Upgrade to Pro — share decks privately, control downloads, hide ads and more …

The future of news: filter bubbles vs. diversity

Avatar for Judithmoeller Judithmoeller
November 14, 2019

The future of news: filter bubbles vs. diversity

Presentation by Sanne Vrijenhoek at the ECP jaarcongres 2019

Avatar for Judithmoeller

Judithmoeller

November 14, 2019

Other Decks in Science

Transcript

  1. De toekomstvan nieuwsvoorziening–filter bubbles versus diversiteit Een project van de

    Universiteit van Amsterdam, RTL Nieuws en het SIDN Fonds
  2. 2

  3. 3

  4. Filter bubbles 4 Eenrisico? “Increasing filtering mechanisms make it more

    likely for people to only get news on subjects they are interested in, and with the perspective they identify with. ... It will also tend to create more insulated communities as isolated subsets within the overall public sphere. … Such developments undoubtedly have a potentially negative impact on democracy.”
  5. Filter bubbles 5 Of eenkans? Audience tracking will “turn[…] journalism

    from elitism of writing for itself and back to writing what people are actually looking for.” (Usher, 2010) “Instead of discussing ways journalism can improve, the solution appears to be to burn down the gates and grant agency to the audience.” (Tandoc &Thomas, 2015)
  6. Filter bubbles 6 Of eenkans? “Als we zien dat je

    al drie uur lang televisie hebt gekeken, moeten we dan stoppen met het doen van nieuwe suggesties, zodat mensen de tv uitzetten en naar buiten gaan? Is dat een van onze verantwoordelijkheden? Of moeten we er alleen voor zorgen dat je aan het beeld gekluisterd blijft? Is dat waar we voor optimaliseren?” - Data scientist for a large media company)
  7. 7 Think of privacy, filterbubbles, democracy, your mission! Go ahead,

    use that data! Beat FB! Clicks sell! Eenvraagstuk voorde media
  8. 8 Project: het ontwikkelen van een tool die mediabedrijven kunnen

    gebruiken voor het meten van diversiteit in hun nieuwsaanbevelingen
  9. Project setup • Scrape nieuwsartikelen van het internet • Alle

    artikelen uit 2017 • nu.nl • Volkskrant • GeenStijl • Simuleer een arbitrair aantal gebruikers en hun recommendations volgens vier verschillende ‘baseline’ principes • Random • Most popular • ‘More like this’ • Politiek • Bereken ‘diversiteit’ en visualiseer 9 Project setup
  10. Expectation management 10 • Nederlands • Heel verschillende schrijfstijlen tussen

    bronnen • Geen geannoteerde data • Titel • Tekst • URL • Bron • Datum van publicatie • Auteur (soms) Expectation management
  11. Project setup • Scrape nieuwsartikelen van het internet • Alle

    artikelen uit 2017 • nu.nl • Volkskrant • GeenStijl • Simuleer een arbitrair aantal gebruikers en hun recommendations volgens vier verschillende ‘baseline’ principes • Random • Most popular • ‘More like this’ • Politiek • Bereken ‘diversiteit’ en visualiseer 11 Project setup
  12. Project setup • Scrape nieuwsartikelen van het internet • Alle

    artikelen uit 2017 • nu.nl • Volkskrant • GeenStijl • Simuleer een arbitrair aantal gebruikers en hun recommendations volgens vier verschillende ‘baseline’ principes • Random • Most popular • ‘More like this’ • Politiek • Bereken ‘diversiteit’ en visualiseer 12 Project setup
  13. “Of” en “hoe” algoritmische aanbevelingen leiden tot meer of minder

    diversiteit is grotendeels afhankelijk van het perspectief op democratie- theorie dat men heeft Afhankelijk van dit perspectief kan diversiteit meerdere doelen en waarden dienen, waar sommigen elkaar ook tegen kunnen spreken. 14 Diversiteiten“democratic theory”
  14. Frameworks of Democracy 15 • Deliberative • Similar to the

    participatory model, but with a focus on ‘finally agreeing on different values and issues’ • Content should challenge the user in particular • Fostering tolerance and open-mindedness • Larger proportion of content featuring different cultures, ethnicities and linguist groups • Impartial, non-provocative style of communication • Critical • Focuses on the importance of conflict and disagreement • Emotional and provocative content • Confront the user with views they did not know existed, or allow minorities to find like-minded individuals • Liberal • Individual freedom, expression, personal development • Recommendations are based on what the user is interested in • Dependence on popularity • Anything goes… • But no anti-democratic hate speech, exclusion of ideas or dissuading people from voting • Participatory • “Shared civic culture and commitment to citizenship” • Common good trumps self-interest • All ideas in society should be represented • Focus on supplying information that is understandable and still complete • Longer articles • Vocabulary matched to user’s level Normatiefkader
  15. Frameworks of Democracy: quantified 16 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  16. Frameworks of Democracy: quantified 17 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  17. Frameworks of Democracy: quantified 18 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  18. 19 “Zahradil zegt tegen NU.nl dat hij denkt dat nieuwkomer

    Forum voor Democratie "een aantal parlementariërs" zal leveren in de Europese verkiezingen. Een probleem daarbij is wel dat ChristenUnie duidelijk heeft laten weten dat het niet heel erg blij is dat het straks met Forum voor Democratie in één groep zit.” https://www.nu.nl/europese-verkiezingen-2019/5867613/frans-timmermans-ik-word-voorzitter-van-europese-commissie.html Stap 1: Named Entity Recognition → Herkent personen, organisaties en locaties in mensgeschreven tekst Data annotatie
  19. 20 “Zahradil zegt tegen NU.nl dat hij denkt dat nieuwkomer

    Forum voor Democratie "een aantal parlementariërs" zal leveren in de Europese verkiezingen. Een probleem daarbij is wel dat ChristenUnie duidelijk heeft laten weten dat het niet heel erg blij is dat het straks met Forum voor Democratie in één groep zit.” https://www.nu.nl/europese-verkiezingen-2019/5867613/frans-timmermans-ik-word-voorzitter-van-europese-commissie.html Stap 1: Named Entity Recognition → Herkent personen, organisaties en locaties in mensgeschreven tekst Stap 2: Wikipedia lookup → Haalt extra gestructureerde informatie over deze entiteiten op van Wikipedia Data annotatie
  20. 21 “Zahradil zegt tegen NU.nl dat hij denkt dat nieuwkomer

    Forum voor Democratie "een aantal parlementariërs" zal leveren in de Europese verkiezingen. Een probleem daarbij is wel dat ChristenUnie duidelijk heeft laten weten dat het niet heel erg blij is dat het straks met Forum voor Democratie in één groep zit.” https://www.nu.nl/europese-verkiezingen-2019/5867613/frans-timmermans-ik-word-voorzitter-van-europese-commissie.html Stap 1: Named Entity Recognition → Herkent personen, organisaties en locaties in mensgeschreven tekst Stap 2: Wikipedia lookup → Haalt extra gestructureerde informatie over deze entiteiten op van Wikipedia Data annotatie
  21. Frameworks of Democracy: quantified 24 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  22. Frameworks of Democracy: quantified 25 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  23. Frameworks of Democracy: quantified 27 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  24. Frameworks of Democracy: quantified 28 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Normatiefkader Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  25. Frameworks of Democracy: quantified 32 • Deliberative model • Focus

    on creating a ‘public sphere’ that stimulates discourse • Larger focus on political articles • Personalization on a user level in order to challenge the user in particular • Impartial, non-emotional, non-provocative communication style • Larger representation of minorities • Critical model • Emotional and provocative content • Confront the user with views they did not know existed • Or • Allow minorities to find like-minded individuals • Liberal model • Distinct personalization on a user level • Popular articles have larger representation • Matches the user’s preferences • Around election time, a larger share of political articles • Participatory model • Stimulates the ‘good citizen’ – less personalization • Larger share of political articles • Longer articles • Matched vocabulary • ‘All ideas should be represented’ Volgendestappen Politicalness Attention spread Concordance to personal preference Complexity Defragmentation Coverage
  26. 33 Emotioneel, neutraal en activerend taalgebruik • Relevant voor deliberative

    en critical model • “Emotive speech contains comparatively less articles and prepositions, and more • pronouns” • Activerend taalgebruik → Gebiedende wijs Nieuwemetrics
  27. 34 Opinion detection • Leer zgn. “taalpatronen” uit geannoteerde data

    en gebruik deze patronen om meningen en standpunten te detecteren in nieuwe teksten • Relevant voor participatory, deliberative en critical model Elon Musk says that Tesla will beat Porsche. Tesla will beat Porsche, according to Elon Musk Nieuwemetrics
  28. 35 Opinion detection • Leer zgn. “taalpatronen” uit geannoteerde data

    en gebruik deze patronen om meningen en standpunten te detecteren in nieuwe teksten • Relevant voor participatory, deliberative en critical model <SOURCE> says that <CONTENT>. <CONTENT>, according to <SOURCE>. Elon Musk says that Tesla will beat Porsche. Tesla will beat Porsche, according to Elon Musk Blue: Source (who holds the opinion) Red: Content (what is the opinion) Black: Cue (pattern indicating an opinion) Nieuwemetrics
  29. 36 Opinion detection • Leer zgn. “taalpatronen” uit geannoteerde data

    en gebruik deze patronen om meningen en standpunten te detecteren in nieuwe teksten • Relevant voor participatory, deliberative en critical model <SOURCE> says that <CONTENT>. <CONTENT>, according to <SOURCE>. Boris Johnson says that the new EU treaty will get Brexit done These proposals are not beneficial to the economic well-being of Northern Ireland, according to DUP. Nieuwemetrics
  30. 37 Opinion detection • Leer zgn. “taalpatronen” uit geannoteerde data

    en gebruik deze patronen om meningen en standpunten te detecteren in nieuwe teksten • Relevant voor participatory, deliberative en critical model Source Content Boris Johnson the new EU treaty will get Brexit done DUP These proposals are not beneficial to the economic well-being of Northern Ireland … … Nieuwemetrics
  31. 38 Identificeer ‘minorities’ • Als concept even abstract als ‘diversiteit’

    zelf • Relevant voor de deliberative en critical modellen • “a non-dominant group of individuals who share certain national, ethnic, religious or linguistic characteristics which are different from those of the majority population. In addition, it has been argued that the use of self-definition which has been identified as ‘a will on the part of the members of the groups in question to preserve their own characteristics’ and to be accepted as part of that group by the other members” • Capotorti, 1991 Nieuwemetrics
  32. 39 Identificeer ‘minorities’ • Als concept even abstract als ‘diversiteit’

    zelf • Relevant voor de deliberative en critical modellen • “a non-dominant group of individuals who share certain national, ethnic, religious or linguistic characteristics which are different from those of the majority population. In addition, it has been argued that the use of self-definition which has been identified as ‘a will on the part of the members of the groups in question to preserve their own characteristics’ and to be accepted as part of that group by the other members” • Capotorti, 1991 Nieuwemetrics
  33. 40 Identificeer ‘minorities’ • Als concept even abstract als ‘diversiteit’

    zelf • Relevant voor de deliberative en critical modellen • “a non-dominant group of individuals who share certain national, ethnic, religious or linguistic characteristics which are different from those of the majority population. In addition, it has been argued that the use of self-definition which has been identified as ‘a will on the part of the members of the groups in question to preserve their own characteristics’ and to be accepted as part of that group by the other members” • Capotorti, 1991 Nieuwemetrics
  34. Follow Up • Werk samen met verschillende mediabedrijven aan de

    identificatie en bruikbaarheid van de tool • Start en faciliteer discussie rondom diversiteit • Geïnteresseerde partners: • RTL (Nederland) • VRT (Vlaanderen) • ZDF (Duitsland) • NPO (Nederland) • BBC (Verenigd Koninkrijk) 41 Volgendestappen
  35. Conclusie 42 Conclusie • Technologische vooruitgang in recommender systems zijn

    zowel een risico als een kans voor democratische warden • Hoewel er nog heel veel ruimte is voor verbetering zijn de eerste resultaten veelbelovend • In de toekomst willen we werken aan de uitwerkingen van drie nieuwe ‘metrics’, en deze in de praktijk testen met partners uit de industrie