witp workbench notes - weblyzard technologyfrom september to december 2004. the project drew upon...

Journal of Information Technology & Politics, Vol. 5(1) 2008Available online at http://jitp.haworthpress.com

© 2008 by The Haworth Press. All rights reserved.doi:10.1080/19331680802149582 121

WITP

WORKBENCH NOTES

An Automated Approach to Investigating the Online Media Coverage of U.S. Presidential Elections

Scharl and Weichselbraun Arno ScharlAlbert Weichselbraun

ABSTRACT. This paper presents the U.S. Election 2004 Web Monitor, a public Web portal thatcaptured trends in political media coverage before and after the 2004 U.S. presidential election. Devel-oped by the authors of this article, the webLyzard suite of Web mining tools provided the requiredfunctionality to aggregate and analyze about a half-million documents in weekly intervals. The studypaid particular attention to the editorial slant, which is defined as the quantity and tone of a Web site’scoverage as influenced by its editorial position. The observable attention and attitude toward the can-didates served as proxies of editorial slant. The system identified attention by determining the fre-quency of candidate references and measured attitude towards the candidate by looking for positive andnegative expressions that co-occur with these references. Keywords and perceptual maps summarized themost important topics associated with the candidates, placing special emphasis on environmental issues.

KEYWORDS. U.S. presidential elections, media monitoring, Web mining, natural languageprocessing, semantic orientation, keyword analysis

Arno Scharl is the Vice President of MODUL University Vienna where he also heads the Department ofNew Media Technology (www.modul.ac.at/nmt). Prior to this appointment, he was a Key Researcher at theAustrian Competence Center for Knowledge Management, held professorships at Graz University of Tech-nology and the University of Western Australia, and had joined Curtin University of Technology and theUniversity of California at Berkeley as a Visiting Fellow. His current research projects focus on the integra-tion of semantic and geospatial technology, human-computer interaction, computer-mediated collaboration,ontology learning, and the various aspects of environmental online communication.

Albert Weichselbraun is an Assistant Professor at the Department of Information Systems and Operations atVienna University of Economics and Business Administration (ai.wu-wien.ac.at). After completing two Master’sdegrees in Economics and Chemical Engineering, his doctoral thesis focused on ontology-based text classifica-tion. Dr. Weichselbraun leads the technical development of webLyzard (www.webLyzard.com) and the IDIOMProject (www.idiom.at) with a special focus on the analytical methods involved. His current research focuses onontology evolution and learning, text mining and the application of semantic technologies to information retrieval.

The IDIOM Research Project (Information Diffusion across Interactive Online Media; www.idiom.at) isfunded by the Austrian Ministry of Transport, Innovation & Technology (BMVIT) and the AustrianResearch Promotion Agency (FFG) within the strategic objective FIT-IT Semantic Systems (www.fit-it.at).The authors would like to thank Arinya Eller and Jamie Murphy, as well as the special issue’s editors andanonymous reviewers for their valuable feedback and suggestions during the preparation of this article.

Address correspondence to: Prof. Arno Scharl, Department of New Media Technology, MODUL UniversityVienna, Am Kahlenberg 1, 1190, Vienna, Austria (E-mail: [email protected]).

122 JOURNAL OF INFORMATION TECHNOLOGY & POLITICS

Representative democracy offers significantpossibilities for exploiting information networks(Holmes, 2002), but there is little agreement onthe specific impacts of these networks. Mostagree that new media represent additional chan-nels for political campaigners to project theirmessages and solicit feedback from citizens(Stromer-Galley & Foot, 2002). Proponents praisethe potential of information networks to increasethe accessibility of information, encourage partici-patory decision-making, and facilitate communi-cation with policy officials and like-mindedcitizens. From the citizens’ perspective, interac-tive Web technologies enable people to find polit-ical soul mates and search for their personal truthsonline (Gibbs, 2004). From the campaigners’ per-spective, disseminating information directly orthrough online news media helps create onlinevisibility and influence public opinion.

Most attempts to monitor the campaignperformance of presidential candidates focus onpublic opinion, which is influenced by the con-sumption of media products (Druckman & Parkin,2005). The analysis of political communication,however, should include both the consumption aswell as the production of media products(Howard, 2003). Monitoring information net-works to study political campaigns provides acomplementary source of empirical data and awindow into the evolving concept of electronicdemocracy (Dutton, Elberse, & Hale, 1999).

A Pew/Internet survey (Horrigan, Garrett, &Resnick, 2004) found that four out of ten U.S.Internet users aged 18 or older accessed politicalmaterial during the 2004 presidential campaign,up 50% from the 2000 campaign. For politicalnews, more than two-thirds of American broad-band users and over half of dial-up users seekWeb sites of national news organizations. Inter-national news sites are the second most popularcategory at 24% and 14%, respectively. Thisshows that traditional media extend their domi-nant position to the online world, and that theirWeb sites reflect the majority of political con-tent that the average Internet user accesses.Therefore, their communication strategiesaffect democratic processes regardless ofwhether authors control editorial slant deliber-ately (Fico & Cote, 1999) or whether readersrecognize the slant (Druckman & Parkin, 2005).

Nearly one third of respondents in a 2002 Stan-ford survey regarded bias relevant for judgingthe credibility of online news (Fogg et al.,2002). Most Americans claim to prefer unbi-ased news sources (Horrigan et al., 2004), asbiased media coverage can polarize groups(Sunstein, 2004) and degrade the climate ofpublic discourse (Horrigan et al., 2004). Conse-quently, many organizations explicitly hold upfairness and balance in reporting (Fico & Cote,1999), although they are usually free to choosewhich candidate to emphasize and how to inter-pret current events (Wayne, 2001).

News media are the most influential stake-holder groups that report or comment on polit-ical campaigns; but, they are not the only one.When corporations and advocacy organizationsembrace networked information technology,intentionally or not, they also influence demo-cratic processes. To capture and understand theinfluence of electronic content published by dif-ferent stakeholder groups, the U.S. Election 2004Web Monitor (www.ecoresearch.net/election2004) gathered and annotated the online cover-age of the presidential candidates (Figure 1). Itpaid particular attention to the editorial slant ofnews media and the diverging perceptions ofenvironmental advocacy organizations and themost influential US corporations.

This article presents computational methodssuited for processing very large documentrepositories. The sheer volume of informationin such repositories and the limits of humancognition preclude manual approaches to ana-lyzing online media coverage and measuringeditorial slant. Therefore, the U.S. Election2004 Web Monitor automatically identifiedattention and attitude toward each candidateand provided additional means for an in-depthinvestigation of the underlying documentrepository: keyword lists summarizing the mostimportant topics associated with a candidate, aquery interface to access and sort sentenceswith candidate references, and special reportsthat aggregate and visual domain-specific cov-erage with a special emphasis on environmentalissues. After presenting the system’s majorcomponents and results, the article concludeswith an outlook on promising future researchavenues, introducing the 2008 edition of the

Scharl and Weichselbraun 123

Election Monitor and outlining how models ofinformation diffusion could help explain theinfluence of Web content on public opinionduring political campaigns.

METHODOLOGY

The volume and dynamic nature of Webdocuments complicate testing the assumptionof organizational bias. To address this chal-lenge, the U.S. Election 2004 Web Monitormirrored 1,153 Web sites in weekly intervalsfrom September to December 2004. Theproject drew upon the Newslink.org,Kidon.com, and ABYZNewsLinks.com direc-tories to compile a list of the most influentialnews media sites—42 from the United Statesand 72 from four other English-speaking coun-tries: Canada, United Kingdom, Australia, andNew Zealand. The study also included 39environmental advocacy organizations operat-ing either nationwide within the United Statesor internationally, and all the Web sites of theFortune 1000 (the largest U.S. corporationsranked by revenue).

A crawling agent mirrored these Web sitesby following their hierarchical structure until

reaching 50 megabytes of textual data for newsmedia and 10 megabytes for commercial andadvocacy sites. These limits helped comparesites of heterogeneous size and reduce the dilu-tion of top-level information by articles foundin lower hierarchical levels.

In the literature, such a collection of recordedcontent used for descriptive analysis is oftenreferred to as a corpus. This research investigatedand visualized regularities in such corpora fromthree different samples of Web sites by applyingmethods from corpus linguistics and textual statis-tics (Biber, Conrad, & Reppen, 1998; Lebart,Salem, & Berry, 1998; McEnery & Wilson,2001). The quantitative textual analysis necessi-tated three steps to yield a useful machine-readable representation (Lebart et al., 1998):

(a) The first step converted hypertext docu-ments into plain text—that is, processingthe gathered data and eliminatingmarkup code and scripting elements.

(b) The second step segmented the textualchain into minimal units by removingcoding ambiguities such as punctuationmarks, the case of letters, hyphens, andpoints in abbreviations. With littlefluctuation across sampling points

FIGURE 1. U.S. Election 2004 Web Monitor one week before the election (October 27, 2004).


between September and December2004, this process yielded up to a half-million documents in weekly intervals,comprising about 125 million wordsin 10 million sentences. The systemidentified and removed redundant seg-ments such as headlines and news summa-ries whose appearance on multiple pageswould otherwise distort frequency counts.

(c) The third step identified groups of iden-tical units and counted their occurrences,thereby creating an inventory of words(“word list”) or multi-word units ofmeaning (Danielsson, 2004). The fre-quency of candidate references pre-sented in the following section is basedon such a word list, which typically usesdecreasing frequency of occurrence asthe primary sorting criterion and lexico-graphic order as the secondary criterion.

FREQUENCY OF CANDIDATE REFERENCES (ATTENTION)

Media coverage and public recognition gohand-in-hand (Wayne, 2001), with strong cor-relations between the attention of news mediaand both public salience and attitudes towardpresidential candidates (Kiousis & McCombs,2004). The U.S. Election 2004 Web Monitor

calculated attention as the relative number ofreferences to a candidate. To determine refer-ences to candidates or environmental topics, apattern-matching algorithm considered commonterm inflections while excluding ambiguousexpressions. Only identifying occurrences ofgeorge w. bush, for example, would ignoreequally valid references to president bush andgeorge walker bush. Yet, a general query forbush would fail to distinguish the president’slast name from references to wilderness areasor woody perennial plants. A detailed list of theterms used to measure the number of candidatereferences is available online at www.ecore-search.net/election2004/candidates.

With regard to the number of documentsmirrored, the Web sites of Fortune 1000 com-panies represented the largest sample withabout 330,000 documents, followed by newsmedia (180,000) and environmental organiza-tions (16,000). As expected, the corporate sec-tor had published relatively few articles on thepresidential candidates—yielding only 484 rel-evant documents (0.15% of total documents) ascompared to 3,834 documents from news media(2.09%) and 709 documents from environmen-tal advocacy organizations (4.50%). The com-parative statistics in Figure 2 highlight thedifferences between samples. In contrast to thefrequently updated coverage of news media,which fluctuated considerably from week to

FIGURE 2. Attention of U.S. News Media, the Fortune 1000, and Environmental Organizationstoward the U.S. Presidential Candidates between September and November 2004.


week, the textual data mirrored from the Websites of Fortune 1000 companies and environ-mental advocacy organizations show a morestatic distribution.

In the week before the election, 58% of themedia references mentioned George W. Bushand Dick Cheney, down one percentage pointfrom the preceding week (see Figure 1).Thirty-nine percent reported on John Kerryand John Edwards. Across all three samples,the Republican candidates dominated the cov-erage, while the independent team of RalphNader and Peter Camejo garnered less than 5%of the attention. The Republican dominancedoes not necessarily represent editorial bias.Incumbents who run for office serve a dual roleas candidates and government officials. Jour-nalists often do not report on the campaign, buton operative decisions and newsworthy day-to-dayactivities. The same phenomenon could beobserved with data from the Fortune 1000 com-panies and environmental organizations, whichdedicated over 80 percent of their coverage tothe incumbent and his running mate. Follow-upstudies should devise automated methods todistinguish general from campaign-specificcoverage in order to investigate how the dualrole of incumbents impacts media attention.

SEMANTIC ORIENTATION (ATTITUDE)

A conclusive measure of organizational biasin media coverage has been elusive (Sutter,2001), but it is essential when investigatingtrends and differing perceptions of interestgroups. The ever-increasing amount of articlesand the limits of human cognition prevent ana-lysts from maintaining a running tally of newsmedia coverage of the candidates during a presi-dential campaign (Watts, Domke, Shah, & Fan,1999). Therefore, the US Election 2004 WebMonitor required an automated method to deter-mine the semantic orientation toward a candidate(Scharl, Pollach, & Bauer, 2003). Calculating thefrequency of candidate references, as described inthe preceding section, only represents a partialanswer to this problem since it disregards the

specific context of these references (Yi, Nasukawa,Bunescu, & Niblack, 2003).

The chosen method is based on the notionthat there is a conceptual connection betweenwords and their adjacent text (Giora, 1996).The semantic orientation toward a target termwithin a sentence is calculated by measuringthe distance (dts) between the target term (t) anda predefined list of sentiment words (s) knownto have positive or negative connotations(Scharl et al., 2003). This list was taken fromthe tagged dictionary of the General Inquirercontaining 4,400 positive and negative senti-ment words (Stone, Dunphy, Smith, & Ogilvie,1966). To a large extent, the validity of thisapproach depends on the size of the tagged dic-tionary. It is therefore essential that allinstances of the sentiment words in the corpusare included in the analysis and not just theirbase forms. Therefore, reverse lemmatizationwas used to add about 3,000 words to the dic-tionary by considering plurals, gerunds, pasttense suffixes, and other syntactical variations(for example, manipulate → manipulates,manipulating, manipulated).

Calculated for each sentence separately, thesemantic orientation (SO) represents the sum ofthe tagged words’ semantic charges (SOs)divided by their distance to the target termbased on the following equation—the exponentλ determines the influence of the distance oncalculating the semantic orientation:

Two sentences from the New Zealand Heraldand the St. Petersburg Times containing refer-ences to oil prices (below) exemplify the differ-ence between positive and negative coverage. Thesentiment charges from the underlined words ofthe tagged dictionary are aggregated to identifyeach sentence’s semantic orientation. In thisexample, a λ of zero (no influence) implies thatthe semantic orientation equals the sum of the sen-timent charges according to the tagged dictionary.

• “US stocks rallied(+1.0) Wednesday,boosted(+1.0) by shares(+1.0) of health and

SOSO

ds

ts

= ∑ λ


defense(−1.0) companies(+0.22) that are seenbenefiting(+1.0) from the re-election ofPresident George W. Bush, but higher oilprices checked advances(+ 0.87)” (NEW

ZEALAND HERALD). ↑ (+ 4.09)• “The dollar hit(−1.0) its lowest(−0.98) level in

more than eight months against the EuroThursday, falling(−0.05) sharply on worries(−1.0)about the economic effects of rising oil pricesand expectations of continued trade andbudget deficits(−1.0) in President Bush’s sec-ond term” (ST. PETERSBURG TIMES). ↓ (−4.03)

The presented method focuses on the lexis oftext and is therefore most appropriate for largedocument archives. Full grammatical parsing ofthe textual data may deliver superior results forindividual sentences—for example, in the caseof sarcasm or complex sentence structures—butsuffers from poor scalability and thus was notcomputationally feasible for the U.S. Election2004 Web Monitor.

Acknowledging the lack of an objectivestandard by which to assess bias, previousresearch recommends focusing on relativecomparisons of coverage (Druckman & Parkin,2005). In line with these recommendations, theUS Election 2004 Web Monitor did not inter-pret a semantic orientation of zero as neutralcoverage, but rather investigated the differ-ences between candidates. The weekly statistics

of Figure 3 show that media coverage initiallyfavored the Republican Party candidates, butthat their Democratic contenders gained ground inSeptember 2004. Kerry’s performance in the firsttelevised debate accelerated these gains in mediaattitude and was followed by a tight race betweenthe two teams in the four weeks preceding theelection. The re-election of George W. Bushunderstandably widened the gap, considering thepositive connotation of terms such as winning andvictory. Sentiment toward Ralph Nader remainedremarkably positive up until the election. Since hereceived very little attention, it can be concludedthat the few outlets reporting about him were sup-porting his environmental and consumer causes,and that his opponents did not deem his candidacyrelevant enough to divert campaign resourcesfrom confronting their main competitors.

While a narrow margin decided the U.S. presi-dential elections in 1996 and 2000, differences inthe candidates’ positions became more pro-nounced in 2004 and the political deliberationmore partisan (Weisberg, 2004). Partisans tend toperceive mass media content as biased againsttheir ideological vantage point (Schmitt, Gunther,& Liebhart, 2004). Explanations for this hostilemedia effect range from selective recall (preferen-tially remembering hostile content), selectivecategorization (perceiving the same content dif-ferently), and conflicting standards (consideringhostile content as invalid or irrelevant).

FIGURE 3. Attitude of U.S. News Media, the Fortune 1000, and Environmental Organizationstowards the U.S. Presidential Candidates between September and November 2004.


Claims of a liberal bias in news coveragehave become quite common in recent U.S. pres-idential campaigns. In 1996, for example,Republican Party candidate Bob Dole blamednews media coverage as the reason for laggingbehind Democratic incumbent Bill Clinton(Watts et al., 1999). The empirical evidence ofthe US Election 2004 Web Monitor contradictsthis notion of “liberal media” and demonstratesthat the study of organizational bias requires amore differentiated approach. It does not sufficeto classify any view that does not comport witha conservative ideological viewpoint as “lib-eral” (Alterman, 2003). Clearly favoring theRepublican candidates, the attention and atti-tude data on the 2004 U.S. election allows twointerpretations: (a) Either mainstream mediacoverage is more sympathetic to conservativecauses, or (b) the incumbent benefits from ajournalistic inclination toward reporting aboutgovernment officials and their “spin” on events(Sutter, 2001). Analyzing elections with liberalincumbents is a promising area for futureresearch to investigate these interpretations.

Compared to the news media, industryand environmental advocacy organizationsshowed a more pronounced articulation oftheir political views. While environmentalorganizations tended to criticize the environ-mental record of George W. Bush (particu-larly abandoning the Kyoto Treaty ratificationand reducing air pollution controls throughthe Clear Skies Act), Fortune 1000 companiespresented the Republican Party candidatesmost favorably. Given the growth of corporateownership of news media and their reliance onrevenues from corporate advertising, the con-servative position of Fortune 1000 companiesmight influence the fundamental rules of newsproduction and help explain the observedmedia bias.

Organizational bias is often specific to acertain topic or domain. Many journalists holdliberal social views, for example, but conserva-tive economics views (Croteau, 1998). The fol-lowing two sections demonstrate howautomated Web site analysis can account forthese differences and provide a more nuancedview of candidate coverage—either on theaggregated level by using keywords to summarize

the unfolding of events, or by providing effec-tive means to query individual sentences thatreport on candidate activities.

KEYWORD ANALYSIS

Complementing measures of attention andattitude toward a candidate, keywords groupedby political party and geographic region high-light the most popular topics in a particularweek. (Keywords are computed by locatingthem in a target corpus and comparing their fre-quency with a reference distribution from ausually larger corpus of text.) To identify thetopics that each stakeholder group associatedwith the candidates, the U.S. Election 2004Web Monitor compared the term frequency dis-tribution in sentences mentioning a candidate(target corpus) with the term frequency distri-bution in the entire sample (reference corpus).The system used a chi-square test with Yates’correction for continuity (Yates, 1934) to ana-lyze the significance of co-occurrence by com-paring term frequencies between target andreference corpus. The ratio between both cor-pora in conjunction with a term’s total counts inthe reference corpus provides the expected fre-quency (Ei). Comparing this value with theobserved frequency (Oi) and applying the chi-square test with Yates’ correction yields thesignificance of the deviation between expectedand observed frequencies.

Table 1 exemplifies the process for a hypotheti-cal target corpus one-fourth the size of the ref-erence corpus, containing all the sentences withat least one reference to George W. Bush. Com-mon words such as prepositions, articles, andconjunctions occur frequently, but in line withexpectations assuming equal term distributionsin the reference corpus and the target corpus.The term Iraq, by contrast, appears three timesas often as expected and thus appears to be agood descriptor of candidate coverage. Termfrequency thresholds prevent the inclusion of

cYates| |2

2

1

0 5=

− −

=∑ ( . )O E

Ei i

ii

N


rare expressions or typographical errors such asWahington [sic].

Table 2 summarizes keywords that U.S.news media associated with the candidates andtheir running mates in the week preceding theelection—that is, those terms that were signifi-cantly over-represented in sentences reportingon a specific candidate. The list ranks keywordsby decreasing significance as computed by thechi-square test. To avoid outliers, the list onlyconsiders nouns with at least 100 occurrencesin the reference corpus.

Keyword analysis distills information fromgigabytes of raw textual data and tells thereader a concise story of what happened on thecampaign trail in a given week. Table 2 shows

that the television debates between the majorcandidates and their running mates remainedtopical up until the election. The war on terrorismand persistent problems in dealing withinsurgents in Iraq dogged Bush, while his chal-lenger’s service in Vietnam continued tooccupy the media. Vice-President and formerCEO of Halliburton, Cheney, was busy travelingto Pensacola, Wyoming, and Wilmington andaddressing media questions about his wifeLynne and his lesbian daughter Mary. A speechby former President Clinton, joining Kerry inhis first appearance after undergoing heart sur-gery, reminded undecided voters of more pros-perous times. At the same time, actor AshtonKutcher hit the campaign trail for John

TABLE 1. Example of Keyword Calculation Including Word Frequency in the Reference Corpus, Expected and Observed Word Frequencies in the Target Corpus, and Chi-Square Significance

Word AND THE WAR IRAQ WASHINGTON WAHINGTON [SIC]

Reference corpus 300 596 200 100 300 4Target corpus (expected) 75 149 50 25 75 1Target corpus (observed) 72 158 127 75 200 2Chi-square significance 0.09 0.44 123.86 113.48 197.73 0.33

TABLE 2. Candidate Keywords Based on U.S. News Media Sites as of October 27, 2004

Republicans Democrats Independents

George Bush Dick Cheney John Kerry John Edwards Ralph Nader Peter Camejo

DEBATE LYNNE NOMINEE CAROLINA BALLOT OPINION5510 | 1650.44 205 | 1935.06 623 | 3516.29 1363 | 1842.84 1333 | 10080.58 1939 | 1893.62CHALLENGER DAUGHTER VIETNAM RUNNING PERCENT RUNNING547 | 1558.19 2090 | 1624.17 1475 | 2652.35 3462 | 1701.16 16270 | 2359.83 3462 | 400.81

IRAQ HALLIBURTON CHALLENGER DEBATE CANDIDACY ELECTORS19156 | 1528.19 377 | 1262.01 547 | 2266.91 5510 | 925.14 197 | 1212.04 116 | 72.77

GORE DEBATE DEBATE NOMINEE ADVOCATE COMMONWEALTH1403 | 1497.44 5510 | 1150.93 5510 | 2109.72 623 | 892.68 356 | 811.17 133 | 63.34

WAR LESBIAN MASSACHUSETTS GEPHARDT SUPREME BALLOT14957 | 1288.99 404 | 117.84 1260 | 1602.77 154 | 406.09 1665 | 574.54 1333 | 54.72

SPEECH RUMSFELD NOMINATION IOWA GORE RESPONDENTS2069 | 1076.86 1278 | 451.90 585 | 1373.27 1489 | 390.60 1403 | 429.32 225 | 37.04

NOMINEE PENSACOLA WAR ASHTON PARTY ENDORSEMENT623 | 839.79 98 | 446.17 14957 | 1133.62 154 | 289.11 8624 | 386.75 323 | 25.50

GUARD RALLY RIVAL NORTH PETITION NOMINEE1771 | 695.87 1342 | 392.09 676 | 889.56 5789 | 281.62 139 | 382.75 623 | 12.75

HUSSEIN WYOMING SPEECH OPTIMISM COURT BATTLEGROUND2609 | 578.92 201 | 365.91 2069 | 592.17 293 | 278.36 5376 | 377.01 681 | 11.59TERRORISM WILMINGTON CLINTON TRAIL PENNSYLVANIA BALANCE2613 | 515.38 115 | 269.22 3499 | 498.88 1372 | 237.17 1503 | 347.22 1437 | 5.00

The values indicate the total number of occurrences in the references corpus and the chi-square significance.


Edwards, senator from North Carolina and run-ning mate of John Kerry. Although the SupremeCourt refused his candidacy in Pennsylvaniaover invalid nominating petitions, Ralph Naderwas on the ballot in more than 30 states.Articles about him reiterated controversies overvote-splitting in the previous election and theSupreme Court’s decision to end the Bush vs.Gore recounts in December 2000.

The results support claims that personalities,strategies, and campaign events dominate oversubstantive policy and governance issues—apossible reason for the average voter’s limitedinterest in and knowledge about political pro-cesses (Haswell, 1999; Watts et al., 1999;Wayne, 2001). The news media’s ongoing huntfor currency and novelty helps explain thisobservation, as the candidates’ positions tend toremain constant throughout a campaign andtherefore lack newsworthiness. The prolifera-tion of polling (Traugott & Lavrakas, 2000) andmarket forces that demand media formats withsignificant entertainment value (Iyengar,Norpoth, & Hahn, 2004) also contribute to thelack of substantive information.

COVERAGE OF ENVIRONMENTAL ISSUES

The Special Reports section of the U.S. Elec-tion Monitor 2004 allowed users to investigatethe candidates’ positions on environmentalissues and shows how these positions were por-trayed by the media. This component used anannotated archive of Web data also referred toas contextualized information space (Scharl,2007). Annotation services added several typesof metadata (topic, time stamp, semanticorientation) to the documents stored in thisarchive. A topic hierarchy with three distinctlayers defined the report type (global warmingand energy policy), relevant subtopics (renew-able energy, fossil fuels, nuclear energy, green-house gases, climate effects), and the respectivetext patterns (regular expressions) that indicatethe topics.

Focusing on energy policy, one such analysisinvestigated Web coverage of renewableenergy, fossil fuels, and nuclear power—crucial

issues considering the global environmentalimpact of U.S. energy policy decisions. TheElection Monitor’s Web site allows users to listsentences containing both candidate referencesand energy-related terms and to sort these sen-tences by semantic orientation. It also providesaggregated data by summarizing the prominenceof energy terms among international news mediasites in the perceptual map of Figure 4. Such mapsaim to provide a high-level view on news mediacoverage of related topics within a specificdomain, shedding light on the relative importanceof these topics as perceived by the media. Such acondensed representation of the information spaceis a valuable tool to classify organizations. Gener-ated repeatedly over time, perceptual maps allowanalysts to track trends in online media coverageand quickly identify unusual developments(Scharl, 2004).

The diagram’s distribution of data points wascalculated by using the correspondence analysismodule of SPSS (Statistical Package for theSocial Sciences; www.spss.com) to process thetable of term frequencies as of September 17,2004. In this snapshot analysis, co-occurringterms appear close to each other in the compu-tationally created two-dimensional space. Thecircular and rectangular markers representenergy-related concepts and media sites,respectively.

When interpreting the diagram, it is impor-tant to note that (a) term frequency rather thanmedia attitude determines the position of datapoints, (b) axes of artificially generated spaceshave no units of measurement, and (c) cross-proximities between row and column points(that is, terms and media sites) originate fromdifferent initial spaces. Thus, the position of aparticular point may be interpreted only withrespect to the other dimension’s set of points,but not with a single point of the other dimen-sion (Lebart et al., 1998).

The distribution of data points in Figure 4shows a tripolar structure: fossil fuels in theupper left, renewable energies in the upperright, and nuclear energy near the bottom of thediagram. The diagram illustrates news organi-zations with distinct content—Fox News andTime, for example, with their coverage ofnuclear energy and wind energy, respectively.


Geographic differences are also apparent. Fos-sil fuels align with many Australian media,reflecting the country’s richness in mineralresources. Most business publications congre-gate around fossil fuels as well, BusinessWeek being a notable exception. Referencesto fuel cells and hydrogen often appeartogether. Their potential use with various energysources explains their slightly isolated position.Combining the energy carrier hydrogen withfuel cell conversion technology yields high effi-ciency and low pollution in applications such aszero-emission vehicles, energy storage, andportable electronics.

While the 2004 edition of the ElectionMonitor only provided a limited number ofreports suggested by domain experts, the2008 edition introduced in the following sectionuses a more flexible framework based on auto-mated ontology extension (Liu, Weichselbraun,Scharl, & Chang, 2005) that identifies relevant

topics and their relations based on unstruc-tured textual data stored in the knowledgerepository.

CONCLUSION AND OUTLOOK

The U.S. Election 2004 Web Monitor pro-vided weekly snapshots of international Webcoverage. It revealed the editorial slant of Websites by measuring attention and attitude towardthe U.S. presidential candidates. The empiricalevidence contradicts the notion of liberalmedia, as the coverage between September andDecember 2004 clearly favored the Republicancandidates. Since organizational bias is oftenspecific to certain topics or domains, the secondpart of this paper introduced more nuanced rep-resentations of candidate coverage: keywordlists and perceptual maps that aggregate semanticassociations within the knowledge base, and

FIGURE 4. Perceptual map of energy terms and international news media.


query interfaces for accessing this knowledgebase on a granular level.

As the coverage of the 2004 U.S. presidentialelection might be atypical in several ways,future research should replicate and extend thisstudy in the context of other campaigns. Forthis purpose, a revised Web portal available atwww.ecoresearch.net/election2008 investigatesthe 2008 U.S. presidential election. Aside frommajor revisions of the underlying annotationservices, the new portal adds a social aspect tothe analysis of aggregated content by invitingusers to cast their votes online. The portal alsofeatures just-in-time information retrievalagents, as well as visual navigational aids suchas information landscapes, ontology views, tagclouds, and geospatial maps (Scharl et al., 2007).

The U.S. Election 2008 Web Monitor is partof the IDIOM research project [Information Dif-fusion across Interactive Online Media(www.idiom.at)], which aims to model theproduction, propagation, and consumptionof electronic content to address four researchquestions: How redundant are online newsmedia, and what are the major determinants ofinformation flows within electronic networks?Can existing models of information propagationsuch as hub-and-spoke, syndication, and peer-to-peer adequately explain these information flows?How does Web content influence public opinionduring political campaigns, and what are appro-priate methods to measure and model the extent,dynamics, and latency of this process? Finally,which content placement strategies increase theimpact on the target audience and support self-reinforcing propagation among individuals?

REFERENCES

Alterman, E. (2003). What liberal media? The truth aboutbias and the news. New York: Perseus Books.

Biber, D., Conrad, S., & Reppen, R. (1998). Corpus lin-guistics—investigating language structure and use.Cambridge, England: Cambridge University Press.

Croteau, D. (1998). Challenging the “liberal media” claim.Extra!, 11(4), 4–9.

Danielsson, P. (2004). Automatic extraction of meaning-ful units from corpora. International Journal of CorpusLinguistics, 8(1), 109–127.

Druckman, J. N., & Parkin, M. (2005). The impact ofmedia bias: How editorial slant affects voters. TheJournal of Politics, 67(4), 1030–1049.

Dutton, W. H., Elberse, A., & Hale, M. (1999). A casestudy of a Netizen’s guide to elections. Communica-tions of the ACM, 42(12), 48–54.

Fico, F., & Cote, W. (1999). Fairness and balance in thestructural characteristics of newspaper stories on the1996 presidential election. Journalism and Mass Com-munication Quarterly, 76(1), 124–137.

Fogg, B. J., Soohoo, C., Danielson, D., Marable, L.,Stanford, J., & Tauber, E. R. (2002). How do peopleevaluate a Web site’s credibility? Palo Alto, CA: Stan-ford University, Consumer Web Watch and SlicedBread Design.

Gibbs, N. (2004). Blue truth, red truth. Time, 164(13),24–33.

Giora, R. (1996). Discourse coherence and theory of rele-vance: Stumbling blocks in search of a unified theory.Journal of Pragmatics, 27, 17–34.

Haswell, S. (1999). The news media’s role in electioncampaigns: A big audience or a big yawn? AustralianJournalism Review, 21(3), 165–180.

Holmes, N. (2002). Representative democracy and theprofession. Computer, 35(2), 118–120.

Horrigan, J., Garrett, K., & Resnick, P. (2004). The Inter-net and democratic debate. Washington, DC: PewInternet & American Life Project.

Howard, P. N. (2003). Digitizing the social contract: Pro-ducing American political culture in the age of newmedia. The Communication Review, 6, 213–245.

Iyengar, S., Norpoth, H., & Hahn, K. S. (2004). Consumerdemand for election news: The horserace sells. TheJournal of Politics, 66(1), 157–175.

Kiousis, S., & McCombs, M. (2004). Agenda-settingeffects and attitude strength—political figures duringthe 1996 presidential election. CommunicationResearch, 31(1), 36–57.

Lebart, L., Salem, A., & Berry, L. (1998). Exploring tex-tual data (vol. 4). Dordrecht, Holland: Kluwer Aca-demic Publishers.

Liu, W., Weichselbraun, A., Scharl, A., & Chang, E.(2005). Semi-automatic ontology extension usingspreading activation. Journal of Universal KnowledgeManagement, 0(1), 50–58.

McEnery, T., & Wilson, A. (2001). Corpus linguistics (2nded.). Edinburgh, Scotland: Edinburgh University Press.

Scharl, A. (2004). Web coverage of renewable energy. InA. Scharl (Ed.), Environmental Online Communication(pp. 25–34). London: Springer.

Scharl, A. (2007). Towards the Geospatial Web: Mediaplatforms for managing geotagged knowledge reposi-tories. In A. Scharl & K. Tochtermann (Eds.), Thegeospatial Web—How geobrowsers, social softwareand the Web 2.0 are shaping the network society(pp. 3–14). London: Springer.


Scharl, A., Pollach, I., & Bauer, C. (2003). Determining thesemantic orientation of Web-based corpora. In J. Liu, Y.Cheung, & H. Yin (Eds.), Intelligent data engineeringand automated learning, 4th international conference,IDEAL-2003, Hong Kong (Lecture Notes in ComputerScience, Vol. 2690) (pp. 840–849). Berlin: Springer.

Scharl, A., Weichselbraun, A., Hubmann-Haidvogel, A.,Stern, H., Wohlgenannt, G., & Zibold, D. (2007).Media watch on climate change: Building and visualiz-ing contextualized information spaces. Paper presentedat the 6th International Semantic Web Conference(ISWC-2007), Busan, Korea.

Schmitt, K. M., Gunther, A. C., & Liebhart, J. L. (2004).Why partisans see mass media as biased. Communica-tion Research, 31(6), 623–641.

Stone, P. J., Dunphy, D. C., Smith, M. S., & Ogilvie, D. M.(1966). The General Inquirer: A computer approach tocontent analysis. Cambridge, MA: MIT Press.

Stromer-Galley, J., & Foot, K. A. (2002). Citizen percep-tions of online interactivity and implications for politicalcampaign communication. Journal of Computer-Medi-ated Communication, 8(1), http://jcmc.indiana.edu/vol8/issuel/.

Sunstein, C. R. (2004). Democracy and filtering. Commu-nications of the ACM, 47(12), 57–59.

Sutter, D. (2001). Can the media be so liberal? The eco-nomics of media bias. Cato Journal, 20(3), 431–451.

Traugott, M. W., & Lavrakas, P. J. (2000). The voter’sguide to election polls. New York: Chatham.

Watts, M. D., Domke, D., Shah, D. V., & Fan, D. P.(1999). Elite cues and media bias in presidential cam-paigns: Explaining public perceptions of a liberalpress. Communication Research, 26(2), 144–175.

Wayne, S. J. (2001). The road to the White House 2000—the politics of presidential elections. New York:Palgrave.

Weisberg, H. F. (2004). The U.S. presidential and con-gressional elections. Electoral Studies, 24, 777–784.

Yates, F. (1934). Contingency table involving small num-bers and the χ2 test. Journal of the Royal StatisticalSociety, 1, 217–235.

Yi, J., Nasukawa, T., Bunescu, R. C., & Niblack, W.(2003). Sentiment analyzer: Extracting sentimentsabout a given topic using natural language processingtechniques. Paper presented at the 3rd IEEE Interna-tional Conference on Data Mining, Florida, USA.

witp workbench notes - weblyzard technologyfrom september to december 2004. the project drew upon...

Documents