Content uploaded by Helen Kennedy
Author content
All content in this area was uploaded by Helen Kennedy on Aug 01, 2016
Content may be subject to copyright.
Content uploaded by Helen Kennedy
Author content
All content in this area was uploaded by Helen Kennedy on Aug 01, 2016
Content may be subject to copyright.
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=hbem20
Download by: [Royal Hallamshire Hospital] Date: 11 December 2015, At: 03:17
Journal of Broadcasting & Electronic Media
ISSN: 0883-8151 (Print) 1550-6878 (Online) Journal homepage: http://www.tandfonline.com/loi/hbem20
Perspectives on Sentiment Analysis
Helen Kennedy
To cite this article: Helen Kennedy (2012) Perspectives on Sentiment Analysis, Journal of
Broadcasting & Electronic Media, 56:4, 435-450, DOI: 10.1080/08838151.2012.732141
To link to this article: http://dx.doi.org/10.1080/08838151.2012.732141
Published online: 12 Dec 2012.
Submit your article to this journal
Article views: 808
View related articles
Citing articles: 1 View citing articles
Perspectives on Sentiment Analysis
Helen Kennedy
One of the consequences of the widespread use of social media is the equally
widespread availability of all sorts of once intimate and private stuff: textual,
visual, and affective. From this, a new form of labor arises: the mining of
social media data. One type of social media data mining is sentiment analysis,
the application of a range of technologies to determine sentiments expressed
within social media about particular topics. This article maps out a range of
emerging perspectives on sentiment analysis and argues that these sometimes-
competing views need to be brought together, so that analyses of new socio-
technical phenomena like sentiment analysis can be rich and rounded.
The Rise of Sentiment Analysis
One of the consequences of the widespread use of social media is the equally
widespread availability of all sorts of once intimate and private stuff: textual, visual,
and affective. From this, a new form of labor arises: the mining of social media
data. Sentiment analysis, one particular form of social media data mining, involves
the application of a range of technologies to determine sentiments expressed within
social media platforms about particular topics, in order to arrive at a measure of
the ambient, or general sentiment (Andrejevic, 2011; Arvidsson, 2011). Sentiment
analysis uses linguistic and textual assessment, such as Natural Language Processing,
to analyze word use, word order, and word combinations and thus to classify
sentiments, often into the categories of positive, negative, or neutral.
There has been a rise in the use of sentiment analysis by branding, marketing,
and advertising companies in recent years such that it is now a significant site of
cultural production. Sentiment analysis has arisen in a context that is said to be
characterized by an overabundance of corporate, marketing messages, a decline of
trust in advertisements, and a growth of trust in peer recommendations. Markets
are conversations, it is claimed (Locke, Searls, & Weinberger, 2000), and through
social media, ‘‘the conversations happen in front of millions of people, and they’re
Helen Kennedy (Ph.D., University of East London) is a senior lecturer of new media at the University of
Leeds in the UK. Her research interests include new media work; specifically Web design and the work
of the social media industries; new media theory, practice and creativity; the myths of new media; media
industries and cultural labor.
©2012 Broadcast Education Association Journal of Broadcasting & Electronic Media 56(4), 2012, pp. 435–450
DOI: 10.1080/08838151.2012.732141 ISSN: 0883-8151 print/1550-6878 online
435
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
436 Journal of Broadcasting & Electronic Media/December 2012
archived for years to come’’ (Zarella, 2009, p. 1). Stories abound of companies
whose reputations have been made or broken by social media chatter, such as the
computer company Dell and the painkiller Motrin (Hunt, 2009).1However dubious
this set of assertions might be, and however difficult it might be to find examples
of bottom line value or prove causal relationships, as a result of such claims and
anecdotes, corporations are keen to know what is being said about their products
in social media, and what sentiments are being expressed.
It is not only commercial companies that are using sentiment analysis. Academics,
politicians, media organizations, and charities also use it or its companion, opinion
mining. Data gathered through sentiment analysis are believed to provide detailed
information about something to which direct access did not previously exist: public
opinion and feeling. Such information could indicate the success or otherwise
of a marketing campaign, for example, which may in turn lead advertisers to
alter their strategies. So sentiment analysis companies gather and analyze social
media sentiment, sell this information to their customers, and this social media
intelligence then becomes the basis for action. Figure 1, a graphic taken from the
company General Sentiment’s Web site, provides a visual illustration of this process.
Alongside commercial products, more freely available tools exist, such as Social
Mention (http://socialmention.com/), a social media search and analysis platform
that ‘‘aggregates global mentions into a single stream of information.’’
Given that sentiment analysis is a fairly recent phenomenon, it is not surprising
that there have been few studies of it from within any academic discipline. Those
who have commented on the phenomenon have expressed concern because of the
monetization of intimacy and the extraction of value that results (Hearn, 2010),
and because of the role it plays in the modulation and control of affect (Andrejevic,
2011). In this article, I suggest weaving these important critical perspectives together
with other perspectives, in order to arrive at rich and rounded understandings of new
socio-cultural phenomena like sentiment analysis. These include the perspectives
of a range of actors, including sentiment analysts themselves, as well as the people
Figure 1
Visualization of the Sentiment Analysis Process, Taken from General Sentiment’s
Web Site (http://www.generalsentiment.com/what-we-do.html)
and Used with Permission
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 437
whose social media activities are mined and monitored. In addition, drawing on
the work of Andrew Sayer (2004) and Russell Keat (2000, 2011), I suggest that
a ‘‘moral economy’’ perspective might also be productive—because of the ways
in which it highlights both the ethical complexity and the practical diversity of
sentiment analysis.
The remainder of the article maps out this range of perspectives. First, I discuss the
issues that concern sentiment analysts themselves, after which I highlight the con-
cerns of scholars working within a critical Marxist tradition who have commented
on this emergent field. I then present some preliminary findings from small-scale
empirical research with people working within sentiment analysis, which points
towards some of the issues revealed by a moral economy approach, highlighting
the moral and ethical concerns of sentiment analysts themselves. Finally, I briefly
discuss the perspectives of other important players, about whom not much is yet
known—social media users.
Perspectives of Sentiment Analysts
Amongst sentiment analysts, key concerns relate to accuracy of analysis, and
quantity and ‘‘cleanliness’’ of data. Accurately identifying social media sentiments
is not easy. This is not only because it is difficult for humans to agree about the
sentiment of a text, but also because of the complex ways in which humans express
sentiment, using irony, sarcasm, humor, or, in social media, abbreviation. In the
field of sentiment analysis, 70% accuracy is considered good—but that is 70%
agreement with human judgment about whether a sentiment is positive, negative,
or neutral, not 70% accurate identification of sentiments. Less accuracy is common,
as a number of my interviewees acknowledged. One respondent stated that 70%
inaccuracy would be a more realistic figure. Because of these problems, two of the
companies whose workers I interviewed use manual, human analysis, rather than
machine analysis, in an effort to achieve a reasonable degree of accuracy.
In this age of big data, the quantity of data analyzed by sentiment analysis
companies is inevitably vast, at least in those that carry out machine analysis.
Indeed, Mark Andrejevic argues that claims made by sentiment analysis companies
about their services refer not to what he calls ‘‘referential accuracy’’ (that is, that the
data can actually be taken to represent sentiments), but rather to the huge quantities
of data analyzed (Andrejevic, 2011). Size makes up for the roughness of the data, he
claims. Nonetheless, it is not always the case that limitless quantities of ‘‘sentiments’’
are there for the analyzing. An academic sentiment analyst who participated in my
research said that he was disappointed by the amount of sentiments he identified
on Twitter. He claimed that in his research into tweets about news, he found that
‘‘people don’t tweet much sentiment about topics that they tweet a lot about.’’ In
other words, there is not always a lot of sentiment to be analyzed in news tweets
on Twitter.
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
438 Journal of Broadcasting & Electronic Media/December 2012
Sentiment analysts are also concerned about the ‘‘cleanliness’’ of data, and talk
about ‘‘cleaning up’’ sentiments once they have been identified. This is necessary
because there is plenty of evidence of ‘‘referential inaccuracy,’’ to adapt Andreje-
vic’s term, or of fake or otherwise unreliable data. For example, Fowler and de
Avila (2009) point out that there is a ‘‘positivism problem’’ in online sentiments and
opinions, as the average rating for all things reviewed, from printer paper to dog
food, is 4.3 out of 5.2They quote Ed Keller of Market Research Group Keller Fay
who says that ‘‘there is an urban myth that people are far more likely to express
negatives than positives,’’ whereas the opposite is true. In their surveys, Keller Fay
found that around 65% of reviews were positive, whereas only 8% were negative.
Likewise, Fowler and de Avila (2009) point out that some We bsites acknowledge
that companies may be submitting reviews of their own products, and that negative
reviews may be suppressed, further indication of this ‘‘positivism problem.’’ In
addition, one of my respondents who runs a reputation services company suggested
that more than 10% of online reviews may be falsely negative and therefore may
provide further evidence of referential inaccuracy.
From the perspective of sentiment analysts, opinion and sentiment are inter-
changeable, as the title of Pang and Lee’s (2008) overview of the field, ‘‘Opin-
ion mining and sentiment analysis,’’ indicates. Thelwall, Buckley, and Paltoglou
(2011) also suggest that they are the same thing. Thus there is surprisingly little
discussion amongst sentiment analysts about whether what is analyzed actually
represents sentiments, or indeed, what a sentiment is. Dictionary definitions of the
term sentiment suggest that it is an attitude based on a feeling, this emotive element
distinguishing it from opinion, which may be more factually based (although plenty
of research has been done to point to the important role of emotions in all kinds of
rational decision-making, such as Nussbaum, 2003). For Pang and Lee (2008), the
important characteristic of both sentiment and opinion is that they are subjective,
private states ‘‘not open to objective observation or verification’’ (p. 9), an assertion
which disregards the emotional and affective character of sentiments. Instead, there
is a rather crude assumption amongst practitioners that, as Thelwall et al. (2011)
state, sentiment analysis gives researchers ‘‘the ability to automatically measure
emotion in online text’’ (p. 408). In response to such assertions, it seems vital to ask
whether such a task is possible at all, never mind as straightforward as implied here,
given the psychological complexity of emotions, and their equally complex cultural
politics (Ahmed, 2004). Such assertions overlook the complex, affective qualities of
sentiments—as does the categorization of sentiments into only three simple types:
positive, negative and neutral.
Extensive research has been carried out into what social media users think they
are doing on social media, which raises further questions about whether we are
dealing with sentiments here. In one example, Marwick and boyd (2010) identified
extensive self-censorship in their research on Twitter users, in which many of their
respondents suggested that they aim for balance in their tweets. Their tweets may
therefore reflect more measured and balanced sentiments than they feel, and what
appears as sentiment may in fact be its performance. If users of social media like
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 439
Twitter self-consciously self-construct, as some have suggested (Hearn, 2008), or
if they conceal more than they reveal, as Marwick and boyd found, then whether
sentiment analysis is taking place at all is debatable. As Hearn (2010) points out,
social media sentiments are conditioned by the systems through which they are
expressed, yet further indication of the need to acknowledge their likely referential
inaccuracy. Yet these issues are rarely discussed by sentiment analysts.
Critical Perspectives
In this section, I briefly sketch three critical perspectives on sentiment analysis,
and then bring these into dialogue with the perspectives of sentiment analysts
discussed above. These include Hearn’s (2010) view that social media monitoring
practices like sentiment analysis represent the extraction of value from feeling,
Andrejevic’s (2011) argument that it represents a form of affective control, and
Turow’s (2012) claim that it results in social discrimination.
Alison Hearn argues that the economic valuation of affect through sentiment
analysis leads to the monetization of intimacy, feeling, and friendship. For Hearn,
sentiment analysis thus represents yet another capitalist mechanism of value extrac-
tion. She describes the people doing this work as ‘‘feeling-intermediaries.’’ ‘‘Feeling-
intermediaries structure feelings into profits for themselves and their clients’’ (2010,
pp. 435–436), she claims, arguing that their systems ‘‘mark the point at which human
feelings are commodified’’ (p. 428). For Hearn, the work of feeling-intermediaries
is a problem in that it produces value from mined sentiments not for the people
who might claim some ownership of them, but for corporations, brand-managers,
and marketers:
what is extracted from the expression of feeling is valuable only to those who
develop, control and license the mechanisms of extraction, measurement and rep-
resentation, and not for the people doing the expressing. (Hearn, 2010, p. 423)
Power in this sector is therefore ‘‘primarily enacted by the media industries’’ (Hearn,
2010, p. 424), a fact of which she is critical. Sites like TripAdvisor, she argues,
operate ‘‘under the guise of serving consumers’ interests’’ by providing them with
a space to express opinion or sentiment whilst ‘‘their corporate clients are able to
access these suggestions, using them to grow brand equity and develop products’’
(2010, p. 431).
Similarly, Mark Andrejevic (2011) is concerned about the role played by sentiment
analysis in the prediction and subsequent control of affect, which he describes,
quoting Massumi (2002), as ‘‘an intrinsic variable of the late capitalist system, as
infrastructural as a factory’’ (Massumi, 2002, p. 45, quoted in Andrejevic, 2011,
p. 609). Andrejevic (2011) argues that affect, ‘‘a circulating, undifferentiated kind
of emotion’’ (p. 608), is an exploitable resource within affective economies, and
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
440 Journal of Broadcasting & Electronic Media/December 2012
its exploitation results in forms of control described by Clough (2003) as ‘‘a never-
ending modulation of moods, capacities, affects, potentialities assembled in genetic
codes, identification numbers, ratings profiles and preference listings’’ (p. 360,
quoted Andrejevic, 2011, p. 608). He proposes that through mechanisms like sen-
timent analysis, emotions are abstracted from individuals, and instead constitute a
kind of background intensity that is of no-one. Marketing strategies focus on the
measurement and surveillance of this background intensity, or ambient sentiment.
Andrejevic points out that sentiment analysis companies use the language of ‘‘listen-
ing’’ to describe the services that they offer, but, he suggests, the goal is not really
to listen, but rather ‘‘to monitor and oversee’’; ‘‘to aggregate and mine [individual
voices] in order to trace signals in the noise and to extract information to improve
: : : marketing campaigns’’ (2011, p. 611).
Discussing a broader range of marketing practices than Hearn and Andrejevic,
Joseph Turow (2012) locates sentiment analysis and other forms of social media
monitoring in the context of the techniques the digital advertising industry has long
used to monitor, measure, and understand its target audiences’ psychographics, in
his book The Daily You: How the New Advertising Industry is Defining Your Identity
and Your Worth (as does Arvidsson, 2011). Turow’s argument is that such practices
ultimately lead to social discrimination. They do this by turning ‘‘individual profiles
into individual evaluations’’ (2012, p. 6): individuals’ marketing value is calculated,
based on behavioral and other forms of tracking, and each individual is categorized
as target or waste. These data define our identity and our worth, suggests Turow,
determining not only what marketing firms do, but also how we see ourselves and
others. This is because those of us who are considered to be waste receive narrowed
options in terms of the advertising messages that are targeted at us, and, according to
Turow, these messages constitute a form of social discrimination that impacts upon
our sense of self. Of course, as Turow points out, advertisers’ notions of individuals
depend on the choices they make about the firms from which to purchase data, as
data and information are transient and there is substantial variability in how profiles
are created.
These critical interventions point to some of the troubling consequences of emer-
gent practices like sentiment analysis. They counterbalance the more celebratory
accounts of the democratic possibilities opened up by social media and their ac-
claimed possibilities for participation (for example in Jenkins, 2008). As Andrejevic
(2011) states, participatory culture ‘‘has the potential to cut both ways: the increasing
influence of participatory consumers on the production process, and the facilitation
of monitoring-based regimes of control’’ (p. 612). Thus these critical perspectives
do the important job of making visible what is largely invisible—that is, the work
of monitoring, mining, and tracking the data, opinion, and sentiment trails that we
leave behind as we move through social media. As Turow points out, the decision
to make such practices invisible was a conscious one. He quotes Richard Smith of
the Electronic Frontier Foundation who asked ‘‘Why are web bugs [used for tracking
consumer activity] invisible on a page?’’ to which Smith answered ‘‘To hide the fact
that monitoring is taking place’’ (quoted in Turow, 2012, p. 61).
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 441
Combining these critical perspectives with the perspectives of key actors within
the field of sentiment analysis, such as sentiment analysts themselves and the people
who produce the content that is mined and monitored, can further enrich under-
standings of the cultural significance of social media monitoring, as it means bring-
ing together structural analyses with a recognition of individual agency within these
structures. This might not be quite what Jeremy Gilbert (2012) recently described as
a necessary ‘‘synthesis between the classical Marxist denunciation of neoliberalism
and the neo-autonomist insistence on recognizing the collective agency of the
multitude,’’ as sentiment analysts can hardly be described as a multitude, but the
productiveness of such as synthesis is clearly applicable here. Many of the issues
that concern sentiment analysts point to the limitations in actually doing sentiment
analysis, as accuracy and cleanliness of data, to use their terms, are difficult to
achieve. Indeed, the interviewee who suggested there is 70% inaccuracy in senti-
ment analysis hinted at both its lack of reliability and its absurdity (though some
practitioners decidedly would not share her cynicism). How significant is it then, as
a cultural phenomenon, and how seriously should we take it? What’s more, what
does the slippage between sentiment and opinion within sentiment analysis mean
for the claims that are made about it—for example, can it be described as affective
control if there is no affective element? These are some of the questions that emerge
from synthesizing critical Marxist perspectives with those that recognize agency. In
the next section, I say more about the agency of individual sentiment analysts,
through a consideration of the issues that arise from using a ‘‘moral economy’’
approach.
A Moral Economy Perspective
In an article entitled ‘‘Moral Economy,’’ Andrew Sayer (2004) argues that eco-
nomic decisions, behaviors, and institutions ‘‘depend on and influence moral/ethical
sentiments, norms and behaviors and have ethical implications’’ (p. 2). ‘‘Ethical and
moral valuation is always either present or latent’’ in economic behavior, he claims
(2004, p. 4). Sayer is one of several British political theorists and philosophers
interested in examining the relationships between ethics and markets. Like him,
Russell Keat has written extensively on the intersection of ethics, morality, and
markets, for example in ‘‘Every Economy is a Moral Economy’’ (2004) and ‘‘Market
Economies as Moral Economies’’ (2011). In these article, Keat argues that the critical
evaluation of market economies must include ethical judgments about the goods
and ills of production, consumption, and exchange. Similarly, John O’Neill (1998)
asserts that non-economic associations are central to economic life, and that social
life ‘‘requires serious commitments which are non-contractual in nature’’ (p. 76).
Elsewhere, other writers have highlighted the ethical character of a range of activ-
ities, some of them work-related, such as Kathleen Gibson and Julie Graham (2006),
in their research into alternative economic models and the post-capitalist politics
they represent. Gibson and Graham acknowledge that through their research, their
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
442 Journal of Broadcasting & Electronic Media/December 2012
intention was to make an intervention, to identify and facilitate openings and pos-
sibilities, such as ‘‘the ethical opening of persons to one another that conversation
provokes and enables’’ (2006, p. 135). For me, engaging in conversation with
sentiment analysts and other people working in this field about the moral and
ethical dimensions of their professional decision-making also represents a search for
ethical openings. How do these workers reconcile concerns about the surveillant
and monetizing dimensions of sentiment analysis discussed in the previous section
within their daily work practices? Addressing these important questions directly with
sentiment analysts themselves is an essential part of the kind of moral economy
approach I propose here.
The writers discussed briefly here make a convincing case that labor not only
results in value for those who own its outputs, but also needs to be understood
as a process which involves a series of judgments based on the values of laborers
themselves. This assertion that value and values come together in labor applies to
sentiment analysis just as much as any other form of work. Value and values merge
not only in the sense that value is made out of consumer values, as Arvidsson (2010)
suggests, but also because producer values play a role in the ways in which feeling-
intermediaries carry out their work. The following paragraphs flesh out this claim by
reflecting on interviews with people working in sentiment analysis and other forms
of social media monitoring, either as sentiment analysts or in other roles (such as
account manager, or head of social media). Eight interviews have been carried out
in two European countries, six in sentiment analysis or social media monitoring
companies, one with an academic sentiment analyst, and one with the head of an
online reputation management company.3I accessed interviewees through existing
networks, a mailing list, and by contacting social media monitoring companies
identified through a Web search. This is ongoing research that clearly represents a
modest amount of data, but I draw on it here to indicate the kinds of issues that
might emerge from taking a ‘‘moral economy’’ perspective.
Before discussing the interviews, a comment on respondents’ backgrounds is
useful. One had experience in what he describes as ‘‘integrity-led’’ communications
consultancy and another worked in Web search engine optimization, where she
actively opposed unethical practices, prior to her current work in social media
monitoring. Three respondents have backgrounds in media and communication
studies; during our interview, one actually said that she had recently been asking
herself, in relation to the work that she does, ‘‘how would Marx frame everything?!’’
These backgrounds in critical media studies and ethical media work are significant,
for, as McRobbie (1999) acknowledges, the fact that they constitute part of the
formation of today’s army of cultural laborers may give us cause to be hopeful
about the ways in which such work gets done.
Ethical considerations could be identified in the ways in which respondents’
companies access sentiment, use sentiment, and intervene in sentiment. When asked
if they have codes of practice regarding which social media sentiments they gather
and analyze, all respondents said that they only look at publicly available data. If
data are behind a firewall, or are not publicly available, then they are considered
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 443
to be off limits. ‘‘If there’s an indication that people don’t want their Web site
looked at, then we don’t look,’’ said one respondent. This is guaranteed by the
fact that the functionality to log-in to closed sites is not built in to her company’s
technology, although it could be. In such cases, although it would be technically
possible to breach systems’ terms of service and ‘‘walk under the fence,’’ as one
respondent put it, the decision was made to adopt a more principled practice.
Therefore, not all sentiments are examined, mined, and monitored. This decision
is clearly commercially wise as well as ethical, as potential clients may be uneasy
with more dubious mechanisms for gathering sentiment.
However, it is well established that what is public and private is complex in social
media environments. danah boyd (2010) uses the notion of ‘‘being private in public’’
to highlight this complexity, comparing social media with corridors. If two people
bump into each other in the public space of a corridor, she argues, one might
say something private to the other that s/he would not want to have publicized.
Instead, the speaker would want ‘‘privacy in context,’’ or what Nissenbaum (2009)
calls ‘‘contextual integrity.’’ I put this to my respondents. Some respondents agreed
that deciding what is public is problematic in social media. They acknowledged
that some information is public both from a systems point of view and because
it is clearly directed at companies and appears to require an answer, but as one
respondent stated, ‘‘It’s not crystal clear what’s public and what is not.’’ Not all
respondents agreed with this, though. One respondent’s company lists amongst its
clients a large multinational for whom she tracked its employees online to identify
‘‘what feeling they have’’ about the organization. She said it is social media users’
responsibility to understand how public their social media messages are. She added:
‘‘If I do a nude photo and upload it on Twitter : : : Whose fault is it? ‘Oh! I didn’t
read that this could be seen by everyone’ : : : So, you should have!’’ When asked
what the ethical limitations of sentiment analysis should be, another respondent
replied ‘‘as far as the user allows,’’ again suggesting that what is made technically
public is fair game to be mined and monitored.
One respondent felt that the use of pseudonyms on forums goes some way towards
guaranteeing the privacy of people expressing sentiments in these environments.
Another felt that the central ethical issue in relation to accessing data was the
widespread ignorance of such practices, suggesting that the ethical ethos of senti-
ment analysis could be something like ‘‘report that it exists.’’ Another interviewee
pointed to the responsibilities that forum administrators and Web site owners have
in protecting the privacy of their systems’ users. The head of the reputation services
company was much more explicitly critical of the organizations from whose Web
sites sentiments are extracted, questioning who should receive criticism in relation
to sentiment analysis, feeling intermediaries themselves, or the firms which make
their users’ feelings open to intermediation. Such companies benefit both from the
opacity of what is public in social media and from the absence of appropriate
legislation, he claimed, and they use the absence of legal liability as an excuse for
failing to exercise due diligence. It is here that critical attention needs to be focused,
he argued, rather than on data miners and sentiment analysts themselves. Thus a
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
444 Journal of Broadcasting & Electronic Media/December 2012
number of social media actors are invoked as having responsibilities in relation to
sentiment analysis, not just the feeling-intermediaries themselves. What’s more, it
is not the case that all social media sentiments are systematically mined regardless
of people’s wishes—some remain the preserve of the sentiment-expresser.
What happens to sentiments once they have been mined? This is another issue
to which some respondents had given consideration. One said that he thought
his company had ethical responsibilities in relation to ‘‘what we do with the data
when we get it.’’ And there are differences in the ways in which sentiment analysis
companies operate in this regard. Some companies advise clients on how to engage
in social media conversation; others do this on clients’ behalves, offering services
known as community management. Other companies choose not to offer such
advice or services. Instead, the companies’ clients decide for themselves what to
do with their data. But even those that offer these services do not always do so
willingly. One respondent whose company engages in ‘‘community management’’
on behalf of clients expressed her ethical unease about her company intervening
in other people’s conversations, saying that she believed that ‘‘we should let our
clients look after their own voice’’ because ‘‘as much work as we can do to get
under the skin of our clients, we are not them’’ and cannot speak with the authentic
voice that they can. And just as some commercial companies do not advise their
clients on strategies for intervening in social media conversations or targeting key
influencers, the academic sentiment analyst amongst my interviewees also used the
notion of non-intervention as an ethical justification for his work:
I don’t monitor people, in the sense that nothing that I do is intended to affect the
people that I’m getting the texts from. So the end result of my analysis is completely
neutral to them. So I won’t try to sell them anything, or try to get them to modify
their behavior in anyway.
Andrejevic’s (2011) claim that emotions are abstracted from individuals in senti-
ment analysis is echoed in this and other interviewees’ comments. Another respon-
dent, for example, expressed the view that what is measured by sentiment analysis
is the general sentiment, not individualized emotions: ‘‘We’re not going too much
into the individual, but trying to catch the feeling,’’ he said. As a result of ‘‘trying
to catch the feeling,’’ as he put it, he felt that his company’s practices were not
harmful or intrusive to individuals. He continued:
We do not place too much attention to the individual statements, but more the total
picture of what is said in a given time period about a company: : : : We can identify
persons if we want to, but we don’t see the relevance of it.
He continued to give an example of a public relations agency that asked his
company to cull data from journalists, so that the agency could inform customers of
what individual journalists were writing about them. ‘‘And we chose not do that,’’
he said, because ‘‘this is not something we want to do. I don’t think it would be in
conflict legally, but we feel it’s one step too far.’’
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 445
Furthermore, some respondents pointed to the benefits that result from some uses
to which their companies’ data are put. The academic sentiment analyst amongst my
respondents felt that the uses of his data were ethically sound, because through his
research, he aimed to generate socially useful knowledge. He said ‘‘my justification
for doing sentiment analysis would be that I’m trying to get findings that are useful
for society.’’ For another respondent, the majority of his company’s clients are other
media organizations, interested not in growing brand equity or changing consump-
tion behavior, but rather in understanding people’s concerns. Another interviewee’s
company works extensively with a world leading co-operative corporation which
promotes ethical business practices. This respondent felt optimistic about the pre-
dictive capacities of sentiment analysis that concern Andrejevic, because they may
contribute to the development of treatments relating to sentiment-related illnesses
such as depression, she suggested. Finally, another respondent commented on how
the data her company gathers get put to all sorts of different uses by a range of
different organizations:
charities use social media monitoring to understand how people feel about world
events and how they can encourage them to contribute to positive change by
donating money or volunteering time.
Because of the opportunities available to engage with and get heard by companies
and brands, consumers now have more control over brands, some respondents
claimed, as evidenced in the stories of Dell and Motrin, outlined in Footnote 1.
‘‘People have more power to impact on how companies behave,’’ said one re-
spondent. Another respondent talked about consumers’ refusal to behave on social
media sites as companies would like them to, which he saw as a form of consumer
power. ‘‘I don’t see that trying to catch the sentiment from these kinds of data is
something that empowers companies more than normal persons,’’ he proposed,
because consumers have the power to say ‘‘we choose to say something opposite
to what they want us to say.’’
Whilst such views may seem to reflect idealized notions of the empowering
potential of participatory cultures which we may want to problematize (and indeed
some respondents did just that), they also point to a belief in the possibility of
social media users’ agency in the production of knowledge about them. Marwick
and boyd’s (2010) findings also point to the agency of social media users, who
make conscious decisions about which sentiments to express or perform, and who
are thus not simply and universally the prey of feeling-intermediaries. Recognition
of such agency is made possible through the approach I am proposing here, which
brings together a range of perspectives on sentiment analysis. These views about
the various ways in which consumers may gain from the increased value attached
to social media chatter confirm Arvidsson’s (2011) proposal that the wrong people
might lose out if social media data were not public and collectable. As he puts it,
‘‘it is crucial that access to the underlying data remains open and free, so that actors
that do not have the economic means to pay for such data, such as activist groups,
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
446 Journal of Broadcasting & Electronic Media/December 2012
consumer cooperatives or other non-profit organizations are able to benefit from its
uses’’ (2011, p. 22).
In addition to these reflections from social media monitoring actors, the appear-
ance of bodies that aim to regulate this emergent sector points to broader, industry-
wide concern with ethical questions. These include, for example, WOMMA, the
Word of Mouth Marketing Association in the US and the UK, which describes itself
as ‘‘the leading voice for ethical and effective word of mouth and social media
marketing’’ (WOMMA, 2012). WOMMA offers, amongst other things, ethics codes
and ethical resources, such as social media disclosure and privacy guides, guides
relating to honesty about return on investment, and ethical assessment tools. The
development of organizations like this and the reflections of individual practitioners
point simultaneously to differences in practice, differences in ethics, and to the
moral economy of sentiment analysis. This is another perspective that contributes
productively to our understanding of it.
Perspectives of Social Media Users
What do social media users think of social media monitoring? Do they consider
it to be more ethical for companies to inform them of their practices, which some
respondents argued should be the guiding ethical principle of sentiment analysis,
or to keep these practices from them, as the academic sentiment analyst amongst
my interviewees does? Little research has been done on this topic, and this is a gap
that needs to be filled.
There are nonetheless resources that provide us with some insight into this ques-
tion. As stated above, there has been a great deal of research exploring what social
media users think they are doing when they use social media, such as the boyd
and Marwick study already cited. Such studies enhance knowledge about what
we can take social media content and activity to represent—feeling, self, identity,
performance?—which in turn contributes to analyses of sentiment analysis practices.
Similarly, Turow and others (Turow, Feldman, & Meltzer, 2005; Turow, Hoofnagle,
King, Bleakley, & Hennessy, 2009) have carried out a number of surveys in the US
exploring consumers’ views of the online tracking activities of digital advertising
companies. These studies offer insight into attitudes to such practices, although
perspectives on social media monitoring may vary, given the specific contexts and
purposes of these media. Turow et al. (2005) argue that people reject digital data
tracking: in one study, 79% of 1500 adults agreed with the statement ‘‘I am nervous
about websites having information about me.’’ However, most other statistics from
these studies point to ignorance rather than concern. For example, the 2005 report
stated that around half of respondents did not know that Web sites are allowed to
share information with affiliates (Turow et al., 2005), and the 2009 study found that
62% falsely believe that if a Web site has a privacy policy, it means that the site
cannot share information about them with other companies (Turow et al., 2009).
Finally, Mark Andrejevic is undertaking a study of public attitudes toward measures
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 447
to regulate the collection and use of online personal information, which again may
offer some indication of social media users’ attitudes to social media monitoring,
but which, at the time of writing, is only just beginning.
Bringing Together Perspectives on Sentiment Analysis
This article is a kind of manifesto for studying sentiment analysis and other forms
of social media monitoring. Companies offering such services could be described as
intermediary social media industries, mediating social media content and activity on
behalf of their clients. Such companies play an important role in the technologically
mediated affective economy (Andrejevic, 2011), and should not be bypassed in favor
of analyses of the Facebooks, Twitters, and other major players in our wired culture.
The article proposes that a range of perspectives can usefully come together in order
to develop rich understandings of such contemporary phenomena. It has suggested
that these might include: the perspectives of sentiment analysts themselves; critical
perspectives deriving from Marxist-influenced structural analyses; a moral economy
perspective examining ethics and agency in sentiment analysis decision-making; and
the perspectives of the social media users without whose activities social media
intelligence could not exist. Bringing together these varied perspectives makes it
possible to acknowledge both the agency of social media (monitoring) actors of all
kinds, and the constraining structural conditions in which these practices take place.
Synthesizing structure and agency in this way, Gilbert (2012) suggests, ‘‘is not just
possible but necessary if we are to understand the genesis and the full complexity
of our historical moment.’’
The concerns of sentiment analysts provide insight into the empirical practicalities
of the field. Their discussions of analytical accuracy and of quantity and cleanliness
of data point to the difficulties inherent in accurately analyzing sentiments, and the
limited availability and reliability of sentiment in social media. This in turn suggests
that the spread and impact of sentiment analysis may be relatively limited and the
need for some caution in what we claim about it.
The critical perspectives of Hearn and Andrejevic to some extent represent Gilbert’s
‘‘classical Marxist denunciation of neoliberalism’’ and, along with Turow’s detailed
analysis, do the important work of drawing attention to both the capitalist structures
within which practices like sentiment analysis cannot help but take place, and the
troubling social consequences of such practices. They remind us of the importance
of being critical, and of taking a normative position that highlights the ways in which
power, control, and discrimination reproduce themselves in innovative ways within
innovative socio-technical systems.
Returning to questions of actor agency, a moral economy perspective highlights
the ethical considerations that can be identified in the decisions that are made
within sentiment analysis practices about how to access data, how to use data,
and whether or not to offer advice about intervening in social media conversations
as a result of data analysis. This article paints an ethically complex picture of how
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
448 Journal of Broadcasting & Electronic Media/December 2012
sentiment analysis gets done. Feeling-intermediaries’ histories inform their practices,
from their educational backgrounds in critical media and cultural studies, to their
experiences of various forms of ethical work. Sentiment analysis practices differ;
some are ethically informed, and some appear to be less so. Some sentiment analysts
make decisions about the ways in which they create value—for themselves and for
their clients—that are based on their values. Some of the time, these decisions limit
their own potential to create value, in ways that practitioners and companies find
ethically acceptable. Social media sentiments also differ—some are meant to be
heard, mined, analyzed, and appear to invite responses. This perspective, then,
points to the importance of being precise about which (kinds of) companies and
sentiments we are addressing in our analyses. This article has dwelt longest on this
perspective, as it is under-represented in debates on sentiment analysis to date.
The views of the social media users whose sentiments are analyzed do not
figure prominently in discussions of sentiment analysis. We simply do not know
the opinions of social media users about the monitoring of their social media
activity. Further research is needed here. The research undertaken by boyd and
Marwick (2010) and others, which aims to unearth the perspectives of social media
users themselves, offers a further perspective that can enhance our understanding
of sentiment analysis. Further research is also needed into the range of practices
carried out under the rubric of sentiment analysis, how the people making and
deploying the related systems think about their work, who their customers are, and
what is done with sentiment analysis data.
These varied perspectives on sentiment analysis attach different meanings to social
media data, sentiment, and activity, which, in turn, contribute to the political,
social, and cultural claims that are made about them, claims that may surprise
their producers, social media users. This points to a bigger question: What is social
media content (or data, sentiment, activity), and what meanings should be attached
to it? Despite widespread research on the topic, the answer to this question remains
unclear. Bringing the perspectives outlined here together, and carrying out further
empirical research with social media users and producers of various kinds of social
media intelligence might help us to arrive at clearer answers. In order to do so, in
this age of e-research, big data, and digital research methods, what is needed is
some old-fashioned and somewhat unfashionable empirical sociology.
The reflections in this article raise broader questions than those that relate to the
best ways in which to make sense of sentiment analysis. The research on which
one part of the article is based adopts an approach which acknowledges our ethical
responsibilities as researchers to enter into dialogue with the people who are the
object of our studies, to put to them some of the criticisms of their work that are
appearing in the pages of academic journals. Such dialogues might be understood
as the kinds of conversations that Gibson and Graham (2006) claim enable ‘‘the
ethical opening of persons to one another’’ (p. 135). In my research, a willingness
to reflect on ethics, to collaborate on future research into the views of social media
users and to consider what these views might mean for their operations, present
amongst some participants, might constitute such an ethical opening.
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
Kennedy/PERSPECTIVES ON SENTIMENT ANALYSIS 449
Notes
1In response to negative blog posts about poor quality customer service and fearful of losing
market share, Dell set up Direct2Dell to encourage customers to share their frustrations directly
with the company. Initially, they received many negative comments, but their responses to
these comments were seen to re-build trust with customers over time. In another incident,
painkiller Motrin released an advertisement targeting mothers carrying their babies in carriers,
which was not well received because of its flippant tone. Negative commentary spread rapidly
on Twitter, after which Motrin removed the ad and apologized. The original advert can
be found here http://www.youtube.com/watch?vDXO6SlTUBA38 and a response here http://
www.youtube.com/watch?vDLhR-y1N6R8Q
2In the US, it’s even higher, at 4.4.
3I thank Cristina Miguel for assistance with some of these interviews.
References
Ahmed, S. (2004). The cultural politics of emotion. New York, NY: Routledge.
Andrejevic, M. (2011). The work that affective economics does. Cultural Studies,25(4–5),
604–620.
Arvidsson, A. (2011). General sentiment: How value and affect converge in the information
economy. Social Science Research Network. Retrieved from http://papers.ssrn.com/sol3/
papers.cfm?abstract_idD1815031.
boyd, d. (2010, March 13). Making sense of privacy and publicity. Keynote Speech, SXSWi
(South by South West Interactive), Austin, TX. Retrieved 20 May 2010 from http://www.
danah.org/papers/talks/2010/SXSW2010.html.
Clough, P. (2003). Affect and control: Rethinking the body ‘beyond sex and gender.’ Feminist
Theory,4(3), 359–364.
Fowler, G. A. & de Avila, J. (2009, October 6). On the Internet, everyone’s a critic but they’re
not very critical. Wall Street Journal. Retrieved from http://online.wsj.com/article/SB12547
0172872063071.html
Gibson, K., & Graham, J. (2006). A postcapitalist politics. Minneapolis, MN: University of
Minnesota Press.
Gilbert, J. (2012, July 13). Moving on from market society: Culture (and cultural studies) in
a post-democratic age. Open Democracy. Retrieved from http://www.opendemocracy.net/
ourkingdom/jeremy-gilbert/moving-on-from-market-society-culture-and-cultural-studies-in-
post-democra.
Hearn A. (2008). Meat, mask, burden: Probing the contours of the branded ‘self.’ Journal of
Consumer Culture,8(2), 197–217.
Hearn, A. (2010). Structuring feeling: Web 2.0, online ranking and rating, and the digital
‘‘reputation’’ economy. Ephemera: Theory & Politics in Organisation, 10(3/4). Retrieved
from http://www.ephemeraweb.org/
Hunt, T. (2009). The whuffie factor: Using the power of social networks to build your business.
New York, NY: Crown Business.
Jenkins, H. (2008). Convergence culture: Where new and old media collide. New York, NY:
New York University Press.
Keat, R. (2000). Every economy is a moral economy. Retrieved from http://www.russellkeat.
net/research/ethicsmarkets/keat_everyeconomymoraleconomy.pdf.
Keat, R. (2011). Market economies as moral economies. Retrieved from http://www.russellkeat.
net/research/ethicsmarkets/keat_marketeconomies_moraleconomies.pdf.
Locke, C., Searls, D., & Weinberger, D. (2000). The clue train manifesto: The end of business
as usual. New York, NY: Basic Books.
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
450 Journal of Broadcasting & Electronic Media/December 2012
Marwick, W. E., & boyd, d. (2010). I tweet honestly, I tweet passionately: Twitter users, context
collapse, and the imagined audience. New Media and Society,13, 96–113.
Massumi, B. (2002). Parables of the virtual: Movement, affect, sensation. London, UK: Duke
University Press.
McRobbie, A. (1999). In the culture society. New York, NY: Routledge.
Nissenbaum, H. (2009). Privacy in context: Technology, policy and the integrity of social life.
Stanford, CA: Stanford University Press.
Nussbaum, M. C. (2003). Upheavals of thought: The intelligence of emotions. Cambridge, UK:
Cambridge University Press.
O’Neill, J. (1998). The market: Ethics, knowledge and politics. London: Routledge.
Pang, B. & Lee, L. (2008). Opinion mining and sentiment analysis. Foundations and Trends in
Information Retrieval,2(1–2), 1–135.
Sayer, A. (2004). Moral economy. Department of Sociology, Lancaster University. Retrieved
from http://www.lancs.ac.uk/fass/sociology/papers/sayer-moral-economy.pdf
Thelwall, M., Buckley, K., & Paltoglou, G. (2011). Sentiment in Twitter events. Journal of the
American Society for Information Science and Technology,62(2), 406–418.
Turow, J. (2012) The daily you: How the new advertising industry is defining your identity and
your worth. London, UK: Yale University Press.
Turow, J., Feldman, L., & Meltzer, K. (2005). Open to exploitation: American shoppers, on-
line and offline. Retrieved from http://www.annenbergpublicpolicycenter.org/Downloads/
Information_And_Society/Turow_APPC_Report_WEB_FINAL.pdf.
Turow, J., Hoofnagle, C. J., King, J., Bleakley, A., & Hennessy, M. (2009). Americans reject tai-
lored advertising and three activities that enable it. Retrieved from http://www.ftc.gov/bcp/
workshops/privacyroundtables/Turow.pdf
WOMMA (2012). Word of Mouth Marketing Association. Retrieved from http://womma.org/
main/
Zarrella, D. (2009) The social media marketing book. Cambridge, CA: O’Reilly Media.
Downloaded by [Royal Hallamshire Hospital] at 03:17 11 December 2015
