of measurement

error. Undoubtedly, words will be occasionally mis-categorized; such is the nature of language. The

goal of validation is to ensure that measurement error is low enough relative to the systematic variation

so that the researcher can make reliable conclusions from the data.

Classification

After choosing a bottom up approach, the next question is determining whether a priori

classifications are available. If the answer is yes, the researcher can use classification,

supervised-learning methods. Here we discuss Naïve Bayes classification, logistic regressions,

and classification trees because of the ease of their implementation and interpretability. We will

also discuss neural networks and k-nearest neighbor classifications, which, as we will describe

below, are more suited for predicting categories of new texts than for deriving theories or

revealing insights.

Naïve Bayes (NB) predicts the probability of a text belonging to a category given its

attributes with Bayes rules- and the “naïve” assumption that each attribute in the word frequency

matrix is independent from each other. NB has been applied in various fields such as marketing,

information, and computer science. Examining whether online chatter affects a firm’s stock market

performance, Tirunilai and Tellis (2012) use NB to classify a user-generated review as positive or
 42

negative. Using star rating to classify reviews as positive or negative a priori, they investigate

language that is associated with these positive or negative reviews. Since there is no complex

algorithm involved, NB is very efficient with respect to computational costs. However, in

situations where words are highly correlated with each other, NB might not be suitable.

Logistic regression is another classification method, and similar to NB, it also takes a word

frequency or characteristic matrix as input. It is especially useful when the dataset is large and when

the assumption of conditional independence of word occurrences cannot be taken for granted. For

example, Thelwall et al. (2010) use it to predict positive and negative sentiment strength for short

online comments from MySpace.

In contrast, a classification tree is based on the concept of examining word combinations in a

piecewise fashion. Namely, it first splits the texts with the word or category that can distinguish the

most variation, and then within each resulting “leaf,” it splits the subsets of the data again with

another parameter. This inductive process iterates until the model achieves the acceptable error rate

that is set by the researcher beforehand (see later sections for guidelines on model validation).

Because of their conceptual simplicity, the classification tree is also a “white box” that allows for

easy interpretation.

There are other classification methods such as neural networks (NN) or k-nearest

neighbor (k-NN) that are more suitable for prediction purposes, but less for interpreting insights.

However, these types of “black box” methods can be considered if the researcher requires only

prediction (e.g., positive or negative sentiments), but not enumeration of patterns underlying the

prediction.

In classifying a training set, researchers apply some explicit meaning based on the words

contained within the unit. Classification is therefore used primarily to study semantics, while

applications of classificatory, bottom-up techniques for analyzing pragmatics and syntax remain
 43

a nascent area (Kuncoro et al. 2016). However, more recent research has demonstrated the utility

of these approaches to study social factors such as detecting politeness (Danescu-Niculescu-

Mizil et al. 2013), predicting lying or deceit (Markowitz and Hancock 2015; Newman et al.

2003), or sentiment analysis that accounts for sentence structures (e.g., Socher et al. 2012).

Topic-Discovery

If there is no a priori classification available, topic models, implemented via unsupervised

learning, are more suitable. Predefined dictionaries are not necessary since unsupervised

methods inherently calculate the probabilities of a text being similar to another text and group

them into topics. Some methods, such as Latent Dirichlet Allocation (LDA) assume that a

document can present multiple topics and estimate the conditional probabilities of topics, which

are unobserved (i.e., latent), given the observed words in documents. This can be useful if the

researcher prefers “fuzzy” categories to the strict classification of the supervised learning

approach. Other methods, such as k-means clustering, use the concept of distance to group

documents that are the most similar to each other based on co-occurrence of words or other types

of linguistic characteristics. We will discuss the two methods, LDA and k-means, in more detail.

LDA is one of the most common topic discovery models (Blei 2012), and it can be

implemented in software packages or libraries such as R and Python. Latent Dirichlet Allocation

(LDA; Blei, Ng, and Jordan 2003) is a modeling technique that identifies whether and why a

document is similar to another document and specifies the words underlying the unobserved

groupings (e.g., topics). Its algorithm is based on the assumptions that 1) there is a mixture of

topics in a document, and this mixture follows a Dirichlet distribution, 2) words in the document

follow a multinomial distribution, and 3) the total of N words in a given document follows a

Poisson distribution. Based on these assumptions, the LDA algorithm estimates the most likely

underlying topic structure by comparing observed words groupings with these probalistic
 44

distributions and then outputs K groupings of words that are related to each other. Since a

document can belong to multiple topics, and a word can be used to express multiple topics, the

resulting groupings may have overlapping words. LDA reveals the underlying topics of a given

set of documents, and the meanings are interpreted by the researcher.

Yet sometimes this approach will produce groupings that don’t semantically hang

together or groupings that are too obviously repetitive. To resolve this issue, researchers will

sometimes use word embedding, a technique for reducing and organizing a word matrix based on

similarities and dissimilarities in semantics, syntax, and part-of-speech that are taken from

previously-observed data. The categories taken from large amounts of previously-observed data

can be more comprehensive as well as more granular than the categories specified by existing

dictionaries such as LIWC’s sentiments. Further, in addition to training embeddings from the

existing dataset, a researcher can download pre-trained layers such as word2vec by Google

(Mikolov et al. 2013) or GloVe by Stanford University (Pennington, Socher, and Manning

2014). In these cases, the researcher skips the training stage and jumps directly to text analysis.

These packages provide a pre-trained embedding structure as well as functions for a researcher to

customize the categories depending on the research context in question. As a supplement to

LDA, once word embeddings have been learned, they can potentially be reused.

In consumer research, LDA is useful in determining ambiguous constructs such as

consumer perceptions, particularly if the corpus is large. For example, Tirunillai and Tellis

(2014) analyze 350,000 consumer reviews with LDA to group the contents into product

dimensions that reviewers care about. In the context of mobile phones, for example, they find

that the dimensions are “portability,” “signal receptivity,” “instability,” “exhaustible,”

“discomfort,” and “secondary features.” LDA allows Tirunillai and Tellis (2014) to

simultaneously derive product dimensions and review valence by labeling the grouped words as
 45

“positive” or “negative” topics. The canonical LDA algorithm is a bag-of-word model, and one

potential area of future research is to relax the LDA assumptions. For instance, Büschken and

Allenby (2016) extend the canonical LDA algorithm by identifying not just words, but whole

sentences, that belong to the same topic. If applying this method to study consumer behavior one

could use topic discovery to identify tensions in a brand community or social network which

could lead to further theorizations about underlying discourse or logic present in a debate.

In cases where a researcher wants to consider linguistic elements beyond word

occurrences, conceptually simpler approaches such as clustering may be more appropriate (Lee

and Bradlow 2011). In addition to word occurrences, a researcher can first code the presence of

syntax or pragmatic characteristics of interest, and then perform analyses such as k-means

clustering, which is a method that identifies “clusters” of documents by minimizing the distance

between a document and its neighbors in the same cluster. After obtaining the clustering results,

the researcher can then profile each cluster, examine its most distinctive characteristics, and

further apply theory to explain topic groupings and look for further patterns through abduction

(Peirce 1957).

Labeling the topics is the last, and perhaps the most critical step, in topic discovery. It is

important to note that, despite the increasing availability of big data and machine learning

algorithms and tools, the results obtained from these types of discovery models are simply sets of

words or documents grouped together to indicate that they constitute a topic. However, what that

topic is or represents can only be determined by applying theory and context-specific knowledge

or expertise when interpreting the results.

Stage 5: Interpretation and Analysis

 46

After operationalizing the constructs through text analysis, the next step is to analyze and

interpret the results. There are two distinct phases of analysis: the text analysis itself and the statistical

analysis, already familiar to many researchers. In this section we discuss three common ways of

incorporating the results of text analysis into research design: 1) comparison between groups, 2)

correlation between textual elements, and 3) prediction of variables outside the text.

Comparison

Comparison is the most common research design amongst articles that use text analysis in the

social sciences, and is particularly compatible with top-down, dictionary-based techniques (see

appendix). Comparing between groups or over time is useful for answering research questions that

relate directly to the theoretical construct of interest. That is, some set of text is used to represent the

construct and then comparisons are made to assess statistically meaningful differences between texts.

For example, Kacewicz (2014) et al. compare the speech of high power versus low power individuals

(manipulated rather than measured), finding that high power people use fewer personal pronouns

(“I”). Investigating the impact of religiosity, Ritter et al. (2013) compare Christians to atheists,

finding that Christians express more positive emotion words than atheists, which the authors attribute

to a different thinking style. Holoein and Fiske (2014) compare the word use of people who were told

to be warm versus people who were told to appear competent, finding a compensatory relationship

whereby people wanting to appear warm also select words that reflect low competence.

Other studies use message type rather than source to represent the construct and thus as the

unit of comparison. Bazarova and colleagues (2012), for example, compare public to private

Facebook messages to understand differences in the style of public to private communication. One

can also compare observed frequency in a dataset to a large corpus such as the standard Corpus of

American English or the Brown Corpus (Conrad 2002; Neuman et al. 2012; Pollach 2012; Wood and
 47

Kroger 2000). In this way, researchers can assess if frequencies are higher than ‘typical’ usage in

English, not just relative to other conditions in their own text.

Comparisons over space and time are also common and valuable for assessing how a

construct can change in magnitude based on some external variable. In contrast to group

comparisons, these studies tend to focus on semantic aspects over pragmatic of syntactic ones. For

example, Dore et al. (2015) trace changes in emotional language following the Sandy Hook School

shooting, finding that emotion words like sadness decreased with spatial and physical distance while

anxiety increased according to distance. Comparing different periods of regulation, Humphreys

(2010) shows how discourse changes over time as the consumer practice of casino gambling

becomes legitimate.

Issues with Comparison. Because word frequency matrices can contain a lot of zeroes (i.e.

each document may only contain a few instances of a keyword), researchers should use caution

when making comparisons between word frequencies of different groups. In particular, the lack

of normally distributed data violates the assumptions for tests like ANOVA, and simple

comparative methods like Pearson’s Chi-squared tests and z-score tests might yield biased results.

Alternative comparative measures such as likelihood methods or linear regression may be more

appropriate (Dunning 1993). Another alternative is using non-parametric tests that do not rely on the

normality assumption. For instance, the non-parametric equivalent of a one-way analysis of variance

(ANOVA) is the Kruskal-Wallis test, whose test statistic is based on ordered rankings rather than

means.

Many text analysis algorithms take word counts or the “term-frequency” (tf) matrix as an

input, but because word frequencies do not follow a normal distribution (Zipf 1932), many

researchers transform the data prior to statistical analysis. Transformation is especially helpful in

comparison because often the goal is to compare ordinally, as opposed to numerically (e.g. document
 48

A contains more pronouns than document B). Typically, a Box-Cox transformation, which is a

𝑥 𝜆 −1
general class of power-based transformation function ( 𝑥 ′ = ) can reduce a variable’s
𝜆

distribution skewness. One easy transformation is arbitrarily setting  = 0, which is equivalent to

taking the logarithmic of the variable for any x greater than 0 (Box and Cox 1964; Osborne 2010).

To further account for the overall frequency of words in the text, researchers will also often

transform the word or term frequency matrix into a normalized measure such as the percent of words

in the unit (Kern et al. 2016; Pennebaker and King 1999) or a Term-Frequency Inverse Document

Frequency (tf-idf) (Spärck Jones 1972).

Common words may not be very diagnostic, and so researchers will often want to weight rare

words more heavily because they are more predictive (Netzer et al 2012). To address this, tf-idf

accounts for the total frequency of a word in the dataset. Specifically, one definition of tf-idf is:

𝑡𝑓 − 𝑖𝑑𝑓 (word 𝑤, document 𝑑, corpus 𝐷) =

Total Number of Documents in 𝐷

[1 + log(Number of Occurrences of 𝑤 in 𝑑)] × log( ),
Number of Documents Containing 𝑤

If number of occurrences of a word is 0, then it is set to 0 (Manning and Hinrich 1999). After

calculating the tf-idf for all keywords in every document, the resulting matrix is used as measures of

(weighted) frequency for statistical comparison. This method gives an extra boost to rare word

occurrences in an otherwise sparse matrix, and as such, statistical comparisons can leverage the

additional variability for testing the hypotheses.

Tf-idf is useful for correcting for infrequently occurring words, but there are other methods

one may want to use to compare differences in frequently occurring words like function words. For

example, Monroe et al. (2009) compare speeches from Republican and Democratic candidates. In

this context, eliminating all function words may lead to misleading results because a function word

like “she” or “her” can be indicative of Democratic Party’s policies on women’s rights. Specifically,
 49

Monroe et al. (2009) first observe distribution of word occurrences in their entire dataset of Senate

speeches that address a wide range of topics to form a prior that benchmarks how often a word

should occur. Then they combine the log-odds-ratio method with that prior belief to examine the

differences between Republican and Democrat speeches on the topic of abortion. Such methods that

incorporate priors account for frequently occurring words and thus complement tf-idf.

Correlation

Co-occurrence helps scholars see patterns of association that may not be otherwise observed,

either between textual elements or between textual elements and non-textual elements such as survey

responses or ratings. Reporting correlations between textual elements is often used as a preliminary

analysis before further comparison either between groups or over time in order to gain a sense of

discriminant and convergent validity (e.g. Markowitz and Hancock 2015; Humphreys 2010). For

example, to study lying Markowitz and Hancock (2015) create an “obfuscation index” that is

composed of multiple measures with notable correlations including jargon, abstraction (positively

indexed), positive emotion and readability (negatively indexed) and find that these combinations of

linguistic markers are indicators of deception. In this way, correlations are used to build higher order

measures or factors such as linguistic style (Ludwig et al. 2013; Pennebaker and King 1999).

When considered on two or more dimensions, co-occurrence between words takes on new

meaning as relationships between textual elements can be mapped. These kinds of spatial approaches

can include network analysis, where researchers use measures like centrality to understand the

importance of some concepts in linking a conceptual network (e.g. Carley 1997) or to spot structural

holes where concepts may be needed to link concepts. For example, Netzer et al. (2012) study

associative networks for different brands using message board discussion of cars, based on co-

occurrence of car brands within a particular post. Studying correlation between textual elements
 50

gives researchers insights about semantic relationships that may co-occur and thus be linked in

personal or cultural associations. For example, Neuman et al. (2012) use similarity scores to

understand metaphorical associations for the words sweet and dark, as they’re related to other, more

abstract words and concepts (e.g. sweetness and darkness).

In addition to using correlations between textual elements in research design, researchers will

often look at correlations between linguistic and non-linguistic elements, on the way to forming

predictions. For example, Brockmeyer et al. (2015) study correlations between pronoun use and

patient reported depression and anxiety, finding that depressed patients use more self-focused

language when recalling a negative memory. Ireland et al. (2011) observe correlations between

linguistic style and romantic attachment and use this as support for the hypothesis of linguistic style

matching.

Issues with Correlation. Well-designed correlation analysis requires a series of robustness

checks, i.e., performing similar or related analyses using alternative methodologies to ensure

results from these latter analyses are congruent with the initial findings. Some of the robustness

checks include: 1) using a random subset of the data and repeating the analyses, 2) examining or

checking for any possible effects due to heterogeneity, and 3) running additional correlation

analyses using various types of similarity measures such as lift, Jaccard distance, cosine distance,

tf-idf co-occurrence, Pearson correlation (Netzer et al. 2012), Euclidean distance, Manhattan

distance, and edit distance.

Generally speaking, results should be congruent regardless of which subset of data or

distance measure is used. However, some distance measures may inherently be more appropriate

than the others, depending on the underlying assumption the distance represents. Netzer et al. (2012)

provides an instructive example of robustness check within the context of mapping automobile

brands to product attributes. Using and comparing multiple methods of similarity, they find that
 51

using Jaccard, cosine, and tf-idf co-occurrence distance measures yield similar results as their

original findings. Pearson correlation, on the other hand, yields less meaningful results due to

sparseness of the data.

It is also important to note that the interpretation of co-occurrence as a measure of

correlation can be biased toward frequent words—words that occur in more documents may

inherently co-occur more frequently than other words. As such, methods such as z-scores or

simple co-occurrence counts may be inappropriate, and extant literature suggests normalizing the

occurrence counts by calculating lift or point-wise mutual information (PMI) using relative

frequencies of occurrences, for example (Netzer et al. 2012). However, one criticism against

mutual information type measurements is that, particularly in smaller datasets, they may

overcorrect for word frequency and thus bias the analysis toward rare words. In these cases, log

likelihood test provides a balance “between saliency and frequency” (Pollach 2012, p.8).

Another issue that arises in correlation, particularly with a large number of categories, is

that many correlations, not all of them theoretically meaningful. To account for the presence of

multiple significant correlations, some of which may be spurious or due to chance, Kern et al

(2016) suggest calculating Bonferroni corrected p-values and including only correlations with

small p-values (e.g. p<.001).

Prediction

Prediction using text analysis usually goes beyond correlational analysis in that it takes other

non-textual variables into account. For example, Ludwig et al. (2016) use elements of email text like

flattery and linguistic style matching to predict deception, where they have a group of known

deceptions. In examining Kiva loan proposals, Genevsky and Knutson (2015) operationalize affect

with percentages of positive and negative words, and they then incorporate these two variables as

independent variables in a linear regression to predict lending rates.

 52

In other contexts, researchers may have access to readily available data such as ratings, likes,

or some other variable to corroborate their prediction and incorporate this information into the model.

Textual characteristics can also be used as predictors of other content elements, particularly in

answering empirical questions. Using a dataset from a clothing store, Anderson and Simester (2014)

identify a set of product reviews that are written by 12,000 “users” who did not seem to have

purchased the products. Using logistic and ordinary least square (OLS) models, they then find that

textual characteristics such as word count, average word length, occurrences of exclamation marks,

and customer ratings predict whether a review is “fake,” controlling for other factors.

Issues with Prediction. When using textual variables for prediction, researchers should

recognize endogeneity due to selection bias, omitted variable bias, and heterogeneity issues. As

previously discussed, samples of text can be biased in various ways and therefore may not generalize

if the sample differs markedly from the population.

When analyzing observational data such as tweets or review posts, a researcher almost

certainly encounters selection bias because the text is not generated by a random sample of the

population nor is a random set of utterances. For instance, reviewers may decide to post their

negative opinions online when they see positive reviews that go against their perspective (Sun 2012).

If a researcher wants to discover consumer sentiment toward a smartphone from CNET, for example,

she may need to consider when and how the reviews are generated in the first place. Are they posted

right after a product has been launched or months afterwards? Are they written when the brand is

undergoing a scandal? By identifying possible external shocks that may cause a consumer to act in a

certain way, a researcher can compare the behaviors before and after the shock to examine the

effects. Combining these contexts with methodological frameworks such as regression discontinuity

(i.e., comparing responses right before and after the treatment) or matching (i.e., a method that

creates a pseudo-control group using observational data) may reduce some of the biases. Future
 53

research using controlled lab experiments or field studies to predict hypothesized changes in written

text can further bolster confidence in using text to measure certain constructs.

Overfitting is another common problem with prediction in text analysis. Because there are

often many independent variables (i.e. words or categories) relative to the number of observations,

results can be overly specific to the data or training set. Kern et al (2016) have suggestions for

addressing the issue by reducing the number of predictors such as applying principle component

analysis (PCA) to the predictors and k-fold cross-validation on hold out sample(s). In general,

developing and reducing a model on a training set and then testing on a sufficient hold-out sample

can increase generalizability and reduce problems with overfitting.

Stage 6: Validation

Automated text analysis, like any method, has strengths and weaknesses. While lab studies

may be able to achieve internal validity in that they can control for a host of alternative factors in a

lab setting, they are, of course, somewhat weaker on external validity (Cook et al. 1979). Automated

text analysis, on the other hand, lends researchers claim to external validity, and particularly

ecological validity, as the data is observed in organically-produced consumer texts (Mogilner et al.

2011). Beyond this, other types of validity such as construct, concurrent, discriminant, convergent,

and predictive validity are addressable using a variety of techniques (McKenny, Short, and Payne

2013).

Construct validity can be addressed in a number of ways. Because text analysis is relatively

new for measuring social and psychological constructs, it is important to be sure that constructs are

operationalized in ways consistent with their conceptual meaning and previous theorization. Through

dictionary development, one can have experts or human coders evaluate wordlists for their construct

validity in pretests. More elaborately, pretests of the dictionary using a larger sample or survey could
 54

also help ensure construct validity (Kovács et al. 2013). Using an iterative approach, one can equally

pull coded instances from the data to ensure that operationalization through the dictionary words

makes sense (for this, Weber 2005 suggests using a saturation procedure to reach 80% accuracy in a

training set). In classification, the selection or coding of training data is another place to address

construct validity. For example, does the text pulled and attributed to brand loyalists actually

represent loyalty? One can use external validation or human ratings for calibration. For example,

Jurafsky et al. (2009) use human ratings of awkwardness, flirtation, etc. to classify the training data.

Convergent validity, the degree to which measures of the construct correlate to each other,

can be assessed by measuring the construct using different linguistic aspects, and by comparing

linguistic analysis with measurements external to text. For example, construal level could be

measured using a semantics-based dictionary (Snefjella and Kuperman 2015) or through pragmatic

markers available through LIWC. Beyond convergent validity in any particular study, concurrent

validity, the ability to draw inferences over many studies, is improved when researchers use standard,

previously used, and thoroughly tested dictionaries. This allows researchers to draw conclusions

across studies, knowing that constructs have been measured with the same list of words. Bottom up,

classificatory analysis does not afford researchers the same assurance.

Discriminant and convergent validity are relatively easy to assess after conducting the text

analysis through factor analysis. Here, bottom-up methods of classification and similarity are

invaluable for measuring the likeness of groups of texts and placing this likeness on more than one

dimension. Researchers can then observe consistent patterns of difference to ascertain discriminant

validity.

Predictive validity, the ability of the constructs measured via text to predict other constructs

in the nomological net, is perhaps one of the most important types of validity to establish the

usefulness of text analysis in social science. Studies have found relationships between language and
 55

stock price (Tirunillai and Tellis 2012), personality type (Pennebaker and King 1999), and box office

success (Eliashberg, Hui, and Zhang 2007). A hold-out sample can be helpful in investigating

whether the hypothesized model is generalizable to new data. There are a variety of ways to do

hold-out sampling and validation such as k-fold cross validation, which splits the dataset into k

parts, and for each iteration, uses k–1 subset for training, and one subset for testing. The process

is iterated until each part has been used as a test sub-set. For instance, Jurafsky et al. (2014) hold

out 20% of the sample for testing; van Laer et al. (2017) save about 10% for testing. The

accuracy rate should be greater than the no-information rate, and it should also be relatively

consistent across all iterations.

Further validation depends on the particular method of statistical analysis used. For

comparison, multiple operationalization using different measures (linguistic and non-linguistic)

can help support the results. If using correlation, triangulation can be accomplished by looking at

correlations in words or categories that one would expect (see e.g. Humphreys 2010; Pennebaker and

King 1999),

Lastly, text analysis conducted with many categories on a large dataset can potentially

yield many possible correlations and many statistically significant comparisons, some of which

may not be actionable, and some of which may even be spurious. For research designs that use

hypothesis testing, Bonferroni-corrected p-values can be used where there is the possibility of

spurious correlation from testing multiple hypotheses (Kern et al 2016). However, some argue

that the test is too stringent and offer other alternatives (Benjamini and Hochberg 1995). While

text analysis provides ample information, giving meaning to it requires theory. Without theory,

findings can be too broad and relatively unexplained, and sheer computing power is never a

replacement for asking the right type of research questions, framing the right type of constructs,
 56

collecting and merging the right types of dataset(s), or choosing the right operationalization

approach (Crawford, Miltner, and Gray 2014).

As we demonstrate, choosing the most appropriate method depends on those high-level

thought processes, which cannot be performed by computers or artificial intelligence alone.

Designing the right type of top-down research requires carefully operationalized constructs and

implementation, and analyzing results and predictions from bottom-up learning requires

interpretation, both of which rely on theory and expertise in a knowledge domain. In short, while

datasets and computing power are as abundant as ever, one still does not gain insight without a

clear theoretical framework. Only through repeated testing through multiple operationalizations

can we separate the spurious from systematic findings.

Ethics

Because the data for text analysis often comes from the internet rather than traditional

print sources, the ethics of collecting, storing, analyzing, and presenting findings from such data

are critical issues to consider and yet still somewhat in flux. Although it can seem

depersonalized, text data comes from humans, and per the Belmont Report (1978), researchers

should minimize harm to these individuals, being mindful of the respect, beneficence and justice

to those who provide underlying data for research. The association of Internet Researchers

provides an overview of ethical considerations when conducting internet research that usefully

apply to most text analyses (Markham and Buchanan 2012), and Townsend and Wallace (2016,

p. 8) provide succinct guidelines for the ethical decisions one faces in conducting social media

research. In general, these guidelines advocate for a case-based approach informed by the

context in which the data exists and is collected, analyzed, and circulated. While few
 57

organizations offer straightforward rules, we find three issues deserve particular consideration

when conducting textual analysis.

The first ethical question is one of access and jurisdiction—do researchers have

legitimate jurisdiction to collect the textual data of interest? Here, the primary concern is the

boundary between public and private information. Given the criteria laid out by the Common

Rule, internet discourse is usually deemed to be public because individuals cannot “reasonabl[y]

expect that no observation or recording is taking place” [45 CFR 46.102(f)]). Summarizing a

report from the US Department of Health and Human Services, Hayden (Hayden 2013)(2013)

says, “The guidelines also suggest that, in general, information on the Internet should be

considered public, and thus not subject to IRB review — even if people falsely assume that the

information is anonymous,” (p. 411). However, technical and socially constructed barriers such

as password protected groups, gatekeepers, and interpersonal understandings of trust also define

participants’ expectations of privacy (Nissenbaum 2009; Townsend and Wallace 2016;

Whiteman 2012). Guidelines therefore also suggest that “investigators should note expressed

norms or requests in a virtual space, which – although not technically binding – still ought to be

taken into consideration.” (HHS 2013, p. 5). The boundary between public and private

information is not always clear, and researcher judgements should be made based on what

participants themselves would expect based on a principle of “contextual integrity” (Marx 2001;

Nissenbaum 2009).

A second question concerns the control, storage, and presentation of textual data. There’s

an important distinction between the ethics of the treatment of aggregated verses individualized

data, with aggregated data being less sensitive than individualized data and individualized data of

vulnerable populations being the most sensitive. Identification such as screenname should be

deleted for those who are not public figures (Townsend and Wallace 2016). Even if anonymized,
 58

individualized textual data is often searchable, and therefore attributable to the original source,

even if the name or other identifying information is removed. For this reason, when presenting

individual excerpts, some researchers will choose to remove words or paraphrase in order to

reduce searchability (Townsend and Wallace 2016).

The durability of textual data also means that comments made by consumers long ago

may be damaging if discovered and circulated. With large enough sample sizes, aggregated data

is less vulnerable to identifiability, although cases in which there are extreme outliers or sparse

data may be identifiable and researchers should take care when presenting these cases. Further,

even in large datasets, researchers have demonstrated that anonymized data can be made

identifiable using meta-data such as location, time, or other variables (Narayanan and Shmatikov

2008, 2010) so data security is of primary importance. Current guidelines suggest that text data

be treated as any human subject data, under password protection, and, when possible, de-

identified (Markham and Buchanan 2012).

A final matter is legitimate ownership or control of the data. Many terms of service

(ToS) agreements prohibit scraping of their data, and while some offer APIs to facilitate paid or

metered access, others do not. While the legalities of ownership are relatively clear, albeit legally

untested, some Communications researchers have argued that control over this data constitutes

an unreasonable obstacle to research that is in the public interest (Waters 2011). Contemporary

legal configurations also mean that consumers who produce the data may not themselves have

access to or control of it. For this reason, it’s important that researchers make efforts to share

research with the population from which the data was used. For many researchers, this also

means clearing permission with the service provider to use data, although requirements for this

vary depending on field and journal.

CONCLUSION
 59

Summary

In this article, we contribute to the growing need for methodologies to incorporate

analysis of textual data in consumer research. Although there are a considerable set of

phenomena that cannot be measured by text, computers can be used to discover, measure, and

represent patterns of language that elude both researchers and consumers. Our goal has been to

provide insight into the relationship between language and consumer thought, interaction, and

culture to inspire research that uses language and then to further provide a roadmap for executing

text-based research help researchers make decisions inherent to any text analysis.

We propose that automated text analysis can be a valuable methodology for filling the gap

between the data available and theories commonly used in consumer research. And while previous

work in psychology has developed a variety of stand-alone approaches advocated by particular

researchers, it has not incorporated linguistic theory into constructs that are meaningful to consumer

researchers, provided guidance about what method to use, discussed when a particular method might be

appropriate, nor tailored approaches to the unique perspective of consumer researchers.

We therefore make three contributions to consumer research in light of previous

treatments of text analysis in the social sciences. First, we provide an overview of methods

available, guidance for choosing an approach, and a roadmap for making analytical decisions.

Second, we address in depth the methodological issues unique to studies of text such as sampling

considerations when using internet data, approaches to dictionary development and validation

and statistical issues such as dealing with sparse data and normalizing textual data. Lastly, and

more broadly, we hope to contribute to growing incorporation of communications and linguistic

theory into consumer research and provide a tool for linking multiple levels of analysis

(individual, group, and cultural dimensions) so crucial to a multi-disciplinary field such as

consumer research.
 60

The focus of this article has been written text. However, visual data such as video contains

tonality and emotion of words as well as visual information such as facial expression and gesture in

addition to textual content. Video data therefore have the potential to provide additional measurement

of constructs and may require more sophisticated techniques and analyses (see e.g. Cristani et al.

2013; Jain and Li 2011; Lewinski 2015). Although beyond the scope of our methodology, some

procedures we discuss here may still apply. For example, if studying gesture, one will need to define

a “dictionary” of gestures that represent a construct or to code all gestures and then group them into

meaningful categories.

Future Directions

Overall, where will text analysis lead us? Consumers are more surrounded by and

produce more textual communication than ever before. Developing methodological approaches

that can incorporate textual data into traditional analyses can help consumer researchers

understand the influence of language and to use language as a measure of consumer thought,

interaction, and culture. Equally, methods of social science are changing as capacities for

collecting, storing, and analyzing both textual and non-textual (e.g. behavioral) data expand. In a

competitive landscape of academic disciplines, it makes sense that consumer research should

incorporate some of this data, as it is useful to exploring the theories and questions driving the

field. As we have argued, questions of attention, processing, interaction, groups, and culture can

all be informed by text analysis.

Text analysis also supports the interdisciplinary nature of consumer research in two ways.

First, it further points to theories in linguistics and communications that inform questions

common to marketplace interaction. There is a growing acknowledgement of the role of

language and communication theory in social influence and consumer life (e.g. Moore 2015;
 61

Packard and Berger 2016, van Laer et al. 2017; Barasch and Berger 2014), and text analysis

methodologically supports the inclusion of these types of theories into consumer research.

Second, text analysis can link different levels of analysis, which is particularly important

in a field that incorporates different theoretical traditions. Studying the self, for example, requires

understanding not only individual psychological processes like cognitive dissonance (Festinger

1962), but also an awareness of how the self interacts with objects (Belk 1988; Kleine, Kleine,

and Allen 1995) and cultural meanings surrounding these objects, meanings that change due to

shifts in culture (Luedicke, Thompson, and Giesler 2010). Similarly, studying social influence

can be informed by psychological theories of power (Ng and Bradac 1993) and processing (Petty

et al. 1983), but also normative communication and influence (McCracken 1986; McQuarrie,

Miller, and Phillips 2013). As Deighton (2007) argues, consumer research is distinct from core

disciplines in that, although it theoretically based, it is also problem oriented in the sense that

researchers are interested in the middle ground between abstract theory and instrumental

implication. Text analysis can lend ecological validity to rigorously conducted lab studies that

illustrate a causal effect or can point toward new discoveries that can be further investigated.

Our intention is not to suggest that text analysis can stand alone, but rather that it is a

valuable companion for producing insight in an increasingly textual world. Social science

research methods are changing, and while laboratory experiments are considered the gold standard

for making causal inference, other forms of data can be used to make discoveries and show the

import of theoretical relationships in consumer life. If we limit ourselves to data that can be gathered

and analyzed in a lab, we are discarding zettabytes of data in today’s digital age, including millions

of messages transmitted per minute online (Marr 2015). And yet automated text analysis is a relatively

new method in the social sciences. It will likely change over the next decade due to the advances in

computational linguistics, the increasing availability of digital text, and interest amongst marketing
 62

professionals. Our aim has been to provide a link between linguistic elements in text and constructs in

consumer behavior and guidance for executing research using textual data. In a world where

consumer texts grow more numerous each day, automated text analysis, if done correctly, can yield

valuable insights about consumer attitudes, interaction, and culture.

 63

APPENDIX

Research Question Text Linguistic Aspect Source

Dictionary-Based - Comparison

How does temporal and spatial distance

affect emotions after a tragic event? Twitter semantic Dore et al. 2015
How do power and affiliation vary by transcripts (chatrooms, State of the
political ideology? Union), news websites semantic Fetterman et al. 2015
What explains representational gender bias
in the media? newspapers phatic Shor et al. 2015

How does personal pronoun use in firm-

customer interactions impact customer Packard, Moore, and McFerran
attitude? transcripts pragmatic 2016
Why don’t major crises like oil spills provoke
broad changes in public discourse
concerning the systemic
risks inherent to a carbon-dependent
economy? newspaper articles semantic Humphreys and Thompson 2014

Do people modify warmth to appear

competent (and v.v.) when doing
impression management? emails semantic Holoien and Fiske 2014
Does social hierarchy affect language use?
In what ways? emails pragmatic Kacewicz et al. 2014
Do Christians and atheists vary in their
language use? Twitter semantic Ritter et al 2013

How does someone’s communication style

change based on private versus public Facebook wall posts and private
communication? messages semantic, pragmatic Bazarova et al. 2012
 64

How do letters to shareholders differ in a

period of economic growth versus
recession? letters to shareholders semantic Pollach 2012
Are people with the same linguistic style
more likely to form a romantic relationship? transcripts, instant messages stylistic, pragmatic Ireland et al 2011
How does happiness change throughout
the lifecycle? Personal blogs semantic Mogilner et al 2011
Dictionary-Based - Correlation

Do depressed patients use more self-

focused language? written essays semantic Brockmeyer et al 2015

Is construal level (physical, temporal, and

social) represented in language and what
are the mathematical properties of concrete
vs. abstract? Twitter semantic, pragmatic Snefjella Kuperman 2015
What neural affective mechanisms prompt
altruism? loan requests (Kiva) semantic Genevsky and Knutson 2014
Does audience size affect the way people Lab experiment (fake emails;
share? conversations; elaborations) semantic, pragmatic Barasch and Berger 2014
Is fame stable or fleeting? newspapers, blogs, television semantic, phatic Rijt et al 2013
How does disease advocacy shape medical
policy? transcripts semantic Best 2012

How did Alan Greenspan’s language

change during periods of economic
expansion versus decline? transcripts semantic Abe 2011
How are markets created? newspaper articles semantic Humphreys 2010
Does language use reflect personality? written essays, journal articles stylistic, pragmatic Pennebaker and King 1999
Dictionary-Based - Prediction

How do cultural artifacts affect social and Internet searches, newspapers,

political change? Twitter semantic Vasi et al 2015
 65

Does affect and linguistic style matching in

online reviews lead to purchase? Book reviews semantic; stylistic Ludwig et al. 2013

How does cognitive complexity relate to the

relationship between CEO facial structure
and performance? (Secondary RQ) letters to shareholders semantic, syntax Wong et al 2012
What makes an article go viral? NYT Articles and top emailed lists semantic Berger and Milkman 2012
Does syntax influence ad recognition or
comprehension? Lab experiment syntax / semantic Bradley and Meeds 2002
Classification - Comparison

Did Shakespeare write this rediscovered

play? plays stylistic, pragmatic Boyd and Pennebaker 2015
How do consumers frame positive and
negative reviews online? Yelp restaurant reviews semantic Jurafsky et al. 2016

What’s the process through which a

concrete word becomes an abstract
concept (“hypostatic abstraction)? text from website semantic, pragmatic Neuman et al. 2012
Classification - Correlation

How to design a ranking system by mining

UGC? Online reviews semantic, stylistic Ghose, Ipeirotis, and Li 2012
How does sentiment change over time
during an election? blogs semantic Hopkins and King 2012

What word associations exist in the poems

of Emily Dickensen that have gone
unnoticed by experts? poems semantic Plaisant et al 2006
Classification - Prediction

How do customers react to actively Field experiments as well as

managed online communities? observational online forum data semantic Homburg, Ehm, and Artz 2015
Which movies generate the most ROI? IMDB movie summaries semantic Eliashberg, Hui, and Zhang (2007)
Can a firm predict a customer’s churn using Coussement, K., & Van den Poel, D.
their complaints? Call center emails semantic 2008
 66

Reviews from Amazon.com,

Epinions.com and Yahoo!Shopping,
Can UGC predict stock performance? 6 markets and 15 firms semantic Tirunillai and Tellis 2012
Topic Discovery - All

Can a brand track its perceived quality

using online reviews? same as Tirunillai and Tellis 2012 semantic Tirunillai and Tellis 2014
What do consumers talk about in hotel
reviews? Hotel review data semantic Mankad et al. 2016

What associative networks do consumers

have for different brands? / How to
construct a perceptual map using web-
available text data? forums semantic Netzer et al 2012
How to analyze market structure using text Online reviews; pros and cons lists
mining of UGC? (explicitly stated by the reviewers) semantic Lee and Bradlow 2011
Does analyzing text data by sentence rather Online reviews from Expedia and
than word improve prediction of sentiment? we8there.com Semantic/syntax Büschken and Allenby 2016
 67

REFERENCES

Alexa, Melina (1997), Computer Assisted Text Analysis Methodology in the Social Sciences:

ZUMA.

Allen, Douglas E (2002), “Toward a Theory of Consumer Choice as Sociohistorically Shaped

Practical Experience: The Fits-Like-a-Glove (Flag) Framework,” Journal of Consumer

Research, 28 (4), 515-32.

Anderson, Eric T and Duncan I Simester (2014), “Reviews without a Purchase: Low Ratings,

Loyal Customers, and Deception,” Journal of Marketing Research, 51 (3), 249-69.

Arnold, Stephen J and Eileen Fischer (1994), “Hermeneutics and Consumer Research,” Journal

of Consumer Research, 21 (1), 55-70.

Arsel, Zeynep and Jonathan Bean (2013), “Taste Regimes and Market-Mediated Practice,”

Journal of Consumer Research, 39 (5), 899-917.

Arsel, Zeynep and Craig J Thompson (2011), “Demythologizing Consumption Practices: How

Consumers Protect Their Field-Dependent Identity Investments from Devaluing

Marketplace Myths,” Journal of Consumer Research, 37 (5), 791-806.

Arvidsson, Adam and Alessandro Caliandro (2016), “Brand Public,” Journal of Consumer

Research, 42 (5), 727-48.

Back, Mitja D, Albrecht CP Küfner, and Boris Egloff (2010), “The Emotional Timeline of

September 11, 2001,” Psychological Science, 21 (10), 1417-19.

--- (2011), ““Automatic or the People?” Anger on September 11, 2001, and Lessons Learned for

the Analysis of Large Digital Data Sets,” Psychological Science, 22 (6), 837-38.
 68

Barasch, Alixandra and Jonah Berger (2014), “Broadcasting and Narrowcasting: How Audience

Size Affects What People Share,” Journal of Marketing Research, 51 (3), 286-99.

Belk, Russell W (1988), “Property, Persons, and Extended Sense of Self,” Proceedings of the

Division of Consumer Psychology, American Psychological Association 1987Annual

Convention, 28-33.

Belk, Russell W, John F Sherry, and Melanie Wallendorf (1988), “A Naturalistic Inquiry into

Buyer and Seller Behavior at a Swap Meet,” Journal of Consumer Research, 14 (4),

449-70.

Bell, Gordon, Tony Hey, and Alex Szalay (2009), “Beyond the Data Deluge,” Science, 323

(5919), 1297-98.

Benford, Robert D. and David A. Snow (2000), “Framing Processes and Social Movements: An

Overview and Assessment,” Annual Review of Sociology, 26 (1), 611-40.

Benjamini, Yoav and Yosef Hochberg (1995), “Controlling the False Discovery Rate: A

Practical and Powerful Approach to Multiple Testing,” Journal of the royal statistical

society. Series B (Methodological), 289-300.

Berger, Jonah and Katherine L Milkman (2012), “What Makes Online Content Viral?,” Journal

of Marketing Research, 49 (2), 192-205.

Bollen, Johan, Huina Mao, and Xiaojun Zeng (2011), “Twitter Mood Predicts the Stock Market,”

Journal of Computational Science, 2 (1), 1-8.

Borgman, Christine L (2015), Big Data, Little Data, No Data: Scholarship in the Networked

World: Mit Press.

Boroditsky, Lera (2001), “Does Language Shape Thought?: Mandarin and English Speakers’

Conceptions of Time,” Cognitive psychology, 43 (1), 1-22.

 69

Boroditsky, Lera, Lauren A Schmidt, and Webb Phillips (2003), “Sex, Syntax, and Semantics,”

Language in mind: Advances in the study of language and thought, 61-79.

Boyd, Ryan L and James W Pennebaker (2015), “Did Shakespeare Write Double Falsehood?

Identifying Individuals by Creating Psychological Signatures with Text Analysis,”

Psychological science, 0956797614566658.

Bradley, Samuel D and Robert Meeds (2002), “Surface‐Structure Transformations and

Advertising Slogans: The Case for Moderate Syntactic Complexity,” Psychology &

Marketing, 19 (7‐8), 595-619.

Brier, Alan and Bruno Hopp (2011), “Computer Assisted Text Analysis in the Social Sciences,”

Quality & Quantity, 45 (1), 103-28.

Brockmeyer, Timo, Johannes Zimmermann, Dominika Kulessa, Martin Hautzinger, Hinrich

Bents, Hans-Christoph Friederich, Wolfgang Herzog, and Matthias Backenstrass (2015),

“Me, Myself, and I: Self-Referent Word Use as an Indicator of Self-Focused Attention

in Relation to Depression and Anxiety,” Frontiers in psychology, 6.

Carley, Kathleen (1997), “Network Text Analysis: The Network Position of Concepts,” in Text

Analysis for the Social Sciences: Methods for Drawing Statistical Inferences from Texts

and Transcripts, ed. Carl W Roberts, Mahwah, NJ: Lawrence Erlbaum.

Carpenter, Christopher J and David Dryden Henningsen (2011), “The Effects of Passive Verb-

Constructed Arguments on Persuasion,” Communication Research Reports, 28 (1), 52-

61.

Chambers, Nathanael and Dan Jurafsky (2009), “Unsupervised Learning of Narrative Schemas

and Their Participants,” in Proceedings of the Joint Conference of the 47th Annual

Meeting of the ACL and the 4th International Joint Conference on Natural Language
 70

Processing of the AFNLP: Volume 2-Volume 2: Association for Computational

Linguistics, 602-10.

Chartrand, Tanya L and John A Bargh (1996), “Automatic Activation of Impression Formation

and Memorization Goals: Nonconscious Goal Priming Reproduces Effects of Explicit

Task Instructions,” Journal of Personality and Social Psychology, 71 (3), 464.

Chartrand, Tanya L, Joel Huber, Baba Shiv, and Robin J Tanner (2008), “Nonconscious Goals

and Consumer Choice,” Journal of Consumer Research, 35 (2), 189-201.

Chomsky, Noam (1957/2002), Syntactic Structures: Walter de Gruyter.

Chung, Cindy K and J. W. Pennebaker (2013), “Counting Little Words in Big Data,” Social

cognition and communication, 25.

Churchill, Gilbert A (1979), “A Paradigm for Developing Better Measures of Marketing

Constructs,” Journal of Marketing Research, 64-73.

Collins, Allan M and Elizabeth F Loftus (1975), “A Spreading-Activation Theory of Semantic

Processing,” Psychological review, 82 (6), 407.

Conrad, Susan (2002), “4. Corpus Linguistic Approaches for Discourse Analysis,” Annual

Review of Applied Linguistics, 22, 75-95.

Cook, Thomas D, Donald Thomas Campbell, and Arles Day (1979), Quasi-Experimentation:

Design & Analysis Issues for Field Settings, Vol. 351: Houghton Mifflin Boston.

Corbin, Juliet M. and Anselm L. Strauss (2008), Basics of Qualitative Research : Techniques

and Procedures for Developing Grounded Theory, Los Angeles, Calif.: Sage

Publications, Inc.
 71

Cristani, Marco, R. Raghavendra, Alessio Del Bue, and Vittorio Murino (2013), “Human

Behavior Analysis in Video Surveillance: A Social Signal Processing Perspective,”

Neurocomputing, 100, 86-97.

Daku, Mark, Lori Young, and Stuart Soroka (2011), “Lexicoder, Version 3.0.”

Danescu-Niculescu-Mizil, Cristian, Justin Cheng, Jon Kleinberg, and Lillian Lee (2012a), “You

Had Me at Hello: How Phrasing Affects Memorability,” in Proceedings of the 50th

Annual Meeting of the Association for Computational Linguistics: Long Papers-Volume

1: Association for Computational Linguistics, 892-901.

Danescu-Niculescu-Mizil, Cristian, Lillian Lee, Bo Pang, and Jon Kleinberg (2012b), “Echoes of

Power: Language Effects and Power Differences in Social Interaction,” in Proceedings

of the 21st international conference on World Wide Web: ACM, 699-708.

Danescu-Niculescu-Mizil, Cristian, Moritz Sudhof, Dan Jurafsky, Jure Leskovec, and

Christopher Potts (2013), “A Computational Approach to Politeness with Application to

Social Factors,” arXiv preprint arXiv:1306.6078.

De Choudhury, Munmun, Hari Sundaram, Ajita John, and Dorée Duncan Seligmann (2008),

“Can Blog Communication Dynamics Be Correlated with Stock Market Activity?,” in

Proceedings of the nineteenth ACM conference on Hypertext and hypermedia: ACM,

55-60.

Deighton, John (2007), “From the Editor: the Territory of Consumer Research: Walking the

Fences,” Journal of Consumer Research, 34 (3), 279-82.

DeWall, C Nathan, Richard S Pond Jr, W Keith Campbell, and Jean M Twenge (2011), “Tuning

in to Psychological Change: Linguistic Markers of Psychological Traits and Emotions

 72

over Time in Popular Us Song Lyrics,” Psychology of Aesthetics, Creativity, and the

Arts, 5 (3), 200.

Doré, Bruce, Leonard Ort, Ofir Braverman, and Kevin N Ochsner (2015), “Sadness Shifts to

Anxiety over Time and Distance from the National Tragedy in Newtown, Connecticut,”

Psychological science, 26 (4), 363-73.

Duggan, Maeve; Brenner, Joanna (2013), “The Demographics of Social Media Users -- 2012,”

Pew Research Center.

Dunphy, D. M., C.G. Bullard, and E.E.M. Crossing (1974), “Validation of the General Inquirer

Harvard Iv Dictionary,” in 1974 Pisa Conference on Content Analysis, Pisa, Italy.

Earl, Jennifer, Andrew Martin, John D McCarthy, and Sarah A Soule (2004), “The Use of

Newspaper Data in the Study of Collective Action,” Annual Review of Sociology, 65-80.

Eichstaedt, Johannes C, Hansen Andrew Schwartz, Margaret L Kern, Gregory Park, Darwin R

Labarthe, Raina M Merchant, Sneha Jha, Megha Agrawal, Lukasz A Dziurzynski, and

Maarten Sap (2015), “Psychological Language on Twitter Predicts County-Level Heart

Disease Mortality,” Psychological science, 26 (2), 159-69.

Eliashberg, Jehoshua, Sam K Hui, and Z John Zhang (2007), “From Story Line to Box Office: A

New Approach for Green-Lighting Movie Scripts,” Management Science, 53 (6), 881-

93.

Ertimur, Burçak and Gokcen Coskuner-Balli (2015), “Navigating the Institutional Logics of

Markets: Implications for Strategic Brand Management,” Journal of Marketing, 79 (2),

40-61.
 73

Farghaly, Ali and Khaled Shaalan (2009), “Arabic Natural Language Processing: Challenges and

Solutions,” ACM Transactions on Asian Language Information Processing (TALIP), 8

(4), 14.

Festinger, Leon (1962), A Theory of Cognitive Dissonance, Vol. 2: Stanford university press.

Frege, Gottlob (1892/1948), “Sense and Reference,” The philosophical review, 209-30.

Fullwood, Michelle (2015), “Parsing Chinese Text with Stanford Nlp.”

Gamson, William A. and Andre Modigliani (1989), “Media Discourse and Public Opinion on

Nuclear Power: A Constructionist Approach,” American Journal of Sociology, 95 (1), 1-

37.

Gardner, Wendi L, Shira Gabriel, and Angela Y Lee (1999), ““I” Value Freedom, but “We”

Value Relationships: Self-Construal Priming Mirrors Cultural Differences in Judgment,”

Psychological science, 10 (4), 321-26.

Gibson, Edward (1998), “Linguistic Complexity: Locality of Syntactic Dependencies,”

Cognition, 68 (1), 1-76.

Goffman, Erving (1959), The Presentation of Self in Everyday Life, Garden City, N.Y.,:

Doubleday.

--- (1979), “Footing,” Semiotica, 25 (1-2), 1-30.

Graesser, Arthur C, Danielle S McNamara, Max M Louwerse, and Zhiqiang Cai (2004), “Coh-

Metrix: Analysis of Text on Cohesion and Language,” Behavior research methods,

instruments, & computers, 36 (2), 193-202.

Graham, Robert J (1981), “The Role of Perception of Time in Consumer Research,” Journal of

Consumer Research, 335-42.

 74

Grayson, Kent and David Shulman (2000), “Indexicality and the Verification Function of

Irreplaceable Possessions: A Semiotic Analysis,” Journal of Consumer Research, 27 (1),

17-30.

Green, Melanie C and Timothy C Brock (2002), “In the Mind’s Eye: Transportation-Imagery

Model of Narrative Persuasion.”

Grice, Herbert P (1970), Logic and Conversation: na.

Grimmer, Justin and Brandon M Stewart (2013), “Text as Data: The Promise and Pitfalls of

Automatic Content Analysis Methods for Political Texts,” Political Analysis.

Gruenfeld, Deborah H and Robert S Wyer (1992), “Semantics and Pragmatics of Social

Influence: How Affirmations and Denials Affect Beliefs in Referent Propositions,”

Journal of personality and social psychology, 62 (1), 38.

Gruhl, Daniel, Ramanathan Guha, Ravi Kumar, Jasmine Novak, and Andrew Tomkins (2005),

“The Predictive Power of Online Chatter,” in Proceedings of the eleventh ACM

SIGKDD international conference on Knowledge discovery in data mining: ACM, 78-

87.

Hall, Stuart (1980), “Encoding/Decoding,” Culture, media, language, 128-38.

Hayden, Erika Check (2013), “Guidance Issued for Us Internet Research: Institutional Review

Boards May Need to Take a Closer Look at Some Types of Online Research,” Nature,

496 (7446), 411-12.

Herring, Susan C (2000), “Gender Differences in Cmc: Findings and Implications,” Computer

Professionals for Social Responsibility Journal, 18 (1), 0.

--- (2003), “Gender and Power in Online Communication,” The handbook of language and

gender, 202-28.
 75

HHS (2013), “Considerations and Recommendations Concerning Internet Research and Human

Subjects Research Regulations, with Revisions,” (Final document, approved at SACHRP

meeting March 12-13, 2013),

https://www.hhs.gov/ohrp/sites/default/files/ohrp/sachrp/mtgings/2013%20March%20M

tg/internet_research.pdf.

Holt, Douglas B (2004), How Brands Become Icons: The Principles of Cultural Branding:

Harvard Business Press.

Holt, Douglas B and Craig J Thompson (2004), “Man-of-Action Heroes: The Pursuit of Heroic

Masculinity in Everyday Consumption,” Journal of Consumer Research, 31 (2), 425-40.

Hopkins, Daniel J and Gary King (2010), “A Method of Automated Nonparametric Content

Analysis for Social Science,” American Journal of Political Science, 54 (1), 229-47.

Hsu, Kean J., Kalina N. Babeva, Michelle C. Feng, Justin F. Hummer, and Gerald C. Davison

(2014), “Experimentally Induced Distraction Impacts Cognitive but Not Emotional

Processes in Think-Aloud Cognitive Assessment,” Frontiers in psychology, 5, 474.

Humphreys, Ashlee (2010), “Semiotic Structure and the Legitimation of Consumption Practices:

The Case of Casino Gambling,” Journal of Consumer Research, 37 (3), 490-510.

Humphreys, Ashlee and Kathryn A Latour (2013), “Framing the Game: Assessing the Impact of

Cultural Representations on Consumer Perceptions of Legitimacy,” Journal of

Consumer Research, 40 (4), 773-95.

Humphreys, Ashlee and Craig J Thompson (2014), “Branding Disaster: Reestablishing Trust

through the Ideological Containment of Systemic Risk Anxieties,” Journal of Consumer

Research, 41 (4), 877-910.

Jain, Anil K and Stan Z Li (2011), Handbook of Face Recognition: Springer.

 76

Jakobson, Roman (1960), “Closing Statement: Linguistics and Poetics,” Style in language, 350,

377.

Jameson, Fredric (2013), The Political Unconscious: Narrative as a Socially Symbolic Act:

Routledge.

Jurafsky, Dan, Victor Chahuneau, Bryan R Routledge, and Noah A Smith (2014), “Narrative

Framing of Consumer Sentiment in Online Restaurant Reviews,” First Monday, 19 (4).

Jurafsky, Dan, Rajesh Ranganath, and Dan McFarland (2009), “Extracting Social Meaning:

Identifying Interactional Style in Spoken Conversation,” in Proceedings of Human

Language Technologies: The 2009 Annual Conference of the North American Chapter

of the Association for Computational Linguistics: Association for Computational

Linguistics, 638-46.

Kassarjian, Harold H. (1977), “Content Analysis in Consumer Research,” Journal of Consumer

Research, 4 (1), 8-19.

Kay, Paul and Willett Kempton (1984), “What Is the Sapir-Whorf Hypothesis?,” American

Anthropologist, 65-79.

Kern, Margaret L, Gregory Park, Johannes C Eichstaedt, H Andrew Schwartz, Maarten Sap,

Laura K Smith, and Lyle H Ungar (2016), “Gaining Insights from Social Media

Language: Methodologies and Challenges.”

Kirschenbaum, Matthew G (2007), “The Remaking of Reading: Data Mining and the Digital

Humanities,” in The National Science Foundation Symposium on Next Generation of

Data Mining and Cyber-Enabled Discovery for Innovation, Baltimore, MD: Citeseer.
 77

Kleine, Susan Schultz, Robert E Kleine, and Chris T Allen (1995), “How Is a Possession “Me”

or “Not Me”? Characterizing Types and an Antecedent of Material Possession

Attachment,” Journal of Consumer Research, 22 (3), 327-43.

Kovács, Balázs, Glenn R Carroll, and David W Lehman (2013), “Authenticity and Consumer

Value Ratings: Empirical Tests from the Restaurant Domain,” Organization Science, 25

(2), 458-78.

Kranz, Peter (1970), “Content Analysis by Word Group,” Journal of Marketing Research, 7 (3),

377-80.

Krippendorff, Klaus (2004), Content Analysis : An Introduction to Its Methodology, Thousand

Oaks Calif: Sage.

Kronrod, Ann, Amir Grinstein, and Luc Wathieu (2012), “Go Green! Should Environmental

Messages Be So Assertive?,” Journal of Marketing, 76 (1), 95-102.

Kuncoro, Adhiguna, Miguel Ballesteros, Lingpeng Kong, Chris Dyer, Graham Neubig, and

Noah A Smith (2016), “What Do Recurrent Neural Network Grammars Learn About

Syntax?,” arXiv preprint arXiv:1611.05774.

Labroo, Aparna A and Angela Y Lee (2006), “Between Two Brands: A Goal Fluency Account

of Brand Evaluation,” Journal of Marketing Research, 43 (3), 374-85.

Lakoff, George (2014), The All New Don’t Think of an Elephant!: Know Your Values and Frame

the Debate: Chelsea Green Publishing.

Lakoff, George and Sam Ferguson (2015), “The Framing of Immigration.”

Lakoff, Robin (1973), “Language and Woman’s Place,” Language in society, 2 (01), 45-79.

Lasswell, Harold D. and Nathan Leites (1949), Language of Politics; Studies in Quantitative

Semantics, New York,: G. W. Stewart.

 78

Lee, Angela Y and Jennifer L Aaker (2004), “Bringing the Frame into Focus: The Influence of

Regulatory Fit on Processing Fluency and Persuasion,” Journal of Personality and

Social Psychology, 86 (2), 205.

Lee, Angela Y and Aparna A Labroo (2004), “The Effect of Conceptual and Perceptual Fluency

on Brand Evaluation,” Journal of Marketing Research, 41 (2), 151-65.

Lee, Thomas Y and Eric T Bradlow (2011), “Automated Marketing Research Using Online

Customer Reviews,” Journal of Marketing Research, 48 (5), 881-94.

Lewinski, Peter (2015), “Automated Facial Coding Software Outperforms People in Recognizing

Neutral Faces as Neutral from Standardized Datasets,” Frontiers in psychology, 6.

Lowe, Will (2006), “Yoshikoder: An Open Source Multilingual Content Analysis Tool for Social

Scientists,” in 2006 APSA Conference.

Lucy, John A and Richard A Shweder (1979), “Whorf and His Critics: Linguistic and

Nonlinguistic Influences on Color Memory,” American Anthropologist, 81 (3), 581-615.

Ludwig, Stephan, Ko de Ruyter, Mike Friedman, Elisabeth C Brüggen, Martin Wetzels, and

Gerard Pfann (2013), “More Than Words: The Influence of Affective Content and

Linguistic Style Matches in Online Reviews on Conversion Rates,” Journal of

Marketing, 77 (1), 87-103.

Ludwig, Stephan, Tom Van Laer, Ko de Ruyter, and Mike Friedman (2016), “Untangling a Web

of Lies: Exploring Automated Detection of Deception in Computer-Mediated

Communication,” Journal of Management Information Systems, 33 (2), 511-41.

Luedicke, Marius K, Craig J Thompson, and Markus Giesler (2010), “Consumer Identity Work

as Moral Protagonism: How Myth and Ideology Animate a Brand-Mediated Moral

Conflict,” Journal of Consumer Research, 36 (6), 1016-32.

 79

Malinowski, Bronislaw (1972), “Phatic Communion,” Communication in face to face

interaction, 146-52.

Markham, Annette and Elizabeth Buchanan (2012), “Ethical Decision-Making and Internet

Research: Recommendations from the Aoir Ethics Working Committee (Version 2.0).”

Markowitz, David M and Jeffrey T Hancock (2015), “Linguistic Obfuscation in Fraudulent

Science,” Journal of Language and Social Psychology, 0261927X15614605.

Martin, Michael K, Juergen Pfeffer, and Kathleen M Carley (2013), “Network Text Analysis of

Conceptual Overlap in Interviews, Newspaper Articles and Keywords,” Social Network

Analysis and Mining, 3 (4), 1165-77.

Martindale, Colin (1975), The Romantic Progression: The Psychology of Literary History:

Halsted Press.

Marwick, Alice E and danah boyd (2011), “I Tweet Honestly, I Tweet Passionately: Twitter

Users, Context Collapse, and the Imagined Audience,” New Media & Society, 13 (1),

114-33.

Marx, Gary T (2001), “Murky Conceptual Waters: The Public and the Private,” Ethics and

Information technology, 3 (3), 157-69.

Mathwick, Charla, Caroline Wiertz, and Ko De Ruyter (2008), “Social Capital Production in a

Virtual P3 Community,” Journal of Consumer Research, 34 (6), 832-49.

McBrian, Charles D (1978), “Language and Social Stratification: The Case of a Confucian

Society,” Anthropological Linguistics, 320-26.

McCombs, Maxwell E and Donald L Shaw (1972), “The Agenda-Setting Function of Mass

Media,” Public Opinion Quarterly, 36 (2), 176-87.

 80

McCracken, Grant (1986), “Culture and Consumption: A Theoretical Account of the Structure

and Movement of the Cultural Meaning of Consumer Goods,” Journal of Consumer

Research, 13 (1), 71-84.

McKenny, Aaron F, Jeremy C Short, and G Tyge Payne (2013), “Using Computer-Aided Text

Analysis to Elevate Constructs an Illustration Using Psychological Capital,”

Organizational Research Methods, 16 (1), 152-84.

McQuarrie, Edward F and David Glen Mick (1996), “Figures of Rhetoric in Advertising

Language,” Journal of Consumer Research, 22 (4), 424-38.

McQuarrie, Edward F, Jessica Miller, and Barbara J Phillips (2013), “The Megaphone Effect:

Taste and Audience in Fashion Blogging,” Journal of Consumer Research, 40 (1), 136-

58.

Mehl, Matthias R (2006), “Quantitative Text Analysis,” Handbook of multimethod measurement

in psychology, 141-56.

Mehl, Matthias R. and Alastair J. Gill (2008), “Automatic Text Analysis,” in Advanced Methods

for Behavioral Research on the Internet., ed. S. D. Gosling & J. A. Johnson,

Washington, DC: American Psychological Association.

Mestyán, Márton, Taha Yasseri, and János Kertész (2013), “Early Prediction of Movie Box

Office Success Based on Wikipedia Activity Big Data,” PloS one, 8 (8), e71226.

Mick, David Glen (1986), “Consumer Research and Semiotics: Exploring the Morphology of

Signs, Symbols, and Significance.,” Journal of Consumer Research, 13 (2), 196-213.

Mikolov, Tomas, Kai Chen, Greg Corrado, and Jeffrey Dean (2013), “Efficient Estimation of

Word Representations in Vector Space.”

 81

Mislove, Alan, Sune Lehmann, Yong-Yeol Ahn, Jukka-Pekka Onnela, and J Niels Rosenquist

(2011), “Understanding the Demographics of Twitter Users,” ICWSM, 11, 5th.

Mogilner, Cassie, Sepandar D Kamvar, and Jennifer Aaker (2011), “The Shifting Meaning of

Happiness,” Social Psychological and Personality Science, 1948550610393987.

Mohr, John W (1998), “Measuring Meaning Structures,” Annual Review of Sociology, 24, 345-

70.

Morris, Charles W (1938), “Foundations of the Theory of Signs.”

Morris, Rebecca (1994), “Computerized Content Analysis in Management Research: A

Demonstration of Advantages & Limitations,” Journal of Management, 20 (4), 903-31.

Namenwirth, J. Zvi and Robert Philip Weber (1987), Dynamics of Culture, Boston: Allen &

Unwin.

Narayanan, Arvind and Vitaly Shmatikov (2008), “Robust De-Anonymization of Large Sparse

Datasets,” in Security and Privacy, 2008. SP 2008. IEEE Symposium on: IEEE, 111-25.

--- (2010), “Myths and Fallacies of Personally Identifiable Information,” Communications of the

ACM, 53 (6), 24-26.

Netzer, Oded, Ronen Feldman, Jacob Goldenberg, and Moshe Fresko (2012), “Mine Your Own

Business: Market-Structure Surveillance through Text Mining,” Marketing Science, 31

(3), 521-43.

Neuman, Yair, Peter Turney, and Yohai Cohen (2012), “How Language Enables Abstraction: A

Study in Computational Cultural Psychology,” Integrative Psychological and

Behavioral Science, 46 (2), 129-45.

 82

Newman, Matthew L., James W. Pennebaker, Diane S. Berry, and Jane M. Richards (2003),

“Lying Words: Predicting Deception from Linguistic Styles,” Pers Soc Psychol Bull, 29

(5), 665-75.

Newsprosoft (2012), “Web Content Extractor,” http://www.newprosoft.com/web-content-

extractor.htm.

Ng, Sik Hung and James J Bradac (1993), Power in Language: Verbal Communication and

Social Influence: Sage Publications, Inc.

Nissenbaum, Helen (2009), Privacy in Context: Technology, Policy, and the Integrity of Social

Life: Stanford University Press.

North, Robert, Richard Iagerstrom, and William Mitchell (1999), “Diction Computer Program,”

ed. ICPSR: Inter-university Consortium for and Political and Social Research, Ann

Arbor, MI: University of Michigan.

Nunberg, Geoffrey (1993), “Indexicality and Deixis,” Linguistics and philosophy, 16 (1), 1-43.

Packard, Grant and Jonah Berger (2016), “How Language Shapes Word of Mouth’s Impact,”

Journal of Marketing Research.

Packard, Grant, Sarah G. Moore, and Brent McFerran (2014), “How Can “I” Help “You”? The

Impact of Personal Pronoun Use in Customer-Firm Agent Interactions,” MSI Report, 14-

110.

Pagescrape (2006), “Pagescrape.”

Parmentier, Marie-Agnès and Eileen Fischer (2015), “Things Fall Apart: The Dynamics of Brand

Audience Dissipation,” Journal of Consumer Research, 41 (5), 1228-51.

Pauwels, Koen (2014), “It’s Not the Size of the Data--It’s How You Use It: Smarter Marketing

with Analytics and Dashboards,” AMACOM Div American Mgmt Assn.

 83

Peirce, Charles Sanders (1957), “The Logic of Abduction,” Peirce’s essays in the philosophy of

science.

Peladeau, N (2016), “Wordstat: Content Analysis Module for Simstat,” Montreal, Canada:

Provalis Research.

Pennebaker, James W (2011), “The Secret Life of Pronouns,” New Scientist, 211 (2828), 42-45.

Pennebaker, James W, Martha E Francis, and Roger J Booth (2001), “Linguistic Inquiry and

Word Count: Liwc 2001,” Mahway: Lawrence Erlbaum Associates, 71, 2001.

Pennebaker, James W., Martha E. Francis, and Roger J. Booth (2007), Linguistic Inquiry and

Word Count (Liwc): Liwc2007, Mahwah, NJ: Lawrence Erlbaum Associates.

Pennebaker, James W. and Laura A. King (1999), “Linguistic Styles: Language Use as an

Individual Difference,” Journal of Personality & Social Psychology, 77 (6), 17p.

Pennington, Jeffrey, Richard Socher, and Christopher D Manning (2014), “Glove: Global

Vectors for Word Representation,” in EMNLP, Vol. 14, 1532-43.

Petty, Richard E, John T Cacioppo, and David Schumann (1983), “Central and Peripheral Routes

to Advertising Effectiveness: The Moderating Role of Involvement,” Journal of

Consumer Research, 10 (2), 135-46.

Piaget, Jean (1959), The Language and Thought of the Child, Vol. 5: Psychology Press.

Plaisant, Catherine, James Rose, Bei Yu, Loretta Auvil, Matthew G Kirschenbaum, Martha Nell

Smith, Tanya Clement, and Greg Lord (2006), “Exploring Erotics in Emily Dickinson’s

Correspondence with Text Mining and Visual Interfaces,” in Proceedings of the 6th

ACM/IEEE-CS joint conference on Digital libraries: ACM, 141-50.

Pollach, Irene (2012), “Taming Textual Data: The Contribution of Corpus Linguistics to

Computer-Aided Text Analysis,” Organizational Research Methods, 15 (2), 263-87.

 84

Pury, Cynthia LS (2011), “Automation Can Lead to Confounds in Text Analysis Back, Küfner,

and Egloff (2010) and the Not-So-Angry Americans,” Psychological science.

Quine, WV (1970), “The Web of Belief.”

Sapir, Edward (1929), “The Status of Linguistics as a Science,” Language, 207-14.

Schatzki, Theodore R (1996), Social Practices: A Wittgensteinian Approach to Human Activity

and the Social: Cambridge University Press.

Schau, Hope Jensen, Albert M. Muniz, and Eric J. Arnould (2009), “How Brand Community

Practices Create Value,” Journal of Marketing, 73 (5), 30-51.

Schmitt, Bernd H and Shi Zhang (1998), “Language Structure and Categorization: A Study of

Classifiers in Consumer Cognition, Judgment, and Choice,” Journal of Consumer

Research, 25 (2), 108-22.

Schouten, John W and James H McAlexander (1995), “Subcultures of Consumption: An

Ethnography of the New Bikers,” Journal of Consumer Research, 22 (1), 43-61.

Schudson, Michael (1989), “How Culture Works,” Theory and Society, 18 (2), 153-80.

Scott, Linda M (1994), “Images in Advertising: The Need for a Theory of Visual Rhetoric,”

Journal of Consumer Research, 252-73.

Sera, Maria D, Christian AH Berge, and Javier del Castillo Pintado (1994), “Grammatical and

Conceptual Forces in the Attribution of Gender by English and Spanish Speakers,”

Cognitive Development, 9 (3), 261-92.

Settanni, Michele and Davide Marengo (2015), “Sharing Feelings Online: Studying Emotional

Well-Being Via Automated Text Analysis of Facebook Posts,” Frontiers in psychology,

6, 1045.
 85

Sexton, J Bryan and Robert L Helmreich (2000), “Analyzing Cockpit Communications: The

Links between Language, Performance, Error, and Workload,” Journal of Human

Performance in Extreme Environments, 5 (1), 6.

Sherry, John F, Mary Ann McGrath, and Sidney J Levy (1993), “The Dark Side of the Gift,”

Journal of Business Research, 28 (3), 225-44.

Shor, Eran, Arnout van de Rijt, Alex Miltsov, Vivek Kulkarni, and Steven Skiena (2015), “A

Paper Ceiling Explaining the Persistent Underrepresentation of Women in Printed

News,” American Sociological Review, 80 (5), 960-84.

Snefjella, Bryor and Victor Kuperman (2015), “Concreteness and Psychological Distance in

Natural Language Use,” Psychological science, 0956797615591771.

Spiller, Stephen A and Lena Belogolova (2016), “On Consumer Beliefs About Quality and

Taste,” Journal of Consumer Research, ucw065.

Stone, Philip J (1966), The General Inquirer; a Computer Approach to Content Analysis,

Cambridge, MA: MIT Press.

Sun, Maosong, Yang Liu, Zhiyuan Liu, and Min Zhang (2015), Chinese Computational

Linguistics and Natural Language Processing Based on Naturally Annotated Big Data:

Springer.

Swanson, Don R (1988), “Migraine and Magnesium: Eleven Neglected Connections,”

Perspectives in biology and medicine, 31 (4), 526-57.

Tausczik, Yla R and James W Pennebaker (2010), “The Psychological Meaning of Words: Liwc

and Computerized Text Analysis Methods,” Journal of Language and Social

Psychology, 29 (1), 24-54.

 86

Thompson, Craig J and Elizabeth C Hirschman (1995), “Understanding the Socialized Body: A

Poststructuralist Analysis of Consumers’ Self-Conceptions, Body Images, and Self-Care

Practices,” Journal of Consumer Research, 22 (2), 139-53.

Thompson, Craig J, William B Locander, and Howard R Pollio (1989), “Putting Consumer

Experience Back into Consumer Research: The Philosophy and Method of Existential-

Phenomenology,” Journal of Consumer Research, 16 (2), 133-46.

Tirunillai, Seshadri and Gerard J Tellis (2012), “Does Chatter Really Matter? Dynamics of User-

Generated Content and Stock Performance,” Marketing Science, 31 (2), 198-215.

Townsend, Leanne and Claire Wallace (2016), “Social Media Research: A Guide to Ethics,”

University of Aberdeen.

Twenge, Jean M, W Keith Campbell, and Brittany Gentile (2012), “Changes in Pronoun Use in

American Books and the Rise of Individualism, 1960-2008,” Journal of Cross-Cultural

Psychology, 0022022112455100.

Valentino, Nicholas A (1999), “Crime News and the Priming of Racial Attitudes During

Evaluations of the President,” Public Opinion Quarterly, 293-320.

Van de Rijt, Arnout, Eran Shor, Charles Ward, and Steven Skiena (2013), “Only 15 Minutes?

The Social Stratification of Fame in Printed Media,” American Sociological Review, 78

(2), 266-89.

Van Laer, Tom, Jennifer Edson Escalas, Stephan Ludwig, and Ellis A Van den Hende (2017),

“What Happens in Vegas Stays on Tripadvisor? Computerized Analysis of Narrativity in

Online Consumer Reviews,” Available at SSRN: https://ssrn.com/abstract=2702484 or

http://dx.doi.org/10.2139/ssrn.84.

Velocityscape (2006), “Webscraper,” Phoenix, AZ: http://www.velocityscape.com/.

 87

Vico, Giambattista (1725/1984), The New Science: Cornell University Press.

Wade, James B, Joseph F Porac, and Timothy G Pollock (1997), “Worth, Words, and the

Justification of Executive Pay,” Journal of Organizational Behavior, 18, 641-64.

Wang, Jing and Bobby J Calder (2006), “Media Transportation and Advertising,” Journal of

Consumer Research, 33 (2), 151-62.

Waters, Audrey (2011), “How Recent Changes to Twitter’s Terms of Service Might Hurt

Academic Research,” ReadWrite,

http://readwrite.com/2011/03/03/how_recent_changes_to_twitters_terms_of_service_mi

/.

Weber, Klaus (2005), “A Toolkit for Analyzing Corporate Cultural Toolkits,” Poetics, 33 (3/4),

26p.

Whiteman, Natasha (2012), “Ethical Stances in (Internet) Research,” in Undoing Ethics:

Springer, 1-23.

Whorf, Benjamin Lee (1944), “The Relation of Habitual Thought and Behavior to Language,”

ETC: A Review of General Semantics, 197-215.

Wilson, Todd (2009), “Screen-Scraper,” Provo, UT.

Wong, Elaine M, Margaret E Ormiston, and Michael P Haselhuhn (2011), “A Face Only an

Investor Could Love Ceos’ Facial Structure Predicts Their Firms’ Financial

Performance,” Psychological science, 22 (12), 1478-83.

Wood, Linda A and Rolf O Kroger (2000), Doing Discourse Analysis: Methods for Studying

Action in Talk and Text: Sage Publications.

Zipf, George Kingsley (1932), “Selected Studies of the Principle of Relative Frequency in

Language.”
 88

Figure 1: Stages of Automated Text Analysis

 89

Table 1: Standardized Dictionaries

Example No. of No. of

Dictionary Description Year Citation
Categories Words Categories
Consumer

A weighted wordlist used to Kovács, B., Carroll, G. R., & Lehman, D. W.

measure authenticity (2013). Authenticity and consumer value ratings:
Authenticity NA 2013 90 1
developed from 35 Empirical tests from the restaurant
participants domain. Organization Science, 25(2), 458-478.
Sincerity, Opoku, Abratt and Pitt (2006), “Communicating
Excitement, brand personality: Are the websites doing the
Wordlist to measure Aaker’s
Brand Competence, talking for the top South African Business
(1997) five traits of brand 2006 833 5
Personality Sophisticatio Schools?.” Journal of Brand Management14, no. 1-
personality
n, 2: 20-39.
Ruggedness
Xu, Z., & Bengston, D. N. (1997). Trends in
Aesthetic,
Measures value systems national forest values among forestry
Environ- Utilitarian,
related to discussion of the 1995 612 4 professionals, environmentalists, and the news
mental Life Support,
natural environment media, 1982–1993. Society & Natural
Moral
Resources, 10(1), 43-59.
General
Sadness,
Pennebaker, J. W., Boyd, R. L., Jordan, K., &
A broad dictionary to Negative
13 Blackburn, K. (2015). The development and
measure parts of speech, Emotion,
categories psychometric properties of LIWC2015. UT
LIWC but also psychological and Overall 2015 6,400
, 68 sub- Faculty/Researcher Works.
social categories and Affect, Verb,
categories
processes and Past
Focus
Zachary, M.A., McKenny, A., Short, J, & Davis,
K.M., Wu, D. (2011). Franchise branding: an
Certainty,
Measures verbal tone, 5 organizational identity perspective. Journal of the
Activity,
developed to study public categories Academy of Marketing Science, 39(4): 629-645.
Diction Optimism, 2015 10,000
communication such as , 36 sub-
Realism,
political speeches categories Yadav, M. S., J. C. Prabhu, & R. K. Chandy.
Commonality
(2007). Managing the Future: CEO Attention and
Innovation Outcomes. Journal of Marketing, 71:
 90

84–101.

Roget, P.M., 1911. Roget’s Thesaurus of English

Space Words and Phrases. TY Crowell Company.
matter, Yarowsky, D., (1992)”Word-sense disambiguation
Wordlist based on Roget’s 6
intellectual using statistical models of Roget’s categories
(1911) thesaurus, which categories
facilities, trained on large corpora,” In Proceedings of the
Roget includes 6 broad categories 1999 100,685 , 1,042
voluntary 14th conference on Computational linguistics-
such as words related to sub-
powers, Volume 2 (pp. 454-460). Association for
space or matter categories
spiritual or Computational Linguistics.
moral powers

Fellbaum, Christiane (2005). WordNet and

wordnets. In: Brown, Keith et al. (eds.),
117,659
A large lexical dictionary that feeling, Encyclopedia of Language and Linguistics, Second
synsets;
categories a variety of adverbs, Edition, Oxford: Elsevier, 665-670.
WordNet 2005 147,278 44
objects, feelings, and possession,
unique
processes animal George A. Miller (1995). WordNet: A Lexical
words
Database for English.

positive,
negative,
A large dictionary that active, Philip J. Stone, Robert F. Bales, Zvi Namenwirth, &
captures a wide variety of passive, Daniel M. Ogilvie (1962). The General Inquirer: A
Harvard IV
semantic and pragmatic pleasure, computer system for content analysis and retrieval
Psychological 1977 11,789 184
aspects of language as well pain, virtue, based on the sentence as a unit of information.
Dictionary
as individually-relevant and vice, Behavioral Science, 7(4), 484-498.
institutionally-relevant words economy,
legal,
academic
Often referred to as
“Laswell’s dictionary”, this power, 8
wordlist contains words respect, categories
1969/ Lasswell, H.D. and Namenwirth, J.Z., 1969. The
Values associated with deference affiliation, 10,628 , 68
1998 Lasswell value dictionary. New Haven.
and welfare. It is often wealth, well- subcatego
combined with the Harvard being ries
IV
Psychological
 91

37,058
words
A weighted wordlist to Brysbaert, M., Warriner, A. B., & Kuperman, V.
and
measure concreteness (2014). Concreteness ratings for 40 thousand
2,896
Concreteness based on 4,000 participants’ NA 2013 8 generally known English word lemmas. Behavior
two-
ratings of the concreteness research methods, 46(3), 904-911.
word
of many common words
express-
ions
Measures words associated need,
Regressive Martindale, C. (1975). Romantic progression: The
with Martindale’s (1975) sensation, 1975,
Imagery 3,200 43 psychology of literary history. Washington, D.C.:
primary or secondary abstract 1990
Dictionary Hemisphere.
processes thought
Anaphore,
Probability
and
Communicati Measures vagueness in Hiller, J.H. (1971). Verbal response indicators of
Possibility,
on verbal and written 1971 196 10 conceptual vagueness. American Educational
Admission of
Vagueness communication Research Journal. 8(1), 151-161.
Error,
Ambiguous
Designation

Transparency Wilson, A. (2006). Development and application of

Operationalizes Fisher and
, Container, a content analysis dictionary for body boundary
Body Type Cleveland’s (1958) theory or 2006 515 12
Protective research. Literary and Linguistic Computing, 21,
self-boundary strength
Surface 105-110.

Sentiment
Word lists for textual
analysis in financial
analyses. Each word is
Negative,
Loughran labeled with sentiment
Positive, Tim Loughran and Bill McDonald, 2014,
and categories as well as
Uncertainty, 2014 85,132 9 “Measuring Readability in Financial
McDonald whether the word is a
Litigious, Disclosures”, Journal of Finance, 69:4, 1643-1671.
Sentiment irregular verb and number of
Modality
syllables. Dictionary was
constructed from all 10-k
documents from 1994-2014
 92

negative, Hutto, C. J., & Gilbert, E. (2014, May). Vader: A

A dictionary- and rule-based neutral, parsimonious rule-based model for sentiment
VADER 2015 9,000 4
tool for measuring sentiment positive, analysis of social media text. In Eighth International
compound AAAI Conference on Weblogs and Social Media.
Bradley, M.M., & Lang, P.J. (1999). Affective
norms for English words (ANEW): Stimuli,
pleasure,
Affective ratings of everyday instruction manual and affective ratings. Technical
ANEW pain, arousal, 1999 1,033 4
words based on surveys report C-1, Gainesville, FL. The Center for
dominance
Research in Psychophysiology, University of
Florida.
A large sentiment dictionary positive
Baccianella, S., Esuli, A., & Sebastiani, F. (2010,
based on the WordNet sentiment,
May). SentiWordNet 3.0: An Enhanced Lexical
Senti dictionary. Each mark is negative 117,659
2010 3 Resource for Sentiment Analysis and Opinion
WordNet marked by a positive sentiment, synsets
Mining. In LREC (Vol. 10, pp. 2200-2204).
sentiment and negative objective
sentiment score language
Sociological
Weber, K. “Converging orders of worth: Adaptive
civic, market, and demographic mechanisms of cultural
industrial, globalization.”
Operationalizes Boltanski
Orders of spiritual, and
and Thevenot’s (2006) 2012 344 5
Worth domestic van Bommel, K. (2014). Towards a legitimate
orders of worth
orders of compromise? An exploration of Integrated
worth Reporting in the Netherlands. Accounting, Auditing
& Accountability Journal, 27(7), 1157-1189.
Liberal,
Conservative,
Measures political position in Laver. M. & Garry, J. (2000). Estimating Policy
Policy Pro-
public policy or other 2000 415 19 Positions from Political Texts. American Journal of
Position Environment,
politically-oriented texts Political Science, 44 (3), 619-634.
Con-
Environment
 93

Headings List

1) ROADMAP FOR AUTOMATED TEXT ANALYSIS

2) Stage 1: Developing a Research Question
2) Stage 2: Construct Identification
3) Attention
3) Processing
3) Interpersonal Dynamics
3) Group and Cultural Level Characteristics
2) Stage 3: Data Collection
3) Identifying Data Sources
3) Preparing Data
3) Unitizing and Storing the Data
2) Stage 4a: Choose an Operationalization Approach
3) Top-Down Approaches
3) Bottom-up Approaches
2) Stage 4b: Execute Operationalization
3) Dictionary- and Rule-based Approaches
3) Classification
3) Topic-Discovery
2) Stage 5: Interpretation and Analysis
3) Comparison
3) Classification
3) Prediction
2) Stage 6: Validation
3) Types of Validity
3) Ethics
1) CONCLUSION
2) Summary
2) Future Directions