Quotatives Indicate Decline in Objectivity in U.S. Political News Tiancheng Hu1 Manoel Horta Ribeiro2 Robert West2 Andreas Spitz3 1ETH Z urich Z urich Switzerland

2025-04-29 1 0 446.49KB 12 页 10玖币

侵权投诉

Quotatives Indicate Decline in Objectivity in U.S. Political News

Tiancheng Hu1, Manoel Horta Ribeiro2, Robert West2, Andreas Spitz3

1ETH Z¨

urich, Z¨

urich, Switzerland

2EPFL, Lausanne, Switzerland

3University of Konstanz, Konstanz, Germany

tiancheng.hu@alumni.ethz.ch, manoel.hortaribeiro@epﬂ.ch, robert.west@epﬂ.ch, andreas.spitz@uni-konstanz.de

Abstract

According to journalistic standards, direct quotes should be at-

tributed to sources with objective quotatives such as “said” and

“told,” since nonobjective quotatives, e.g., “argued” and “in-

sisted,” would inﬂuence the readers’ perception of the quote and

the quoted person. In this paper, we analyze the adherence to

this journalistic norm to study trends in objectivity in political

news across U.S. outlets of different ideological leanings. We

ask: 1) How has the usage of nonobjective quotatives evolved?

2) How do news outlets use nonobjective quotatives when cov-

ering politicians of different parties? To answer these questions,

we developed a dependency-parsing-based method to extract

quotatives and applied it to Quotebank, a web-scale corpus of

attributed quotes, obtaining nearly 7 million quotes, each en-

riched with the quoted speaker’s political party and the ideolog-

ical leaning of the outlet that published the quote. We ﬁnd that,

while partisan outlets are the ones that most often use nonob-

jective quotatives, between 2013 and 2020, the outlets that in-

creased their usage of nonobjective quotatives the most were

“moderate” centrist news outlets (around 0.6 percentage points,

or 20% in relative percentage over seven years). Further, we ﬁnd

that outlets use nonobjective quotatives more often when quot-

ing politicians of the opposing ideology (e.g., left-leaning out-

lets quoting Republicans) and that this “quotative bias” is rising

at a swift pace, increasing up to 0.5 percentage points, or 25% in

relative percentage, per year. These ﬁndings suggest an overall

decline in journalistic objectivity in U.S. political news.

1 Introduction

Journalistic objectivity is the notion that news should con-

tain accurate information and not convey the personal opin-

ions or emotions of the writer (Ryan 2001; Calcutt and Ham-

mond 2011). Historically, objectivity emerged alongside the

conception of journalism as a profession (Schudson 1981)

and has shaped many of the practices and norms in mod-

ern journalism (Boudana 2011). In the context of U.S. pol-

itics, with its two major political parties, this can also be

interpreted as “equal treatment” of both parties (D’Alessio

and Allen 2006). Bias in the news could affect public opin-

ion (O’Connell 1999; Kahn and Kenney 2002) and lead to

changes in voting behavior (DellaVigna and Kaplan 2007;

Bernhardt, Krasa, and Polborn 2008).

“Absolute” objectivity has been criticized as unattainable,

as structural biases would creep into news production (Be-

harrell et al. 2009), or even as harmful, as the excessive

balance of viewpoints could create an illusion of credibil-

ity for dubious or unsupported positions (Dixon and Clarke

2013). However, amidst the fragmented media ecosystem

that emerged from the digitization of news outlets and

the algorithmic serving of content (Thurman, Lewis, and

Kunert 2019), journalism scholars have argued that objectiv-

ity has become ever more important to consumers of journal-

ism (Boudana 2011; McNair 2017). This opinion is also held

by the public worldwide, who, as of 2018, overwhelmingly

agree that news media should be unbiased in its coverage of

politics (Mitchell et al. 2018).

One of the concrete ways in which journalists have sought

to report the news objectively is through the usage of di-

rect quotes (Brooks et al. 2007; Stenvall 2008). Since jour-

nalists almost never directly observe the events they report,

using quotes lends them more reliability and factuality than

their own words (Van Dijk 1988). Furthermore, direct quotes

would let people “speak for themselves,” following one of

the golden rules of journalism (Ingram and Henshall 2012).

However, even when using a direct quote, journalistic objec-

tivity can still be compromised by the use of certain quota-

tives that relay the emotions of reporters to readers (Mencher

and Shilton 1997) or the attempt to describe the speaker’s

state of mind (Gidengil and Everitt 2003). For example, in

the direct quote

“New York is not afraid of terrorists,” boasted Rep. Jerrold

Nadler, a Democrat representing Manhattan,

the quotative (boasted) carries an illocutionary force from

the reporter that inﬂuences how the reader perceives the

quote itself, possibly distorting its original meaning (Caldas-

Coulthard 1992). Objective quotatives, like “say” or “tell,”

on the other hand, are considered neutral, as they imply lit-

tle about the presumed intent or the fashion in which the

quote was uttered (Sonoda 1997; Bell 1991).

Recent years were marked by increased political polar-

ization (Abramowitz and Saunders 2008), mistrust in me-

dia (Brenan 2022), increased negative tone by politicians

(K¨

ulz et al. 2022), and the perception that the public de-

bate around politics has become less respectful and less fact-

based (Doherty et al. 2019). Solutions to these issues are

complex, but analyzing the bias and the departure from jour-

nalistic objectivity in political news coverage can help in-

arXiv:2210.15476v2 [cs.CY] 16 May 2023

form new practices and interventions that seek to improve

the political news media ecosystem. Quotatives, in this con-

text, are a powerful instrument to measure bias. Studying

how journalists deviate from the standard usage of quota-

tives – e.g., “say” and “tell” (The Associated Press 2020) –

allows researchers to quantitatively assess adherence to jour-

nalistic objectivity (Lee 2017) and reveal biases in journal-

istic coverage of politics (Gidengil and Everitt 2003).

Present work. This paper analyzes quotatives to study ob-

jectivity and media bias in political journalism. We ask:

•RQ1 How has the usage of nonobjective quotatives

evolved in U.S. political journalism?

•RQ2 How do news outlets use nonobjective quotatives

when covering politicians of different parties?

To answer these research questions, we developed a method-

ology to extract quotatives from a large-scale news corpus.

We then performed a comprehensive study on how (and

which) quotatives are used in direct quotes from U.S. politi-

cians between 2013 and 2020, leveraging a large dataset of

quotes from English-language media linked with relevant

speaker metadata (Vaucher et al. 2021) and enriched with

the political leanings of different U.S. outlets. By counting

the usage of nonobjective quotatives like “shout” or “assert,”

we analyze how U.S. news outlets of different political incli-

nations adhere to basic journalistic objectivity principles and

how this adherence has evolved. Further, analyzing how out-

lets of different political inclinations use quotatives to talk

about politicians of different parties, we study the evolution

of quotative bias in news outlets.

Summary of ﬁndings. We ﬁnd that the usage of nonob-

jective quotatives varies across different outlet categories.

Overall, the more ideologically extreme an outlet is, the

more nonobjective quotatives it uses. However, we also ﬁnd

that centrist outlets are experiencing a signiﬁcant increase

in the usage of nonobjective quotatives over the last years

(about 0.6 percentage points, or 20% in relative percentage),

suggesting that they may be “catching up” to the more bi-

ased outlets, which are not experiencing such signiﬁcant in-

creases (RQ1). We also ﬁnd evidence of “quotative bias,”

i.e., outlets tend to use nonobjective quotatives, especially

when referring to politicians of opposing ideology. For in-

stance, left and right-leaning outlets use nonobjective quota-

tives up to 2% more often when referring to Republican and

Democrat politicians, respectively (RQ2). Last, we ﬁnd that

this quotative bias is increasing at a swift pace, increasing as

much as 0.5 percentage points per year in absolute percent-

age, or 25% in relative percentage, for left-leaning outlets,

suggesting a rapid increase in polarization (RQ1 and RQ2).

Implications. Our ﬁndings indicate a decline in journalis-

tic objectivity in U.S. political news, particularly from cen-

trist outlets. This suggests that centrist outlets may play a

role in the increasingly less respectful and fact-based debate

around politics (Doherty et al. 2019). Further, we also ﬁnd

evidence of an increasing quotative bias, which could further

erode trust in the media (Brenan 2022).

2 Background and Related Work

When a quote occurs in the news, three elements are typi-

cally involved: the source, i.e., the speaker who uttered the

quote (underlined in the examples); the quoted content itself

(in italic); and the quotative that introduces the quote (also

known as a cue, reporting verb, speech verb, or attribution

verb; in bold). Quotes can be classiﬁed as either direct, in-

direct, mixed, or pure (Cappelen and Lepore 1997), where

only the former three types are typically of concern to jour-

nalists. We give an example of a direct quote in the introduc-

tion and of mixed and indirect quotes below.

Indirect quote: Sen. Ron Wyden of Oregon, the chairman of

the Senate Finance Committee, indicated that in 2019, about

100 to 125 corporations reported ﬁnancial statement income

greater than 1B USD.

Mixed quote: Catsimatidis said he’d serve for 99 cents “be-

cause I’m a grocer.”

In direct and mixed quotes, a pair of quotation marks

are used, and we can infer that the speaker uttered the

quoted words, whereas indirect quotes may paraphrase the

speaker’s words. Therefore, journalists have the most free-

dom in word choice in indirect quotes, as they can, to some

extent, rewrite what the speaker said. In contrast, in direct

quotes, journalistic norms require them to report the quoted

words verbatim (Harry 2014).

In our work, we focus on direct quotes for the fol-

lowing reasons: Quotebank does not contain indirect

quotes (Vaucher et al. 2021); automatically extracting the

quotative in mixed quotes is technically challenging and,

in some instances, impossible as there is no quotative, e.g.,

John will not help as he has “done more for this house than

all of us combined”.

Measuring media bias. Previous work has studied me-

dia biases: how journalists’ and editors’ personal opinions,

beliefs, and ﬁnancial incentives shape what is considered

newsworthy (McCombs and Shaw 1972) and how issues are

covered (Iyengar 1994). Scholars argue that partisan media

bias can harm democracy by distorting citizens’ political

knowledge and increasing polarization (Bernhardt, Krasa,

and Polborn 2008; Boudana 2011; McNair 2017). Thus,

measuring media bias is the ﬁrst step to improving our in-

formation ecosystem (Watts, Rothschild, and Mobius 2021).

Early studies in media bias required extensive manual an-

notation. For instance, Kobre (1953) studied how the press

in Florida covered the U.S. 1952 presidential campaign by

coding the number of inches of text given to each party,

the position of pictures, etc., across hundreds of newspa-

pers. However, in recent times and with the digitization of

news, various methods have been developed to automat-

ically measure media bias (Hamborg, Donnay, and Gipp

2019). Some of these methods are audience based, mea-

suring how segregated news consumers are across outlets,

e.g., Zhou, Resnick, and Mei (2011) use votes on Digg, a

social news aggregator, to classify political articles. Others

are content based, quantifying media bias by analyzing pub-

lished content directly. For instance, Gentzkow and Shapiro

(2010) measured bias using the frequency at which outlets

reproduce partisan phrases in congressional speeches.

According to Budak, Goel, and Rao (2016), both content

and audience-based approaches suffer from distinctive limi-

tations. On the one hand, audience-based approaches do not

scale beyond outlets for which detailed readership informa-

tion can be obtained. On the other hand, content-based ap-

proaches struggle to generalize well across different types of

news and outlets, e.g., methods that try to match politicians’

speeches to news only apply to a minority of news articles,

limiting the scope of the results obtained.

Quotatives and bias. Quotatives can impact how readers

perceive a news story and the involved speakers (Geis 1987;

Just, Crigler, and Buhr 1999). For instance, Cole and Shaw

(1974) carried out an experiment in which they changed “ob-

jective” quotatives like “said” for stronger verbs like “ar-

gued” or “insisted” and asked participants to rate stories

across a variety of criteria. They found that in the modi-

ﬁed versions, stories were perceived as more exciting and

less objective, and speakers were perceived as more rash.

Through quotatives, journalists can “paint reports on speech

with any brush they like” (Geis 1987), which would not only

reveal the beliefs and preferences of the writer (Gidengil and

Everitt 2003) but also subtly inﬂuence the reader (Cole and

Shaw 1974). In this context, quotatives have been used to

measure political bias, sometimes referred to as “attribu-

tion bias.” This line of work dates from the 1960s when

Merrill (1965) studied how Time magazine used quota-

tives (among other things) when referring to U.S. Presi-

dents Kennedy, Truman, and Eisenhower. More recently, Gi-

dengil and Everitt (2003) analyzed differences in quotative

usage between male and female party leaders on Canadian

television, ﬁnding that female leaders’ speech was reported

with more negative and aggressive quotatives. With a similar

methodology, Lee (2017) studied differences in nonobjec-

tive quotatives between ofﬂine and online newspapers, ﬁnd-

ing the former to adhere better to journalistic standards.

Quote attribution and analysis. Previous work in natural

language processing has studied the problem of quote attri-

bution (see Vaucher et al. (2021) for a review), an important

task in understanding dialogue structure and developing bet-

ter conversational agents. For each quote, the goal is to ex-

tract the speaker of the quote, either at the mention or entity

level. This task is challenging as the speaker could be men-

tioned implicitly or require anaphora resolution. The task

can be further combined with entity linking to extract unique

IDs of speakers ( ˇ

Culjak et al. 2022). Most prior work, how-

ever, has not dealt with the problem of quotative extraction.

Nonetheless, several datasets annotated for attributional

relationship exist (Pareti 2012, 2016; Newell, Margolin, and

Ruths 2018) that could be considered in this context. These

datasets contain labels for the content, source, and cue for

each attributional relationship. They can be viewed as an ex-

tension of The Penn Discourse TreeBank 2.0 (Prasad et al.

2008) that provides annotation of discourse relations and ar-

gument structures. While these datasets can potentially be

used as resources for training a supervised model for quota-

tive extraction, the attributional relationship they considered

is much broader than quotation and thus not suitable for our

study.

Existing work has also analyzed quotes from different

perspectives. Niculae et al. (2015) found a systematic pat-

tern in the outlets’ quoting behavior when covering the exact

same event. Lazaridou, Krestel, and Naumann (2017) found

that a machine learning classiﬁer could reliably predict one

of two news outlets based solely on the quotes they re-

port, demonstrating media bias. Tan, Peng, and Smith (2018)

showed a declining trend of bipartisan quote coverage with

a bipartite graph of media outlets and the sentences they

quoted. K¨

ulz et al. (2022) analyzed the quotes of U.S politi-

cians between 2008 and 2020 and found a decrease in nega-

tivity during Obama’s tenure and a sudden increase starting

from Trump’s presidential primary campaign in 2015.

Relationship with prior work. In this paper, we set out to

investigate how the usage of nonobjective quotatives evolved

in U.S. political journalism (RQ1) and how it is modu-

lated by media biases (RQ2). We do so by using depen-

dency parsing to extract quotatives from a large dataset (see

Sec. 3). Our method is related to quote attribution, a prob-

lem widely studied in natural language processing, with the

key difference that previous methods aim to attribute quotes

to speakers instead of ﬁnding the quotative used. Further,

our approach is similar to previous work that derives au-

tomated media bias measurements (Budak, Goel, and Rao

2016). However, in contrast to previous work, we automate

the measurement of quotative bias instead of relying on tra-

ditional manual annotation (Cole and Shaw 1974). Due to

the scalability of our approach, we obtain results that help

further understand the political news ecosystem (see Sec. 5).

Namely, while previous work often attributes the decrease in

journalistic objectivity to the rise of partisan media (McNair

2017), we ﬁnd that centrist outlets in our dataset have sys-

tematically departed from journalistic standards.

3 Materials and Methods

3.1 Data and Data Processing

To study quotative usage across various news outlets, we

use the Quotebank dataset (Vaucher et al. 2021), a web-

scale corpus of quotes. Quotebank contains over 235 million

unique quotes, extracted from 196 million English news ar-

ticles from 377 thousand web domains between September

2008 and April 2020. We additionally obtain a list of cur-

rent and former U.S. politicians with their party afﬁliations

from Wikidata, in the same fashion as K¨

ulz et al. (2022).

We ﬁlter Quotebank to consider the period containing the

best-quality speaker attributions (May 2013 to 2020) and re-

tain only quotes from politicians on this list. In cases where

quotes are attributed to more than one speaker in Quote-

bank (which happens to 8.13% of speakers in 12.25% of the

quotes), we heuristically attribute the quote to the speaker

with the alphanumerically smallest Wikidata identiﬁer. We

validate speaker attribution in our ﬁltered dataset on a manu-

ally annotated sample of 100 quotes and ﬁnd that combining

the speaker names provided by Quotebank with this heuris-

tic yields 86% accuracy in identifying the correct ID.

To ensure the validity of our ﬁndings, we preprocess

Quotebank as depicted in Figure 1. We 1) use heuristics to

retain only direct quotes; 2) extract quotatives and remove

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

QuotativesIndicateDeclineinObjectivityinU.S.PoliticalNewsTianchengHu1,ManoelHortaRibeiro2,RobertWest2,AndreasSpitz31ETHZ¨urich,Z¨urich,Switzerland2EPFL,Lausanne,Switzerland3UniversityofKonstanz,Konstanz,Germanytiancheng.hu@alumni.ethz.ch,manoel.hortaribeiro@ep.ch,robert.west@ep.ch,andreas.spitz@un...

展开>> 收起<<

Quotatives Indicate Decline in Objectivity in U.S. Political News Tiancheng Hu1 Manoel Horta Ribeiro2 Robert West2 Andreas Spitz3 1ETH Z urich Z urich Switzerland.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Quotatives Indicate Decline in Objectivity in U.S. Political News Tiancheng Hu1 Manoel Horta Ribeiro2 Robert West2 Andreas Spitz3 1ETH Z urich Z urich Switzerland

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: