When no news is bad news -- Detection of negative events from news media content
Kristoffer L. Nielbo, Frida Haestrup, Kenneth C. Enevoldsen, Peter B. Vahlstrup, Rebekah B. Baglini, Andreas Roepstorff
WWhen no news is bad news - Detection of negativeevents from news media content
Kristoffer L. Nielbo a,b , Frida Haestrup a , Kenneth C. Enevoldsen a , PeterB. Vahlstrup a , Rebekah B. Baglini b and Andreas Roepstorff b a Center for Humanities Computing Aarhus, Jens Chr. Skous Vej 4, Building 1483, 3rd floor, DK-8000 AarhusC, Denmark b Interacting Minds Centre, Jens Chr. Skous Vej 4, Building 1483, 3rd floor, DK-8000 Aarhus C, Denmark
Abstract
During the first wave of Covid-19 information decoupling could be observed in the flow of news mediacontent. The corollary of the content alignment within and between news sources experienced byreaders (i.e., all news transformed into Corona-news), was that the novelty of news content wentdown as media focused monotonically on the pandemic event. This all-important Covid-19 newstheme turned out to be quite persistent as the pandemic continued, resulting in the, from a newsmedia’s perspective, paradoxical situation where the same news was repeated over and over. Thisinformation phenomenon, where novelty decreases and persistence increases, has previously been usedto track change in news media, but in this study we specifically test the claim that new informationdecoupling behavior of media can be used to reliably detect change in news media content originatingin a negative event , using a Bayesian approach to change point detection.
Keywords
Newspapers, Pandemic Response, Bayesian Change Detection, Information Theory
Introduction
A peculiar behavior could be observed in news media when the first wave of Covid-19 virusspread across the world. In response to this pandemic event, the ordinary rate of change innews content was disrupted because every story became associated with Covid-19. On the onehand, content novelty went down, because every story became more similar to previous stories,but on the other hand, the Covid-19 association became more prevalent, resulting in, at leastinitially, an increase in content persistence. A recent study [1] argues that this behavior is anexample of the news information decoupling (NID) principle, according to which informationdynamics of news media are (initially) decoupled by temporally extended catastrophes suchthat the content novelty decreases as media focus monotonically on the catastrophic event, butthe resonant property of said content increases as its continued relevance propagate throughoutthe news information system. The authors further argued NID can be used to detect significantchange in news media that originate in catastrophic events.Previous studies have shown that variation in newspapers’ word usage is sensitive to the " [email protected] (K.L. Nielbo); [email protected] (F. Haestrup); [email protected] (K.C.Enevoldsen); [email protected] (P.B. Vahlstrup); [email protected] (R.B. Baglini) ~ https://knielbo.github.io/ (K.L. Nielbo) (cid:18) © a r X i v : . [ c s . C Y ] F e b ynamics of socio-cultural events [2, 3, 4], can detect event-driven shifts [5], and accuratelycan model effects of change in comprehensive collections of newspapers [6]. Furthermore, theassociative structure of newspapers has been shown to accurately capture thematic develop-ment [7], and, when modelled dynamically, is indicative of the evolution of cultural valuesand biases [3, 8]. Adaptive fractal analysis of word frequencies over time has been used todiscriminate between different classes of catastrophic events that display class-specific frac-tal signatures in, among other things, word usage in newspapers [9]. Several studies haveshown that information theoretical construct can be used to detect fundamental conceptualdifferences between distinct periods [2], concurrent normative and ideological movements [10],and even, development of ideational factors (e.g., creative expression) in temporally dependentwritings [11, 12, 13]. More specifically, a set of methodologically related studies studies haveapplied windowed relative entropy to thematic text representations to generate signals thatcapture information novelty as a reliable content difference from the past and resonance asthe degree to which future information conforms to said novelty [10, 11]. Three recent studieshave found that successful social media content show a strong association between novelty andresonance [14], that legacy news media under normal conditions display a remarkably similarmedium to strong association between novelty and resonance across the political spectrum [1],and, finally, that variation in the novelty-resonance association can predict significant changepoints in historical data [15].This study specifically tests the claim of [1] that NID-like behavior can provide input forchange point detection algorithms. Specifically, we propose to test the claim that two changepoints are observable in news media during the first phase of Covid-19, Lockdown and
Opening respectively, using a Bayesian approach to change point detection.
Results
Figure 1 displays a prototypical example of NID during the first phase of Covid-19 [1]. AlthoughCovid-19 news items date back to December 2019, ‘
W uhan ’, newspaper content is not impacteduntil the period after the first national outbreak (in this case in Denmark). ‘
V irus ‘. Fromthe phase 1 lockdown ‘
Lockdown ‘ to the opening, ‘
Opening ‘, the newspaper shows a valley innovelty and, initially, a peak in resonance until both processes approximately return to normalafter the opening.Source Class
NID
Start
NID
End
NID
Berlingske B .
07 [03 . , .
09] 04 .
28 [04 . , . T rue BT T .
10 [12 . , .
01] 07 .
25 [04 . , . F alse
Ekstrabladet T .
28 [01 . , .
17] 05 .
08 [01 . , . F alse
Jyllands-Posten B .
10 [03 . , .
14] 05 .
25 [05 . , . T rue
Kristligt Dagblad B .
07 [03 . , .
12] 04 .
15 [04 . , . T rue
Politiken B .
13 [03 . , .
13] 04 .
08 [04 . , . T rue
Table 1: Estimated temporal change points at 94% high density intervals for novelty. Columnone contains the name of the newspaper, columns two its class ( B roadsheet or T abloid), NIDStart and End is the beginning and end of the lockdown as represented in the newspaper, andthe final column indicated if the specific source supported the NID principle.To validate the observed behavior, we tested for two change points in novelty using a Bayesian igure 1: Novelty (upper panel) and resonance (lowerpanel) for the center-left newspaper
Politiken beforeand during Covid-19 phase 1. Trend lines in the upper and middle panel are estimated using a nonlinearadaptive filter suggested in [1]. model, see Appendix A for methods. The first change point, ‘
N ID
Start‘ should separate pre-lockdown from lockdown centered on week 11, and the second lockdown, ‘
N ID
End‘ from postopening (centered on week 16). Table 1 shows the estimated change points for six nationalnewspapers, two of which are T abloid newspapers (Class) and the remainder B roadsheet.From the model, it can be observed that all broadsheet newspapers seem to support the NIDprinciple in novelty. The first change point is placed in weeks 10-11, the second, however, ismore a matter of contention. The opening change point lies within April and displays a month’sdelayed response. Finally, it can be observed that tabloid press shows no indication of NIDbehavior. Table 2 and figure 2 show the posterior distributions, their means and highest densityintervals, for four broadsheet and one tabloid newspaper, clearly indicating that broadsheetnewspapers do conform to NID, while tabloids do not.Source N pre N NID N post Berlingske 0 .
36 [0 . , .
37] 0 .
29 [0 . , .
31] 0 .
34 [0 . , . .
29 [0 . , .
30] 0 .
23 [0 . , .
24] 0 .
27 [0 . , . .
27 [0 . , .
28] 0 .
19 [0 . , .
21] 0 .
26 [0 . , . .
27 [0 . , .
28] 0 .
15 [0 . , .
17] 0 .
26 [0 . , . igure 2: Posterior distributions of novelty at high density intervals for newspapers Berlingske, andPolitiken, see table 2. the medium to strong association between novelty and resonance is momentarily weakened.Following [1], we inspected the time-windowed linear fits of resonance on novelty, N × R , inorder to confirm this, see figure 3. All broadsheet newspapers display a slope decrease duringthe lockdown, thereby conforming to the NID principle 3. Tabloids on the other hand, followan inverse pattern, such that the N × R slope increases during the lockdown period.Source N × R pre N × R NID N × R post Berlingske 0 .
33 [0 . , .
51] 0 .
16 [ − . , .
38] 0 .
44 [0 . , . .
49 [0 . , .
66] 0 .
55 [0 . , .
83] 0 .
26 [0 . , . .
55 [0 . , .
72] 0 .
65 [0 . ,
1] 0 .
57 [42 , . .
42 [0 . , .
63] 0 .
31 [0 . , .
56] 0 .
39 [0 . , . .
57 [0 . , .
78] 0 .
43 [0 . , .
78] 0 .
76 [0 . , . .
39 [0 . , .
61] 0 .
16 [ − . , .
37] 0 .
43 [0 . , . Concluding Remarks
In conclusion, this study sought to validate the news information decoupling (NID) principleon a sample of six national Danish newspapers during the first phase of Covid-19. Using aBayesian approach to change point detection, we found that content novelty in broadsheetnewspapers does indeed display statistically reliable points of change during the Covid-19lockdown. NID was further corroborated by the N × R pre slopes that indicated a decouplingof resonance from novelty during the lockdown. Several observations can be made from the igure 3: N × R slopes before during and after the lockdown for Berlingske (upper row), Ekstrabladet(middle row), and Politiken (lower row) during Covid-19 phase 1. findings. First, the estimated change points for the ‘Pre-lockdown → Lockdown’ are spreadover a two week interval, which reflects that a lockdown could be reasonably predicted alreadyfrom the first Covid-19 incident in Denmark. Second, in a similar vein the ‘Lockdown → Opening‘ change points are spread over an entire month from April 8 to May 8. The Danishgovernment during the period was center-left and model’s uncertainty in determining the open-ing may reflect political observations [1], where center-right newspapers (e.g., Berlingske andJyllands-Postern) were more sceptical towards the government’s implementation of an openingthan the center-left (e.g., Politiken). In other words, the center-right might have been morereluctant to acknowledge the opening as a return to normal. Third, tabloid newspapers do notshow any indication of a news decoupling, on the contrary, their N × R pre slopes momentarilyincreases during the lockdown . This increase in slopes does, however, not provide much usefulnformation, because, as shown by the change point detection model, the periodization is notmeaningful to the two tabloid newspapers.Validation of the NID principle is still needed for multilingual data and its value for crisismanagement should be further tested. For change detection, the scope of the principle needsadditional testing; does NID generalize beyond a small set of negative events to, for instance,temporally extended significant events (e.g., moon landing, fall of the Berlin Wall). Finally,several comparisons already hinted that left vs. right-wing newspapers, tabloid vs. broad-sheet newspapers, silly season and other seasonal effects, are interesting venues for media andjournalism researchers. AppendixA. Methods
Data and Normalization
The data set consists of all linguistic content (title and body text) from front pages of sixDanish national newspapers Berlingske, BT, Ekstrabladet, Jyllands-Posten, Kristligt Dagblad,and Politiken. The newspapers were sampled during December 1, 2019 to July 1 2020. Contentnot produced by the newspaper, e.g., advertisements, was excluded from the sample. In orderto normalize linguistic content, numerals and highly frequent function words were removed, andthe remaining data were lemmatized and casefolded. Subsequently, the data were representedas a bag-of-words (BoW) model using latent Dirichlet allocation in order to generate a denselow-rank representation of each article. Note that with a few modifications to equations (4) and(5), the approach works for any probabilistic or geometric vector-representation of documents.Novelty and resonance were estimated for in windows of one week ( w = 7). Novelty and Resonance
Two related information signals were extracted from the temporally sorted BoW model:
Nov-elty as an article s ( j ) ’s reliable difference from past articles s ( j − , s ( j − , . . . , s ( j − w ) in window w : N w ( j ) = 1 w w ∑︂ d =1 J SD ( s ( j ) | s ( j − d ) ) (1)and resonance as the degree to which future articles s ( j +1) , s ( j +2) , . . . , s ( j + w ) conforms toarticle s ( j ) ’s novelty: R w ( j ) = N w ( j ) − T w ( j ) (2)where T is the transience of s ( j ) : T w ( j ) = 1 w w ∑︂ d =1 J SD ( s ( j ) | s ( j + d ) ) (3)The novelty-resonance model was originally proposed in [10], but here we propose a sym-metrized and smooth version by using the Jensen–Shannon divergence ( J SD ): SD ( s ( j ) | s ( k ) ) = 12 D ( s ( j ) | M ) + 12 D ( s ( k ) | M ) (4)with M = ( s ( j ) + s ( k ) ) and D is the Kullback-Leibler divergence: D ( s ( j ) | s ( k ) ) = K ∑︂ i =1 s ( j ) i × log s ( j ) i s ( k ) i (5)Finally, in order to describe the information states before and after an events (e.g., Lockdown,Opening), we fit resonance on novelty to estimate the N × R slope β in the specific timewindows: R i = β + β N i + ϵ i , i = 1 , . . . , n. (6) Bayesian Change Point Detection
For the estimation of change points, a Bayesian approach was used. Following previous con-siderations, we assume that the time series contains two change points, τ and τ . Aside fromchange points, the series is assumed to be stable and follow a normal distribution with var-ied mean, µ i , and singular variance, σ . This gives us the following model given the observedNovelty, N i : N t = ⎧⎪⎨⎪⎩ N ( µ , σ ) for t < τ N ( µ , σ ) for τ ≤ t < τ N ( µ , σ ) for t ≥ τ (7)for which we wish to estimate the location of the change points τ i , means µ i and variance σ , i.e. the following posterior: P ( µ i , σ, τ i | N t ) = P ( µ , µ , µ , σ, τ , τ | N t )For estimation of the posterior, we have used NUTS sampling as implemented in pyMC3 [16]using 4000 samples. The estimation was done using using naive to slightly conservative priorsassuming that the change points, τ i , can be anywhere in the sequence (with τ > τ ) and thatthe variance, σ , is stable across change points. Note that the half Cauchy prior distributionhas series of beneficial properties [17, 18] including its fat tail which allows for extreme values.These assumptions were modelled using the following priors: µ i ∼ N (0 , . σ ∼ Half Cauchy(0 . τ ∼ Uniform(0 , max( N t )) τ ∼ Uniform( τ , max( N t )) B. Online Resources
All data are proprietary and have been collected through Infomedia’s API: https://infomedia.dk/. For inquiries regarding models and derived data, please contact [email protected]. Theource code for methods is available on Github: https://bit.ly/3beahFd. More details on NIDdetection can be found at NeiC’s NDHL website: https://bit.ly/3bfeW9C.
Acknowledgments
This research was supported the ”HOPE - How Democracies Cope with COVID-19”-projectfunded by The Carlsberg Foundation with grant CF20-0044, NeiC’s Nordic Digital HumanitiesLaboratory project, and DeiC Type-1 HPC with project DeiC-AU1-L-000001. The authorswould like to thank Berlingske Media, JP/Politkens Hus, and Kristeligt Dagblad for providingaccess to proprietary data.
References [1] K. L. Nielbo, R. B. Baglini, P. B. Vahlstrup, K. C. Enevoldsen, A. Bechmann, A. Roep-storff, News Information Decoupling: An Information Signature of Catastrophes in LegacyNews Media, arXiv:2101.02956 [cs] (2021). URL: http://arxiv.org/abs/2101.02956, arXiv:2101.02956.[2] J. Guldi, The Measures of Modernity: The New Quantitative Metrics of Historical ChangeOver Time and Their Critical Interpretation, International Journal for History, Cul-ture and Modernity 7 (2019) 899–939. URL: https://brill.com/view/journals/hcm/7/1/article-p899 42.xml. doi: .[3] J. van Eijnatten, R. Ros, The Eurocentric Fallacy. A Digital-Historical Approach to theConcepts of ‘Modernity’, ‘Civilization’ and ‘Europe’ (1840–1990), International Jour-nal for History, Culture and Modernity 7 (2019) 686–736. URL: https://brill.com/view/journals/hcm/7/1/article-p686 33.xml. doi: .[4] J. Daems, T. D’haeninck, S. Hengchen, T. Zere, C. Verbruggen, ‘Workers of the World’ ?A Digital Approach to Classify the International Scope of Belgian Socialist Newspapers,1885–1940, Journal of European Periodical Studies 4 (2019) 99–114. URL: https://ojs.ugent.be/jeps/article/view/10187. doi: .[5] M. Kestemont, F. Karsdorp, M. D¨uring, Mining the Twentieth Century’s Historyfrom the Time Magazine Corpus, in: Proceedings of the 8th Workshop on LanguageTechnology for Cultural Heritage, Social Sciences, and Humanities (LaTeCH), Asso-ciation for Computational Linguistics, Gothenburg, Sweden, 2014, pp. 62–70. URL:http://aclweb.org/anthology/W14-0609. doi: .[6] P. Bos, H. Wijfjes, M. Piscaer, G. Voerman, Quantifying “Pillarization”: ExtractingPolitical History from Large Databases of Digitized Media Collections, Proceedings of the3rd HistoInformatics Workshop (2016) 10.[7] D. J. Newman, S. Block, Probabilistic topic decomposition of an eighteenth-centuryAmerican newspaper, Journal of the American Society for Information Science and Tech-nology 57 (2006) 753–767. URL: http://doi.wiley.com/10.1002/asi.20342. doi: .[11] J. Murdock, C. Allen, S. DeDeo, Exploration and Exploitation of Victorian Sciencein Darwin’s Reading Notebooks, arXiv preprint arXiv:1509.07175 (2015). URL: http://arxiv.org/abs/1509.07175.[12] K. L. Nielbo, M. L. Perner, C. P. Larsen, J. Nielsen, D. Laursen, Automated Compo-sitional Change Detection in Saxo Grammaticus’ Gesta Danorum, in: DHN, 2019, pp.320–332.[13] K. L. Nielbo, K. F. Baunvig, B. Liu, J. Gao, A curious case of entropic decay: Persistentcomplexity in textual cultural heritage, Digital Scholarship in the Humanities 34 (2019).URL: https://doi.org/10.1093/llc/fqy054. doi: .[14] K. Nielbo, P. Vahlstrup, A. Bechmann, Trend Reservoir Detection: Minimal Persistenceand Resonant Behavior of Trends in Social Media, Proceedings of Computational Hu-manities Research 1 (2021).[15] E. Vrangbæk, K. Nielbo, Composition and Change in De Civitate Dei: A Case Study ofComputationally Assisted Methods, Studia Patristica (2021).[16] J. Salvatier, T. V. Wiecki, C. Fonnesbeck, Probabilistic programming in python usingpymc3, PeerJ Computer Science 2 (2016) e55.[17] A. Gelman, Prior distributions for variance parameters in hierarchical models, BayesianAnalysis 1 (2006) 515–533. doi: .[18] N. G. Polson, J. G. Scott, On the Half-Cauchy Prior for a Global Scale Param-eter, Bayesian Analysis 7 (2012) 887–902. URL: https://projecteuclid.org/euclid.ba/1354024466. doi:10.1214/12-BA730