Social solidarity with Ukrainian and Syrian refugees in the twitter discourse. A comparison between 2015 and 2022

Since the beginning of the Russian invasion of Ukraine in February 2022, many Ukrainian residents have fled to European countries, most importantly to Poland, Hungary, Romania but also Germany. In the process, the refugees currently encounter a wave of social support and solidarity in the receiving countries. In the media, this high level of support and solidarity towards Ukrainian refugees is seen in stark contrast to the level of solidarity experienced by earlier groups of incoming refugees i.e. from Syria (but also Afghanistan, Iraq, etc.) during the so-called ‘immigration crisis’ in 2015 (Hardman 2022; Sharma 2022). This observation has raised concerns about a ‘hypocritical selectivity of Western hospitality’ (Pratt and Laroche 2022, no page number), or, more generally ‘solidarity bias’ (Visconty and Kyriazi 2022) in Europe. Solidarity bias privileges groups that are considered to resemble the native population in the receiving countries racially, culturally, and/or geographically (Pratt and Laroche 2022; Trope and Liberman 2010). So far, the debate on solidarity bias towards different groups of refugees has mostly relied on expert assessment and subjective impressions. It also contrasts with the predominant scientific portrayal of solidarity with incoming groups of refugees in the summer of 2015 (Della Porta 2018; Feischmidt et al. 2018; Gerhards et al. 2020; Youkhana and Sutter 2017; Vandevoordt and Verschraegen 2019), in spite of its geographical variation and fluctuation over time (Eger et al. 2022; Lahusen and Grasso 2018; Liebe et al. 2018).

We aim to inform the debate on potential bias in refugee solidarity based on quantitative data on (anti-)solidarity statements posted on Twitter during both waves of refugee immigration, in 2015 and 2022. Since previous research has argued that social behavior in online social networks impacts public opinion and political mobilization (Fenton 2008; Margolin and Liao 2018; Santhanam et al. 2019; Tufekci 2014), we want to analyze how solidarity with fleeing Ukrainians is expressed in online media, compared to solidarity expressions towards Syrian refugees. We build on our previous research on solidarity toward immigrants and refugees during the COVID-19 pandemic (Eger et al. 2022; Ils et al. 2021).

To assess social solidarity, we use the definition of social solidarity provided by Stjerno (2005: 2), which has been applied widely by other scholars (for example, Lahusen and Grasso 2018; Ils et al. 2021). According to this definition, social solidarity is ‘the preparedness to share one's own resources with others, be that directly by donating money or time in support of others or indirectly by supporting the state to reallocate and redistribute some of the funds gathered through taxes or contributions’ (Lahusen and Grasso 2018: 4). Based on this definition, we label Twitter posts (tweets) concerning refugees. Contestations of this solidarity are labeled and analyzed as expressions of anti-solidarity (cp. Burgoon and Rooduijn 2021; Cinalli et al. 2021).

In order to compare the current social media debates on refugees with previous ones, we assess the discourses about the influx of refugees to European countries in the context of the so-called ‘immigration crisis’ (2015) and the current influx of refugees from Ukraine. Based on theories of conditional solidarity (Montgomery et al. 2018) and previous research on the fluctuation of (anti-)solidarity discourses in online media, we test several empirical expectations regarding differences and similarities in discourses relating to the current and previous wave of refugees in Europe. By comparing social solidarity and anti-solidarity statements regarding different groups of refugees, our paper also makes a contribution to the discussion on ‘good’ and ‘bad’ refugees or who is welcome in host countries (Czymara and Schmidt-Catran 2017). While most previous research on solidarity towards different social groups has treated these groups as homogenous i.e. the disabled, unemployed or refugees in general, our analyses highlight that differences in the depiction and attribution of deservingness exist within such categorizations.

Our contributions are: Firstly, we present novel empirical evidence regarding changes in European solidarity debates. Secondly, we show to what extent solidarity with the fled Ukrainians has been expressed since the Russian invasion. Thirdly, we compare the current situation to the so-called ‘immigration crisis’ in 2015 and assess whether perceived group differences related to the country of origin among both waves of refugees explain differences in social solidarity expressed. In addition, we demonstrate how social media data can be used to assess shifts in discourse.

Arguably, different present and past waves of mass immigration in European countries are the result of distinct historical, social and political developments (cp. Buonanno 2017; Pries 2019). Related to this, European countries and the EU treat different groups of asylum seekers and refugees differently, based on their country of origin and the nature of immigration i.e. documented/undocumented and immigration reasons. This differential institutional treatment has raised much concern among academic scholars and the public (Niemann and Zaun 2018; Pries 2019; Visconty and Kyriazi 2022). It is unclear, however, how this differential institutional treatment affects social solidarity among the population in the receiving countries. Some have argued that what sets the 2015–16 ‘immigration crisis’ apart from other waves of mass immigration was that civil society compensated for nation states’ and the EU's failure to provide the aid needed at the time (Pries 2019). Social solidarity may thus be especially strong when institutional support is lacking.

Over the second half of 2015, an unprecedented number of 1.3 million refugees and asylum seekers, mostly from Syria, but also from Afghanistan, Iraq and other countries arrived in Europe (Verschraegen and Vandevoort 2019). Most of them (roughly one million) registered for asylum in Germany (BAMF 2016: 89). Whereas the majority of asylum seekers from Syria, Afghanistan and Irak fled the civil war raging in their home country and can thus be considered refugees, other asylum-seeking groups immigrated for work, education and other personal reasons originating in global socio-economic inequalities (Agustín and Jørgensen 2018; Buonanno 2017). A large number of these immigrants came as undocumented migrants. According to most research conducted in the social sciences, public discourse at the time was marked by a culture of welcome, signifying general support for incoming groups in need (Della Porta 2018; Feischmidt et al. 2018; Gerhards et al. 2020; Vandevoordt and Verschraegen 2019). Some studies, mostly based on the analysis of media coverage, suggest that the public debate was more ambivalent than this research suggests (Eberl et al. 2018; Greussing and Boomgaarden 2017; Jäckle and König 2017; Vollmer and Karakayali 2018). The divergent findings may at least partly stem from differences in study design (country/case selection; timeframe covered; method of analysis) and the nature of data analyzed (media analysis vs. surveys).

Studies tend to agree that both attitudes held by the general population and media discourse changed over the period between spring 2015 and spring 2016 towards less solidarity (Greussing and Boomgaarden 2017; Koos and Seibel 2019; Liebe et al. 2018; Vollmer and Karakayali 2018). Analyses of print media suggest a change in discourse away from characterizing refugees as victims towards stereotyping them as criminals (Greussing and Boomgaarden 2017). This shift in media coverage may have altered attitudes in the general population (Cinalli et al. 2021; Czymara and Dochow 2018). For Germany, scholars sometimes link this attitudinal turn with the high number of sexual assaults allegedly committed by male asylum seekers from African countries on New Year's Eve 2015/16. Evidence shows that subsequent declines in solidarity in the population were not directed towards immigrants per se. Instead, solidarity became more selective, signifying continued support for refugees and declining support for other immigrant groups (Czymara and Schmidt-Cantran 2017). Other singular events have allegedly contributed to this climatic change in Europe more generally, such as the terrorist attacks committed by asylum seekers in Paris and Brussels (Agustín and Jørgensen 2018). Whereas such events clearly altered the media discourse (Eberl et al. 2018), longitudinal evidence covering more than a snapshot of attitudinal change in the general population is scarce.

Since the Russian invasion of Ukraine in February 2022, millions of Ukrainians have been fleeing, most of them to the neighboring countries of Poland, Hungary, and Romania. Germany registered more than one million Ukrainian refugees between the end of February 2022 and March 2023 in the Central Register of Foreigners (BMI 2023). Whereas this number is comparable in size to the asylum seekers registered in 2015, the situation regarding the current wave of Ukrainian refugees in Germany and other EU countries is markedly different from the earlier wave of immigration in 2015–16. First, the incoming group is much more homogeneous and geographically as well as socio-economically close in comparison, seeking to escape a war unfolding within Europe instead of the Global South (Morrice 2022). In this sense, Ukrainian refugees ‘[…] are the archetypal refugees imagined by European nations when, in the wake of the holocaust and large-scale displacement following World War II, they established the Refugee Convention in 1951‘ (Morrice 2022: 251). Second, Ukrainian refugees enjoy free movement within the EU and have better institutionalized access to work and education in the receiving countries, compared to refugees from Syria and other non-European countries (De Haene et al. 2018; Morrice 2022). This facilitates a smoother transition and arrival in the receiving countries. Third, European countries today are better prepared to cope with a mass influx of refugees, compared to 2015–16, based on institutional processes of ‘learned solidarity’ (Agustín and Jørgensen 2018) during the earlier immigration wave.

To be sure, these differences should affect the extent and nature of public support and solidarity expressed on social media platforms. Crucially, however, such differences may or may not affect the nature of solidarity expressed i.e. solidarity bias. Following public claims of solidarity bias (Pratt and Laroche 2022), indicators of such bias would include statements relating to resemblance and difference between refugees and the native population in the receiving countries, either racially, culturally, and/or geographically. We use references to refugees’ countries of origin as our key indicator of solidarity bias.

Longitudinal research on attitudinal solidarity bias, or more generally attitudinal change towards refugees in Europe during 2015 and thereafter remains scarce in the social sciences, where evidence is mostly based on cross-sectional data i.e. qualitative expert interviews or survey responses of the population in the receiving countries. An exception is a study by Czymara (2021) who analyzed attitudes towards refugees in 22 countries using the seventh and eighth waves of the European Social Survey. His findings show that lower acceptance of refugees over time was not driven by trends in asylum applications but rather by general forces of social and cultural diversification among incoming groups. Importantly, in an earlier study, Czymara and Dochow (2018) provided evidence that media coverage on migration issues in Germany affected individual perceptions and concerns regarding the topic, with more coverage leading to more concerns. Liebe et al. (2018) conducted a repeated stated choice experiment regarding individuals’ preferences for establishing a new refugee or migrant home in their neighborhood as part of a multifactorial survey in Germany. Respondents were interviewed twice, first in November 2015 and again in November 2016. The findings show that a majority of respondents rejected the idea of building a refugee home in their vicinity in both waves. Only 20 percent of respondents initially approved the idea and were found to adopt a more negative attitude towards building a refugee/migrant home in 2016 (Liebe et al. 2018). Czymara and Schmidt-Catran (2017) examined changes in the public acceptance of immigrants in Germany between the summer of 2015 to after the sexual assaults of New Year's Eve 2015/16. They compared changes in public acceptance along different immigrant characteristics and found that refugees, including Muslims, were highly accepted across both panel waves, whereas the acceptance of immigrants from Arab or African countries dwindled. In addition, female respondents stated preference for male immigrants in the summer of 2015 disappeared after the New Year's Eve events 2015/2016. Another longitudinal study, which was conducted jointly by social and computational scientists, assessed change in migrant and refugee solidarity and anti-solidarity in Europe before and after the onset of the COVID-19 pandemic (Eger et al. 2022). Eger and colleagues collected and analyzed a large text corpus of migration- and refugee-related tweets between September 2019 and June 2021. In contrast to earlier studies, which compared a few data points, they also used daily data to assess change over time. They found both solidarity and anti-solidarity towards immigrants and refugees to fluctuate markedly, with peaks during the onset of the pandemic and in relation to politically salient events of high media coverage, which related to the issue of migration. Importantly, neither this paper nor a joint paper published at an earlier stage of this research (Ils et al. 2021) collected data on refugee solidarity during the onset of the ‘long summer of migration’ in 2015, or after June 2021.

Another research gap to which our paper speaks concerns the explicit comparison of solidarity expressed towards different groups of immigrants over time. Apart from the aforementioned study by Czymara and Schmidt-Catran (2017), longitudinal research on this issue remains scarce. Bansak et al. (2016) conducted a cross-sectional conjoint experiment in which eligible voters in fifteen European countries evaluated profiles of asylum-seekers that randomly varied their personal characteristics. They found more support for asylum-seekers with higher employability, more severe vulnerabilities, and Christian rather than Muslim religious background. In line with this latter finding, Czymara (2020) found that Native Europeans’ attitudes toward Muslim immigrants were especially more hostile in countries where political elite discourses were more exclusionary. In contrast, views on ethnically similar immigrants were hardly affected by elite discourse. Another comparative study by Cinalli et al. (2021) highlights the role of public news media in shaping public opinion towards refugees (see also Baltov 2022; Czymara and Dochow 2018; Fotopoulos et al. 2022).

Existing literature in (computational) social science has mostly analyzed (anti-)solidarity expressions in online media by focusing on a particular crisis event for a very short period of time (days or weeks) and selecting specific topics. The study by Eger et al. (2022) provides a recent review regarding migration-related (anti-)solidarity assessments in Natural Language Processing – NLP (see Khatua and Nejdl 2022 for a broad review regarding social media and migration). One strand of this research has a strong focus on mainstream media, investigating, for example, differences in perspectives and narratives between leftist and conservative mainstream media and social media discourses on platforms such as Twitter (Cinalli et al. 2021; Nerghes and Lee 2019; Wallaschek 2019; Wallaschek 2020). Another strand of research investigated stances towards refugees and migrants by focusing on social media platforms, such as Facebook (Capozzi et al. 2020; Hrdina 2016), Instagram and Pinterest (Guidry et al. 2018) and, most frequently Twitter (Aswad and Menezes 2018; Calderón et al. 2020; Gualda and Rebollo 2016; Khatua and Nejdl 2022). In addition, one study from computational science assessed solidarity expressed by social media users by means of emojis, shortly after two selective crisis events (Santhanam et al. 2019). This study did not focus on refugee solidarity but on solidarity expressed towards the victims of Hurricane Irma in 2015 and the people affected by the terrorist attacks in Paris in 2017. The authors performed hashtag-based manual annotation, thereby ignoring the actual content of the tweets. They utilized an LSTM (Hochreiter and Schmidhuber 1997) network for automatic classification. Most of this research employed techniques of supervised and unsupervised machine learning (cp. Khatua and Nejdl 2022). In addition, the vast majority of studies investigated positive or negative online behavior only after the onset of a crisis and for a very short time, without considering important fluctuation in pro- or anti-social behavior (Eger et al. 2022). Research focusing on (anti-)social online behavior towards ethnic, religious, or other minority groups tends to be biased towards investigating antisocial behavior, such as the rise of hate speech, without considering the possibility that the salience of these groups in political discourse may increase both pro- and antisocial online behavior simultaneously (Ibid.). Thus, research started highlighting the importance of longitudinal and long-term observation and other design choices, including sampling, to assess attitudes towards migrants, refugees, and other vulnerable groups (Eger et al. 2022; Czymara and Schmidt-Catran 2017; Khatua and Nejdel 2022).

There is very little research on (anti-)solidarity in the NLP community, but plenty of research on social biases relating to gender, ethnicity, etc. (Bolukbasi et al. 2016; Caliskan et al. 2017; Sweeney and Najafian 2019; Nadeem et al. 2021)), also over time (Walter et al. 2021). This research makes use of unsupervised techniques, e.g. word representations derived from massive amounts of text data and queries them for social biases or makes use of partially masked text templates (such as ‘He is very [MASK]’) which language models must complete. Social biases and statements of (anti-)solidarity are related because biases are oftentimes the reason for statements of anti-solidarity.

In sum, the state of research suggests that social solidarity bias may exist in Europe, though direct evidence comparing social solidarity towards refugees from Syria (and other salient countries) in 2015 and thereafter and Ukraine since February 2022 is lacking. A notable exception is the media study by Baltov (2022) who assessed differential construction patterns relating to the refugee waves from Syria and Ukraine over the last 10 years in five Bulgarian online media platforms. This study, however, did not focus on changes in social solidarity towards refugees. In the absence of robust longitudinal data covering both waves of intensified refugee migration to the EU countries, accounts of attitudinal differences and social bias towards refugees mostly rely on media interviews with experts and practitioners (for example, Johnson and Bräuer 2016; Schöppner 2016; Ritzmann 2016; or Hardman 2022; Sharma 2022).

Empirical studies investigating ‘conditional solidarity’ and ‘hierarchies of solidarity’ suggest that solidarity bias may exist, but this research has not looked at within-group differences, such as different groups of refugees (Kyriakidou 2021; Montgomery et al. 2018). Instead, empirical evidence points to different hierarchies or conditions of solidarity with different groups in need, for example pitting opinions regarding the welfare deservingness of the unemployed, the disabled and refugees against each other (Montgomery et al. 2018; van Oorschot 2000; 2006). This research can nevertheless inform our empirical expectations because it has identified factors potentially influencing solidarity towards those in need.

In particular, research by van Oorschot (2006) and Montgomery et al. (2018) suggests that solidarity is higher towards social groups that are perceived as (1) being in need but not responsible for their situation; (2) finding themselves in a highly precarious position; (3) having supported society earlier through own contributions; (4) culturally similar; and (5) sharing similar co-orientations. According to the first three deservingness criteria, social solidarity is conditional and hierarchical, but not necessarily biased. The last two deservingness criteria, in contrast, point to solidarity bias, favoring social groups that are perceived as either racially, culturally, and/or geographically close.

Based on the state of research, we test several empirical expectations regarding differences and similarities in discourses relating to the current and previous wave of refugees in Europe. We argue that especially the first two criteria of conditional solidarity apply to refugees. Whereas refugees from both immigrant waves can be considered being in need but not responsible for their situation (criterion 1), refugees from Syria (and other countries hit by civil war) found (and still find) themselves in a more precarious position when trying to enter EU countries compared to Ukrainian refugees. More specifically, refugees from Syria and other non-European countries often travel to the Eurozone under life-threatening circumstances and find themselves trapped at the borders in overcrowded refugee camps for extended periods of time. In contrast, Ukrainian refugees are free to enter the EU and travel within it, with better access to accommodation, work and education upon their arrival. Based on the different degrees of precariousness (deservingness criterion 2) among refugees upon their arrival in Europe, we would expect to find a higher share of refugee related solidarity tweets (relative to all tweets) during the 2015–16 wave of immigration compared the current wave of immigration (Hypothesis 1a).

To be sure, this hypothesis runs counter to the expectation of social solidarity bias based on perceived socio-cultural distance. In line with this latter perspective, the conditional solidarity literature suggests that perceived cultural similarity (deservingness criterion 4) and the perception of shared co-orientations (deservingness criterion 5) might lead to more solidarity with the current wave of incoming refugees, because this group is considered socio-culturally and geographically closer than refugees in 2015–16. Based on these latter criteria, we thus expect to find a lower share of refugee related solidarity tweets (relative to all tweets) during the 2015–16 wave of immigration compared the current wave of immigration (Hypothesis 1b).

Earlier research based on social media data shows that the salience of (anti-)solidarity discourses related to the precarious situation of migrants and refugees fluctuates markedly over time, sometimes disrupting the longer-term trends in public opinion for a couple of days (Eger et al. 2022). Both solidarity and anti-solidarity statements tend to skyrocket simultaneously and in relation to singular political events. It is thus likely that such patterns also mark the two waves of refugee immigration we study in this paper. We thus expect to find similar incidences of fluctuation in (anti-)solidarity over time for both waves of refugees (Hypothesis 2).

Whether and to what extent social solidarity is driven by the perceived lack of control and neediness of the group in need (deservingness criteria 1 and 2) or by perceived psychological closeness i.e. regional closeness or distance (deservingness criteria 4 and 5), is an important but empirically open question. If the latter is the case, evidence would point to solidarity bias. We would thus expect a higher share of anti-solidarity statements in the 2015–16 wave of refugees emphasizing refugees’ country of origin (3a) and/or a higher share of solidarity statements in the current wave of refugees emphasizing refugees’ country of origin (3b).

To assess our empirical expectations, we collected tweets about migration and refugees posted between 01.01.2015 and 31.08.2022. To crawl the data, we utilized the Twitter Academic API (an official Twitter interface for downloading tweets). We used a query to retrieve tweets with specific hashtags related to refugees and migration. The selection of hashtags we utilized was adopted from Ils et al. (2021). These hashtags were optimized by Ils and colleagues for the immigration discourse for the period 2019–2020. Since we focus on Europe, we only considered tweets posted by users from Europe or for which the geotag of the tweet indicated a European location. This results in a data set with 556,000 tweets in English. Additionally, we collected German-language tweets, with no country-level limitation, as we assume that most German-language tweets are sent by users in Europe. There are 1,764,601 German language tweets in the time span using the hashtags. In total, we collected 2,320,601 tweets for English and German together. The used hashtags are presented in the appendix.

We first annotated a subsample of the collected tweets for solidarity and anti-solidarity between January 2015 and July 2022. The annotation guideline is available in the appendix. We distinguish three categories: solidarity, anti-solidarity, and other (Table 1). A tweet is annotated as showing signs of solidarity if it indicates support for people in need, including willingness to help and/or share own resources as well as tweets expressing gratitude towards others sharing resources and/or helping. In addition, tweets criticizing European institutions for not offering enough support for refugees are also labeled as solidarity. Tweets that contradict the aforementioned criteria and/or show nationalistic tendencies in relation to refugees and/or advocate closed borders are classified as anti-solidarity. In addition, the category other is used for tweets that can be interpreted ambiguously or for tweets that do not fit into the (anti-)solidarity categories. In total, we annotated 6912 tweets, 3106 in English and 3806 in German from January 2015 to July 2022. We oversampled tweets from September 2015 and March 2022. The annotation was performed by a student assistant who was trained in advance. The kappa agreement with previously collected expert annotations is 0.63, which is considered an acceptable agreement but lower than the previous agreements reported in Eger et al. (2022). To ensure the accuracy of labels for the anti-solidarity training examples, we implemented a verification process. An expert reviewed all tweets that were previously marked as anti-solidarity by the student assistant.

The annotated tweets were used to train a supervised machine learning model. To this end, we employed state-of-the-art language models, e.g. XLM-R (Conneau et al. 2020). This is a pre-trained deep learning model, which we fine-tuned with our annotated data to obtain a high-quality classifier. With that classifier, we can automatically label all collected tweets.

We follow Eger et al. (2022) and Ils et al. (2021) to induce our classifiers and leverage previously annotated data from them, obtaining 6,912 human annotated tweets together with our newly collected annotations. To get more stable training results, we made three random splits, each of which consists of a training set of size 4,147, a development set of size 1,382, and a test data set of size 1,383. We fine-tuned xlm-roberta-base models (Conneau et al. 2020) and performed grid search over the following hyperparameters: a maximal learning rate of 2e-5 or 4e-5 for the linear learning rate scheduler with 100 warmup steps, a layer-wise learning rate decay of 1, 0.9 or 0.8, a dropout rate on the classifier layer of 0.1 or 0.3, and a training epoch number of 3, 4, 5, or 6. We trained all models with a batch size of 64, a maximal token length of 128 and an AdamW optimizer, using HuggingFace Trainer API; we kept the other settings default. As a result, 48 models were trained on each split. The best model was selected according to the average macro F1 on the developemnt set over the three splits, which is the one trained with 4e-5 learning rate, 0.1 dropout rate, 0.9 layer-wise learning rate decay and 6 epochs; it obtains an average macro F1 of 0.717 on dev set and 0.693 on test set. These numbers are lower than in Eger et al. (2022) but we consider them acceptably high for a difficult social science problem. We observe that the lower macro F1 is due to the models’ bad performance on category ‘Other’ (0.489 F1 score on test set); for the other two categories, the F1 scores are much higher (0.805 for ‘Solidarity’ and 0.784 for ‘Anti-solidarity’). In a next step, we used the best model identified to automatically label the unlabeled tweets. To validate the legitimacy of our classifier, we drew 132 random samples from the automatically labeled tweets for our target time period, namely September 2015 and March 2022, to manually evaluate the classifier. The results indicate acceptable F1-Scores for the specified period, ranging from 0.71–0.78. These scores contribute to a macro average of 0.73.

On the automatically labeled data, we performed a trend analysis to show peaks in the Twitter discourse. Therefore, we plotted the number of tweets that are classified as including solidarity and anti-solidarity expressions. This trend analysis allows us to assess change over time. We use (anti-)solidarity statements relating to refugees’ countries of origin as our key indicator of solidarity bias. We use keywords relating to the countries of origin and nationalities to depict group differences. Table 2 provides an overview of the collected data and the predicted labels. Overall, there are more German than English tweets, and the number of tweets classified as solidarity statements is higher than the number of tweets showing anti-solidarity.

In 2015, many Syrians fled their homes and around 1.3 million fled to the EU. In comparison to previous years, the number of incoming refugees to the EU increased strongly. The collected tweets demonstrate a welcoming climate in Europe and Germany in the summer of 2015. The number of tweets that express their solidarity with refugees and migrants during the summer months of 2015 is the highest for the period of collected tweets from 2015 to 2022. In September 2015, for both German and English language tweets, the highest total number of tweets related to migration and refugees can be observed. In total, our dataset contains 240,369 tweets in that month that were classified as expressing solidarity with refugees. Over the same period, only 17,592 tweets were classified as showing anti-solidarity. The ratio between solidarity and anti-solidarity tweets is 13.7, meaning that the number of solidarity tweets is fourteen times higher than for anti-solidarity tweets. The log scale graphs in the appendix allow for assessing differences in the ratio over time (Figure A1).

The ratio between solidarity and anti-solidarity tweets declined to 2.9 after the sexual assaults on New Year's Eve 2015/16 in Cologne (‘Köln’), from January 2016 onwards. Especially in Germany, anti-solidarity tweets peaked at the beginning of 2016. Plotting the most often used terms in the tweets from January 2016, we can observe that especially words related to Cologne appear in the collected tweets and that a high share of German words occur in that month (Figure 1). In the summer of 2018, from June to August especially the issue of sea rescue (‘Seenotrettung’) was discussed. The issue was particularly brought to the attention of the European Union by Italy's then Interior Minister Salvini, who wanted to prevent rescue ships from docking in Italy. In March 2020, at the onset of the COVID-19 pandemic, Turkey's President Erdogan declared to open borders for refugees. This was a consequence of a dispute with the EU on how to handle Syrian refugees. At the end of March 2020, the EU announced funding for new refugee camps in Greece. The conditions and status of refugee camps were also mentioned in tweets in September 2020, when the Greek refugee camp Moria was destroyed by a fire (cp. Eger et al. 2022).

In sum, given the recent media debates claiming solidarity bias in favor of Ukrainian refugees over Syrian refugees, the peak at the onset of Ukrainian migration towards their neighboring EU countries appears surprisingly small in magnitude and short in duration.

Table 3 shows the ratio between solidarity and anti-solidarity tweets in September 2015 and March 2022 as well as the ratio for the whole period covered in our data. To assess our hypotheses 1a and 1b, as well as 3a and 3b, we decided to focus on one specific month during each immigration wave to capture a similar phase in the immigration process. Higher values mean that a relatively higher share of solidarity tweets than anti-solidarity statements was posted. For the two months where a peak in Twitter activity was observed, we find that tweets in September 2015 are far more likely to express solidarity than anti-solidarity, especially if they are written in English. But also German tweets are over ten times more likely to state solidarity than anti-solidarity. For March 2022, after the beginning of the Russian invasion of Ukraine, tweets related to migration and refugees show a lower ratio of solidarity compared to September 2015. These findings lend support to hypothesis 1a, suggesting more solidarity expressions in favor of the incoming, mostly non-European refugees in 2015, compared to incoming refugees in 2022 who mostly fled the Ukraine. This general picture is confirmed when considering the full period covered in our data. For the entire time span covered (January 2015 to September 2022), we observe that tweets declaring solidarity are more frequent. Overall, solidarity tweets are 4.6 times more frequent than anti-solidarity tweets. In general, English-language tweets with a European geotag have a higher degree of solidarity than German-language tweets without a geotag. Interestingly, even the solidarity ratio for the complete time span covered among English-language tweets is higher, compared to March 2022, after the beginning of the Russian invasion in February. In contrast, German language tweets show a higher solidarity ratio in March 2022 than over the full period from 2015 to 2022.

To account for the prevalence of duplicate tweets, particularly retweets, within the collected dataset, Table A3 in the appendix provides the ratio calculations after removing text duplicates. Overall, the ratios continue to exhibit a general alignment in terms of direction. However, noteworthy differences emerge. For instance, when text duplicates are eliminated, tweets written in English demonstrate a lower solidarity ratio in September 2015. Additionally, in March 2022, tweets are comparatively more inclined to express solidarity statements (across languages). In sum, also these findings lend support to hypothesis 1a and lead us to reject hypothesis 1b.

Table 3 includes two additional columns to present the ratio for tweets explicitly mentioning nationalities or countries. The analysis focuses on two specific periods: September 2015, when a particularly high share of individuals fled from Syria, Iraq, and Afghanistan, and March 2022, when people began to flee in high numbers from the Ukraine to EU countries. Particularly for tweets written in English in March 2022, the ratio of tweets referencing Ukrainians is significantly higher compared to tweets within the same period that do not mention Ukrainians (179.10–30.39). Similarly, German tweets also exhibit a higher ratio, although the difference is comparatively smaller (6.95–4.27). In contrast, for the month of September 2015, the observed differences in ratios are minor (Table 3).

In addition to analyzing the proportion between solidarity and anti-solidarity statements posted, it is crucial to explore whether social solidarity is primarily driven by the perceived lack of control and neediness of the group in need (deservingness criteria 1 and 2) or by the perceived psychological closeness, such as regional proximity or distance (deservingness criteria 4 and 5). We expected a higher share of solidarity statements in the current wave of refugees emphasizing refugees’ country of origin (hypothesis 3b) as well as, or alternatively, anti-solidarity statements in the 2015–16 wave of refugees emphasizing refugees’ country of origin (hypothesis 3a), as indicating solidarity bias related to psychological closeness. We use a simple setup to examine the countries and nationalities mentioned in tweets. The primary objective is to ascertain the prevalence of specific groups within tweets categorized as either anti-solidarity or solidarity for the two months of interest.

Table 4 displays the percentages of tweets, categorized by period and label, mentioning nationalities and/or countries. For simplicity, we restrict our analyses to the most salient origin countries in both refugee waves. The percentages of posts mentioning Syria, Iraq, and Afghanistan are combined, while separate percentages are shown for the Ukraine.

In September 2015 the relative number of solidarity tweets mentioning Syria, Afghanistan or Iraq, is slightly higher than the number of anti-solidarity tweets. This finding is not in line with the expectations based on our first solidarity-bias hypothesis (3a). However, in March 2022, the percentage of solidary tweets mentioning the Ukraine is much higher than the percentage of Ukraine-related anti-solidarity tweets, in line with our second solidarity-bias hypothesis (3b). In addition, though Syria, Afghanistan and Iraq are mentioned rarely during this month, (anti)solidarity shows the opposite pattern in these tweets. Manual inspection of 200 randomly selected anti-solidarity Tweets from March 2022 indicates that roughly 20 percent of these Tweets favor Ukrainian refugees over other asylum seeking groups. Taken together, these findings point to a solidarity discourse favoring the geographically and culturally closer Ukrainians over refugees from other world regions, at least in March 2022. The observed differences between German and English tweets in March 2022 in Table 4 indicate variations in the discourse surrounding the topic. Although both languages exhibit higher percentages of solidarity tweets, German tweets labeled as anti-solidarity also contain a significant proportion of references to Ukraine. This can be traced back to the discussion of the perceived misuse of Ukrainian's right to enter the EU by individuals from other nationalities. In the case of the UK, the discussion primarily revolves around the limited assistance offered by the Conservative Government to the fleeing Ukrainians during the specified period.

Three example tweets depict the situation in March 2022:

German tweets classified as anti-solidarity translated: ‘Just a small number are authentic Ukrainian #refugees. They have newly issued, yet legitimate, Ukrainian passports.’ or ‘#Ukraine #refugees: The label reads “Poor Ukrainians” yet inside are unappreciative Arabs’

English tweet classified as solidarity: ‘Why does the UK mandate visas for #Ukraine #Refugees, barring entry even to those with family ties there?.. Why?’

Analyzing 2.3 million tweets regarding migration and refugees, we found differences in how often tweets are expressing solidarity and anti-solidarity over time. From January 2015 to August 2022, we registered several peaks in the Twitter discourse and examined differences between September 2015, the onset of the so-called European ‘immigration crisis’, and March 2022, just after Russia started to invade Ukraine. Importantly, our findings show that the ratio of tweets expressing solidarity relative to anti-solidarity was much higher in September 2015 than in March 2022.

Clearly, this finding runs counter to the expectation of social solidarity bias favoring Ukrainian refugees over refugees from non-European countries per-se. Instead, it points to the importance of the higher degree of precariousness among refugees from Syria, Afghanistan and Iraq upon arrival in Europe, compared to the degree of precariousness among Ukrainian refugees, in line with our first hypothesis (Hypothesis 1a). The counter-expectation of finding a lower share of refugee-related solidarity tweets during the 2015–16 wave of immigration compared to the current wave of immigration (Hypothesis 1b) has to be rejected. These findings tie in with earlier studies that have established the importance of precariousness through no fault of one's own for being considered deserving of social solidarity (van Oorschot 2006). Our findings go beyond previous research by indicating that deservingness matters not only when considering solidarity with different groups in need (for example, the unemployed vs. immigrants), but also when it comes to solidarity differences within groups (different immigrant groups).

Our interpretation is backed up when comparing the political events and their media coverage at the time. The summer of 2015 was marked by several political key events. The then chancellor of Germany, Angela Merkel, gave a speech on August 31st, in which she uttered the well-known phrase ‘Wir schaffen das!’ (‘We can do it)’, a clear commitment of Germany's willingness to take in large numbers of refugees. The following day, in opposition to Germany's position, Hungary closed its capital city Budapest's main train station to refugees, leading some of them to set off on foot towards Austria. On September 4th, both Germany and Austria opened their borders. In several German cities, trains with refugees were received with friendly applause from the locals in the stations. Those events were covered by the mass media to a large extent. By comparison, media coverage in March 2022 focused to a greater extent on the ongoing war, rather than the situation of Ukrainian refugees. This shift in media attention was probably related to the fact that fleeing Ukrainians were unanimously welcomed in Poland and other countries, with little political concern regarding how their influx should be handled. Likely related to this, a lack of media coverage regarding the situation of refugees at the borders might have played a role in Twitter users’ lower activity to express their solidarity and anti-solidarity, compared to the situation in 2015.

Comparing the Twitter discourses regarding both waves of incoming refugees, we expected to find similar incidences of fluctuation in (anti-)solidarity over time for both waves of refugees (Hypothesis 2). In line with this expectation, we found fluctuation and an escalated rise of both solidarity and anti-solidarity statements immediately following politically salient events with high media coverage, but the frequency and magnitude of these peaks were much higher in 2015–16, compared to 2022. This finding leads us to reject Hypothesis 2. We believe that differences in media coverage might explain this. However, further analyses linking news media reports to the rise and fall of (anti-)solidarity expressions would be needed to validate our interpretation.

Our findings suggest that it is not the onset of immigration itself, but the political events surrounding it that trigger expression of (anti-)solidarity in large numbers. This finding confirms the dynamics that have been identified in earlier research on (anti-)social behavior on social media platforms (Awal et al. 2020; Eger et al. 2022). Though we could not examine this directly, our finding is also in line with studies highlighting the role of news media reporting on public opinion towards refugees (Cinalli et al. 2021; Fotopoulos et al. 2022). Going beyond the state of research, our findings show how much more intense Twitter activities were during the summer of 2015, compared to the spring of 2022. The number of solidarity-related tweets was 13 times higher. All in all, our results thus confirm the body of research suggesting that public discourse in the summer of 2015 was marked by a culture of welcome, signifying support for those in need (Della Porta 2018; Feischmidt et al. 2018; Gerhards et al. 2020; Vandevoordt and Verschraegen 2019).

Finally, we assessed the nature of differences in social solidarity towards different groups of refugees, focusing on perceived psychological closeness i.e. regional closeness or distance, our indicator of solidarity bias. Whereas our results do support the expectation of a higher share of solidarity statements in the current wave of refugees emphasizing refugees’ country of origin (hypothesis 3a), we did not find more anti-solidarity in tweets mentioning refugees’ countries of origin in 2015–16. Still, the percentage of solidarity tweets mentioning Ukraine in March 2022 is much higher, compared to the percentage mentioning Syria, Iraq and Afghanistan in September 2015. In addition, and this is noteworthy, among tweets mentioning refugees’ country of origin in March 2022, a very high share is solidary with Ukrainians while anti-solidarity prevails among tweets mentioning Syria, Iraq or Afghanistan. Whereas, this finding is limited to a selective subset of tweets, we also found evidence that about 20 percent of anti-solidarity tweets in March 2022 combined pro-Ukrainian statements with negative stereotypes against other groups of refugees, indicating solidarity bias.

Taken together, our findings speak to the literature on solidarity bias, confirming the role of groups’ vulnerabilities, as well as of socio-cultural preferences in the support of migrants and refugees (Bansak et al. 2016; Czymara 2021). We conclude that the present wave of incoming Ukrainian refugees does not receive more support among the Twitter-using public, compared to refugees from Syria, Iraq and Afghanistan in 2015–16. Rather the opposite is the case and according to our data, the higher vulnerability among refugees from Syria, Iraq and Afghanistan helps explaining this. However, given the present situation, in which refugees from the Ukraine as well as Syria, Iraq and other countries continue to seek shelter in EU countries, we find some indication of solidarity bias among tweets in favor of Ukrainians, which seems to be related to the higher degree of perceived psychological closeness of EU citizens with Ukrainians.

This study clearly has limitations, especially due to the long-term focus on one social media platform. Over time, different users engaged with the platform, and both general and individual usage patterns may change. Moreover, Twitter users and people who write tweets are not representative of the population. In addition, we work with two different sets of tweets, that is all tweets in German language, and tweets in English language for which geographical information points to a European location. It has already been shown that English language tweets posted by users in countries in which English is not an official language contain on average a higher share of solidarity statements (Eger et al. 2022). Another limitation is that the hashtags we collected for this analysis have originally been focused on the immigration discourse as it unfolded prior to the war in Ukraine. Whereas keeping hashtags constant across both immigration waves is a sensible design choice to facilitate comparison over time, we might have missed a substantial part of the current immigration discourse, which centers on the Ukrainian refugees using different hashtags. To be sure, using a broader list of hashtags or migration-related words in the search query for the Twitter API might have changed the results. Because the academic Twitter API is no longer available, we cannot investigate this aspect further. However, we have conducted a sensitivity analysis of the hashtags we used (cp. Beese et al. 2022), by removing one hashtag at a time from our list of hashtags and considering hashtags individually (see online appendix). Whereas results based on removing single hashtags are in line with the findings reported in this paper, namely more solidarity in 2022 than in 2015, our sensitivity analysis further shows that individual hashtags (e.g. ‘#refugee(s))’ have different solidarity ratios in 2015 vs. 2022 i.e. the tweets containing those hashtags show more solidarity in 2022 than in 2015. Thus, in some cases, findings based on individual hashtags are contrary to the overall trend we reported in this paper, and thus more in line with the country- and nationality-specific analysis presented in Table 4. This indicates a nuanced picture overall and emphasizes the dependence of our results on the selection of data: selective subsets may yield different conclusions.

Further research should examine the peaks in (anti-)solidarity discourses in more detail and link traditional media coverage to Twitter behavior. In addition, the extension to several European languages, such as Polish, would be useful in order to gain further insights into the role played by the national context regarding solidarity levels and bias. In general, there is much potential to apply advanced deep learning strategies to enhance our substantial understanding of migrant and refugee-related social media discourses.

At present, the current wave of migration from Ukraine is well underway and it will take time to see how solidarity discourses develop in comparison to the discourses that surrounded the ‘long summer of migration’ in 2015. In sum, our findings highlight that Twitter using populations in Europe continue to show more solidarity than anti-solidarity to those in need, even though solidarity bias seems to emerge when culturally different immigrant groups are compared.

The authors would like to thank the participants of the ConTrust lunch seminar 04–2022, the participants of the Workshop ‘The Comparative Politics of Solidarity’, held at Politicologenetmaal 2023 in Leuven, Belgium, and anonymous reviewers for their constructive comments on earlier drafts of this paper.

No potential conflict of interest was reported by the author(s).

The Python code for data cleaning and analysis is available at https://zenodo.org/record/8232678.