Europe in the Headlines: What Two Decades of French News Reveal about EU Sentiment

Abstract
This policy brief documents the treatment of European affairs by French newspapers over the last two decades. Based on a new dataset of 400,000 articles and natural language processing methods, we demonstrate that the interest in European issues has remained stable since 2005 and is primarily driven by the European Parliament elections every 5 years. Our analysis also reveals that French media sentiment towards the EU deteriorated significantly following the financial and sovereign debt crises, mirroring the trend observed in opinion on EU sentiment derived from the Eurobarometer surveys. However, from 2013 onward, a divergence emerged since sentiment in the press gradually returns to pre-crisis levels, while public image of the European Union in the opinion remains below these levels.

Introduction

Defining the “European public opinion” remains a challenge. The 2005 referenda on European treaties in France and the Netherlands bear the memory of a disjunction between supposed public opinion on European affairs and results in votes, raising the complexity of establishing a common understanding of what EU citizens think. Nonetheless, the inclusion of citizens in European Union (EU) policy shaping is nowadays considered as an essential element of transparency, accountability and finally acceptance of EU policies. The understanding of citizens’ sentiment towards EU affairs is therefore of utmost importance.

The European Commission set up in the 1970s a unique collection of opinion surveys, the Eurobarometer, to help assessing the opinion of European citizens. This source, based on interviews conducted with a representative sample of citizens in each European country, has remained largely unchallenged and is therefore the main dataset used in most of the literature on European sentiment (see for instance Ingelhart and Reif, 1991, or Teperoglou and Belchior, 2024). In a way, it almost embodies the “European public opinion” itself (Signorelli, 2012). Nonetheless, the Eurobarometer is not without political and methodological limitations, notably concerning sampling issues in certain countries or challenges linked to the cross-national intelligibility of questions across the 27 Member States (Aldrin and de Lassale, 2011), but also because it might potentially also be instrumentalized (Nissen, 2014).

Building on recent advances in natural language processing (NLP), this article precisely aims at providing another source of understanding of the sentiment toward the European affairs in France by drawing on information from the French press. Newspaper readers, who are often older and more educated than the average, are not representative of the French population, and therefore the press cannot claim to be fully representative of citizens’ opinions, being also challenged by social media. However, despite the decline of its readers, written press remains an important factor of public opinion building, due to its specific position as actor of the public sphere where opinions are shaped. It can therefore provide a valuable source of information for understanding public sentiment, as a useful complement to opinion derived from surveys (Vliegenthart and Walgrave, 2008, Mourlon-Druol et al., 2022, among others).

Using a large language model (LLM), we built a new dataset covering around 400,000 articles related to the European Union published in France between 2005 and 2023 from more than 100 local and national newspapers as well as magazines. In particular, we have full coverage of the six national daily newspapers throughout the entire period. For each article, we determine both the topic addressed and the tone towards the EU (positive, neutral or negative).

Compared to previous studies, we enhance the methodology by going beyond traditional keywords- and dictionary-based approaches (see for instance Bortoli et al. 2018) and leveraging large language models (LLMs) to improve classification performance. These new techniques enable a more direct processing of the original text as input, reducing dependence on external datasets typically used in dictionary-based approaches. The accuracy of our methods in both identifying EU-related articles and assessing their topic and tone is improved. In addition, our dataset offers a substantially broader coverage of newspapers than most of the existing literature, which usually focuses on only one or two newspapers (see, for example, Papadia et al., 2019 focusing only on Le Monde newspaper for France).

EU salience in French newspapers

The salience of EU affairs remained relatively stable over the period, around 0.75% of published articles (Figure 1). The main identified peaks can be directly linked to the 2005 European Constitution referendum and to the different European Parliament elections, which have an amplifying effect on news coverage of the European Union. This greater media coverage echoes what is observed by Broc and Verdier (2019) on the treatment of European elections in French TV news. However, such interest remains quite short-lived: there is no persistent effect in news coverage after the elections. Other major European events (such as the Brexit referendum, the 2015 refugee crisis or the 2020 NGEU programme) failed to trigger any major interest.

Figure 1. Monthly share of EU-related articles in the French press

Source: Jehle and Le Gallo (2025), based on Factiva data
Note: The share of EU-related articles is obtained by the number of articles related to the European Union (obtained via LLM computations) by the total number of articles published. Vertical bars signal European elections and the 2005 European Constitution referendum.

What are we talking about? EU covered topics

Using an LLM, we define the 10 most frequent topics in EU-related articles. They account for half of the articles of the database (Figure 2).  The above-mentioned strong interest in European elections logically leads to the most frequent topic being related to politics. In second place is the topic of school and cultural exchanges for young people, notably through references to the famous Erasmus programme. The third most frequent topic concerns agricultural affairs, highlighting notably the significance of the Common Agricultural Policy in French news. We then find several topics linked to the crises recently experienced by the European Union: the migrant crisis, the Euro area crisis, Brexit or the invasion of Ukraine war.

Figure 2. Top 10 topics mentioned in EU articles

Source: Jehle and Le Gallo (2025), based on Factiva data
Note: for each topic, the four most representative words are retrieved. UMP stands for Union pour un Mouvement Populaire, name of the French right-wing party between 2002 and 2015. CAC refers to the benchmark French stock market index CAC40.

Thanks to our dataset covering both national and local newspapers, we can show that the local newspapers devote more articles to everyday and concrete subjects. The topic on cultural exchanges is for instance first and foremost a local one, with some 70% of articles on this topic coming from local newspapers (although they only account for 24.5% of overall articles). This is also the case for the third topic, dealing with agricultural affairs, which is also mainly a local one (35% of the articles published in the local press).These findings echo Mendez et al. (2020), which found, based on a sample of British and Spanish newspapers, that local newspapers concentrate on EU practical policies aligned with local concerns. Our analysis reveals then the multilayered and multicentred coverage of EU affairs when enlarging the scope to all written news media.

An EU sentiment index

On the overall dataset, a relative majority of articles are labelled as negative (36.4%, against 30.0% as positive and 33.6% as neutral). It is in line with the findings of Alarcón (2010) who demonstrates, on a much smaller dataset, that newspapers publish more negative than neutral and positive news in their coverage of the European Union. We then compute a sentiment index of French newspaper towards the EU as the difference between the number of positive and negative articles over the total number of articles (Figure 3).

Figure 3. Monthly French newspapers’ EU sentiment index

Source: Jehle and Le Gallo (2025), based on Factiva data
Note: The blue curve represents French newspapers’ sentiment as a ratio between the net balance (number of positive – negative articles) and the total number of articles. The orange curve shows the annual moving average of the blue line.

The index peaked in May 2017 corresponding to the period following the 2017 French presidential election. The victory of Emmanuel Macron as President of the French Republic, with very affirmed pro-European positions, are reflected in the many positive articles.

The index exhibits distinct phases: stable from 2005–2009, sharply declining until 2012, unevenly recovering through 2017, declining again in 2018–2019, and stabilizing at unprecedented levels since 2021. The sharp drop in sentiment between 2009 and 2012 can be explained by the various crises that the European Union experienced during this period (financial crisis, sovereign debt crisis…). As a result, we can see that topics related to the Euro area and public deficits, associated at that time with rather negative tone, take on a more prominent role during this period. These latter topics contribute largely to the degradation of the balance of our sentiment index.

If the evolution of the sentiment index between the different categories of newspapers is similar, local newspapers publish constantly more positive articles about the European Union. This pattern can be accounted for by two factors. First, local newspapers tend to focus on concrete EU achievements (school cooperation, etc), notably local ones, which are generally covered in a more favourable tone. Second, the difference between national and local newspapers can also be explained by the latter’s more favourable treatment of each topic. For 90% of the topics, the average sentiment index of local press articles is higher than that of articles from national newspapers.

Comparison with the Eurobarometer

Finally, we aim to assess whether the sentiment indicator, constructed from the press using LLM techniques, sheds light on French citizens’ opinion towards the European Union as set out by surveys. To do this, we compare our sentiment index with the results obtained from the French part of the Eurobarometer Standard survey. We more specifically focus on the following question: In general, does the EU conjure up for you a very positive, fairly positive, neutral, fairly negative or very negative image?

By aggregating fairly positive/very positive and fairly negative/very negative responses, we can then compute a sentiment index of French citizens towards the European Union and compare it with our newspaper sentiment index (Figure 4).

Figure 4. Public attitudes toward the European Union and newspaper sentiment index

Source: Jehle and Le Gallo (2025), based on Factiva data.
Note: The blue curve (lhs) represents French citizens’ perception of the EU as measured by the Eurobarometer (net balance of positive and negative opinions), while the red curve (right-hand side) shows the annual moving average of our sentiment index computed from the French written press.

We can broadly distinguish two periods. First an initial period from 2005 to 2015, where the two sentiment indexes broadly follow the same trends (stability between 2005 and 2009, followed by a sharp decline during the financial and sovereign debt crises). Second, from 2015 onwards, a form of disconnection is observed: while the index obtained from the press returns to – and then exceeds – the levels reached previously, the French citizens’ image of the European Union no longer changes significantly and remains globally stable, below pre-crisis levels.

One assumption on this disconnection is a form of public distrust towards the opinions expressed by the press regarding the European Union. Such a possibility could be reflected in the decline in the level of trust that French citizens have in newspapers. Indeed, this trust in the written press (as measured by an additional question in the Eurobarometer) dropped sharply during the 2008 crisis and has since remained at this lower level. An additional assumption would be a decline in the representativeness of the print media. Since 2011, the proportion of daily readers in France has significantly declined — from 30% to just 18% — especially when compared to the growing number of social media users.

In fact, when computing a sentiment index specific to respondents who trust the press or read it on a daily basis, we do not observe such a divergence between the sentiment index derived from the press and the Eurobarometer opinion survey. The views of those who trust or read the press broadly continue to align with the image conveyed by the press itself. While it is not possible to determine which of these two factors is most decisive (either distrust to press or less press readers), both may help certainly explain the divergence observed between our press sentiment index and French citizens’ opinions of the European Union.

Conclusion

By combining a vast new media dataset with modern AI language analysis, this paper demonstrates how machine learning can uncover richer and sometimes different narratives about public attitudes than those captured by traditional surveys alone. In particular, these results illustrate the multifaceted relationship between French citizens and the European Union as reflected through the written press when globally taken into consideration.