• No results found

Tiago Devezas

INESC TEC [email protected]

Sérgio Nunes

INESC TEC/Faculdade de Engenharia/ Universidade do Porto

[email protected]

Resumo:

A filtragem algorítmica de notícias pode ajudar a combater a sobrecarga de informação, mas também encerrar efeitos potencialmente nefastos. Estes incluem a disseminação de notícias falsas e as chamadas bolhas de filtro, que isolam os consumidores de diversas perspetivas e podem reforçar pre- conceitos. Em aplicações deste tipo, as consequências podem ser minoradas operando de forma transparente e cedendo controlo aos utilizadores. De forma a informar o desenvol- vimento de uma aplicação algorítmica de notícias focada na transparência e no controlo dos utilizadores, desenvolve- mos um questionário para examinar a importância relativa de vários critérios que podem ser usados para controlar e personalizar os resultados. Com base nas respostas de 432 utilizadores, um subconjunto dos critérios mais valorizados foi implementado num protótipo funcional de um agregador algorítmico de notícias. A aplicação usa algoritmos de agru- pamento (clustering) para apresentar grupos de notícias relacionadas por tópico. São utilizadas estratégias para per- mitir o acesso em tempo útil a notícias produzidas por profis- sionais, maximizar a exposição a diferentes perspetivas, for- necer controlo aos utilizadores sobre os parâmetros e fontes

de notícias usadas para gerar a lista, e exibir informação de transparência sobre a sua operação. Ao apresentar um pro- tótipo de uma aplicação de notícias que incorpora uma série de abordagens baseadas nas respostas de utilizadores para aumentar o controlo e a transparência, pretendemos forne- cer uma referência potencialmente útil para investigadores e implementadores de sistemas com objetivos similares. Palavras-chave:

Personalização de notícias; transparência algorítmica; con- trolo dos utilizadores

Abstract:

Algorithmic filtering of news content can help combat infor- mation overload but also entail potentially nefarious effects. These include the dissemination of fake news and the so- called filter bubbles, which isolate consumers from diverse perspectives and possibly reinforce their biases. In such ap- plications, these consequences might be lessened by operat- ing transparently and yielding control to the users. In order to inform the development of an algorithmic news applica- tion focused on transparency and user control, we developed a questionnaire to examine the relative importance of mul- tiple criteria which could be used to control and personalize its output. Based on the answers from 432 users, a subset of the most highly rated criteria was implemented in a function- al prototype of an algorithmic news aggregator. The applica- tion employs clustering algorithms to present groups of top- ic-related news articles. It employs strategies to allow timely access to professionally produced news articles, maximize exposure to different perspectives, provide user control over the parameters and news sources used to generate the feed, and present transparency information about its operation. By presenting a prototype of a news application that embod- ies a set of approaches based on user feedback to increase

Tiago Devezas | Sérgio Nunes

user control and transparency, we aim to provide a poten- tially useful reference for researchers and implementers of systems with similar goals.

Keywords:

News personalization; algorithmic transparency; user control User control and transparency in an algorithmic news application

386 387

Introduction

Algorithmic content personalization and curation can help re- duce information overload by tailoring the information displayed based on the system’s assertion of what the user wants and in- crease the user experience (Rader, 2017). However, this approach can introduce new problems. In addition to potential biases in the code, news personalization algorithms can limit users’ expo- sure to different perspectives and therefore amplify biases and isolate them in their own ideological and cultural bubbles.

There’s also the issue of veracity, particularly in systems with- out any human intervention, such as Facebook’s News Feed. The widespread dissemination of fake, but highly shared, news on Facebook related with the 2016 USA presidential election (Sil- verman, 2016) led to discussions about the potential impact of false news on the electoral outcome. It has been indicated that most of the fake news stories disseminated on social media fa- voured the winning candidate, Donald Trump, over the run- ner-up, Hillary Clinton (Allcott & Gentzkow, 2017). Some com- mentators have expressed their belief that fake news shared on social media helped elect Donald Trump (Parkinson, 2016) but recent research indicates that the role of fake news on the elec- tion’s outcome is unclear (Allcott & Gentzkow, 2017).

Even if the 2016 election’s outcome wasn’t determined by the spread of misinformation in social media, it has been shown that algorithmic manipulation can indeed shift voting preferences. Epstein and Robertson (2015) conducted a large-scale study which estimated that the manipulation of a search engine’s rank- ing algorithm could change the outcome of more than 25% elec- tions worldwide.

These potentially far-reaching effects have led to a growing discussion about how to achieve algorithmic transparency and accountability in several fields, including news media. However, this is still an ongoing debate. Normative approaches are lack- ing, and concerns about possible negative impacts on the users’ experience have been mentioned. Some potential strategies have

Tiago Devezas | Sérgio Nunes

been discussed in the literature. These include disclosing the al- gorithms’ existence and influence, providing transparent infor- mation about how they operate, and let users manipulate the system, not the other way around (Diakopoulos & Koliska, 2016).

In the context of this work, which is based on the first author’s Master’s dissertation supervised by the second author (Devezas, 2017), a questionnaire was developed to measure the importance of several criteria that could be used to control an algorithmic news application. They were divided into five categories: news content, diversity of perspectives, sources’ attributes, algorith- mic information, and personalization type. Participants were asked to rate each criterion on a 5-level scale with the following labels: ‘Not important’, ‘Little important’, ‘Moderately import- ant’, `Important’, and ‘Very important’.

Based on a quantitative analysis of the answers provided by 432 participants from the University of Porto (U.Porto) commu- nity, we developed a functional prototype of an algorithmic news aggregator. It aims to combat information overload and provide user control and transparency information, while providing a positive user experience based on the requirements expressed by the participants.

This paper discusses the implementation of several features intended to address the aforementioned issues. We start by pre- senting the news collection platform in which the prototype was integrated. We then describe the study’s methodology, followed by a discussion how the findings from both the literature and the analysis of the participants’ responses were integrated in the de- veloped front-end application in order to achieve the proposed goals of alleviating information overload, providing user control, and present algorithmic transparency information.