A Text Mining Approach to Determinants of Attitude Towards Syrian Immigration in the Turkish Twittersphere

Introduction

In the first half of 2020, three intertwined events occurred in Turkey over a very short span of time: the emerging COVID-19 pandemic, the Turkish offensive to fight combatant armies on Syrian ground, and the subsequent cross-border movement of refuge-seeking migrants from Syria in Turkey. This combination of events significantly reshaped the public attitude towards immigrants in Turkey.

Taking the specific context of the relationship between the host-society and migrants into consideration, we trained a classifier model to identify attitudes towards immigrants in a large number of public Twitter messages (i.e., tweets). Then, we used topic modeling techniques to identify major themes within the tweet corpus. Utilizing deep learning-based novel computational methods, this paper aims to: (i) reveal discursive construction of “immigrant talk” in social media, and (ii) discover the association between major macro-level, real-life events and social media discussions.

A strong understanding of the sociopolitical context of Syrian immigration to Turkey is required for our objectives in this paper. The Syrian refugee crisis—an unintended consequence of peaceful protests that turned into civil war—has affected the social and political landscape of the Middle East and Europe through the cross-national movement of millions of immigrants since 2011. In 2015, Europe faced one of the largest waves of migratory movements, as refugees became more visible to the European public (BBC News, 2020). Syrian migrants were met with mixed reactions in host countries, depending on their local political context.

In Turkey, immigration has found its way into increasingly heated public discussions in a politically divided society, during a time when the migration agreement between the European Union and Turkey in 2016 brought Syrian refugees to the domestic political stage (European Council, 2016). Turkish political leadership instrumentalized their ostensibly liberal policy of admitting high numbers of refugees to take the moral high ground over European countries (BBC News, 2015; Nebehay & Farge, 2019; Voice of America, 2018). When expressing his views on the Syrian migrant crisis, Turkish president Erdogan used civilizationist rhetoric, which incorporated Islamist anti-colonial themes with Turkish nativist populism to forge a new civilizational identity against the “inhumane,” “corrupt,” “profit-oriented,” and “self-interested” Western civilization in a similar tone to European far-right populist movements that contradictorily merged elements of Christianism, secularism, and liberalism as the European civilizational identity othering Islam and immigrants from Muslim-dominant countries (Brubaker, 2017).

Beyond the political discourses that instrumentalized Syrians, there are serious obstacles to the integration of immigrants and refugees into the wider Turkish society at the social policy level. According to the Migrant Integration Policy Index, Turkey ranked last consecutively in 2010 and 2015. The latest index in 2020 reports considerable improvements in Turkey, particularly in the areas of education and access to healthcare. Nevertheless, Turkey has not yet overcome its challenges regarding civic engagement and naturalization of immigrants (MIPEX, 2015, 2020). Historical and ideological factors play a significant role in the failure of a successful immigrant and refugee integration in these specific areas of social policy. This research examines the context of reception and will contribute to the conceptual discussion on the underlying reasons behind the shortcomings of Turkey’s current social policies.

From a historical perspective, Turkish society’s peculiarities are particularly relevant to refugee and immigrant integration. Inheriting a greatly diverse population from the heydays of the multiethnic Ottoman Empire, the Turkish power elite has carried out a massive ethnic homogenization in the nation-building project. The fin-de-siècle Ottoman Turkey witnessed a series of conflicts, massacres, forced displacements, and population exchanges that paved the way for today’s modern Turkish nation-state (Boyraz, 2017; White, 2014). The state-sanctioned nationalist ideal of this era has long denied or dismissed visible manifestations of ethnic identity and cultural markers (Keyman, 2013). Specifically after enacting radical Westernizing reforms, Turkish nationalism reinforced anti-Arab sentiments by stigmatizing Arabs with their “inevitable backwardness” (Eldem, 2010). The public response in Turkey to Syrian immigrants was hence channeled through this ideological constellation of Turkish modernization.

Therefore, this study considers both the cultural and political climate of the host society when analyzing immigrant receptivity on Twitter, where the data is accessible and searchable. To bridge the disciplinary gap between social scientific literature on immigration and the burgeoning field of computational textual analysis, this paper interweaves a historically grounded approach with empirically evidenced research. We delve into the construction of competing narratives of the recent Syrian immigration to Turkey and power relationships expressed through those narratives on Twitter. This study aims to discover how those narratives change over time through linking them with real-life events happening outside the social media platforms to have a solid understanding of the determinants of attitude towards immigration in Turkey.

Literature Review

There is no remarkable socio-historical account of how Turks and Arabs see each other after the dissolution of the Ottoman Empire. Common cognitive categories, ethnic schemas, frames, and stereotypes in the dominant Turkish imaginary have rarely been addressed in scientific literature. Additionally, most studies on the Turkish-Syrian relationship either merely focused on Ottoman imperial history or modern political history. However, studies on the media portrayal of Arabs in early Republican Turkey suggest that the Turkish media in the interbellum era exhibited certain Orientalist features, either embracing Orientalist notions regarding the East or projecting the Orientalist depiction to a “more Eastern” society (Eldem, 2010; Lafi, 2014).

The proliferation of information and communication technologies has restored the long-broken ties between modern Arab and Turkish societies. The political movement of the Arab Spring in the early 2010s drew ample attention in the Turkish media. Most of the media outlets covered events in the Middle East as freedom protests while pointing Turkey as a role model for democracy and modernization in the region (Dağtaş, 2013). However, as the Syrian Civil War yielded deeper troubles, the refugee problem appeared as the primary matter of concern for the media. Turkish media has been sharply divided over the portrayal of Syrian refugees. Pro-government newspapers applauded the Turkish government for addressing a humanitarian crisis with an emphasis on the nation’s moral and religious obligations (Efe, 2015). They covered Turkmens, Syrian citizens of Turkish origin, to create an ethnic affinity with the Turkish audience. The anti-government Turkish media, however, instrumentalized the problems of refugees to criticize the government for mishandling the situation. Safety concerns and the economic burden of immigrants were two prominent themes in those media outlets. A newspaper close to a far-right nationalist ideology, for example, highlighted the economic burden of Syrians as soon as their number was less than a hundred thousand in Turkey (Efe, 2015). Similarly, popular Turkish print media represented Syrian immigrants as an economic burden while Turks were portrayed as the victims of immigration (Onay-Coker, 2019). Syrian immigrants, according to this media framing, were the “cowards” who did not fight for their country and were instead motivated by self-interest and opportunism. Syrian immigrants were also associated with crime and violence in Turkish news, which provokes a negative sentiment in addition to the evoked feelings of burden and cowardice (Onay-Coker, 2019). One print media analysis study (Doğanay & Keneş, 2016) identified three major themes in select Turkish newspapers. The first theme framed the Syrian presence as a “threat,” while the second common media discourse objectified Syrian people in numbers and economic figures, with the sentimentalization of Syrian reality being the last discursive dimension.

The media portrayal of Syrian refugees in European and North American societies highlighted similar themes. Vollmer and Karakayali (2018) studied the agenda-setter print and online media in Germany. They concluded that the distinction of “deserving” and “undeserving” migrants is an important theme in the media portrayal of Syrian refugees in Germany (Vollmer & Karakayali, 2018). Covering six print media outlets and 4000 articles, a Canadian study found “conflict,” “citizenship,” “family,” “services,” “human rights,” and “religion” as themes that represent lenses for understanding the refugee crisis (Wallace, 2018). Another study explored two opposing perspectives in depicting the Syrian refugee crisis in the US media. The mainstream media outlets embraced the liberal narrative and a humanistic perspective to depict Syrian refugees as victims and hardworking people, while a few mainstream, but mostly alternative media grounded in conservative ideology portrayed Syrian refugees as a threat to the American way of life (Bhatia & Jenks, 2018). In a cross-national design, Abid et al. (2017) compared the use of metaphors in the representation of Syrian immigrants in corpora of the websites of Western, Turkish, and Arab media outlets. They found a number of common water mass metaphors such as “wave,” “stream,” “influx,” “flow,” “pour,” or “absorb” on those websites. Their statistical analysis suggested that media in refugee-hosting societies significantly used more water metaphors to depict the exodus of Syrians (Abid et al., 2017).

Despite the prevalence of social media in conveying anti-immigration opinions, few studies focused on Turkish content on social and digital media outlets. Öztürk and Ayvaz (2018) used a text-mining approach to analyze two million tweets that include the keywords “Syrian” and “refugee” in Turkish and English. Similarly, they were able to analyze a wider variety of opinions since the conventional mainstream media does not always cover opinions opposing their editorial policy. Özdemir and Öner-Özkan (2016) explored the representation of Syrian immigrants in a small sample of Turkish online boards and discovered several themes for positive and negative attitudes towards immigrants. For the negative attitude those themes were: the perception of “troublemakers” regarding immigrants, the economic impact hosting immigrants, and the exploitation of immigration for political gain in favor of the government either in the form of whitewashing the ruling party or enfranchising migrants for wider electoral support. The humanistic representation of immigrants with an emphasis on their innocence and anti-racist remarks were common among posts with a positive attitude towards immigrants (Özdemir & Öner-Özkan, 2016).

Nevertheless, social media studies in Europe can inform research on Turkish public response to Syrian refugees. A European study examined the portrayal of male Syrian refugees on social media platforms such as Twitter and Pinterest. The research revealed the gender aspect of representation where male refugees are either portrayed as rapists and terrorists or effeminate cowards who flee their country instead of fighting for it (Rettberg & Gajjala, 2016). Another study looked at the representation of Syrian refugees by Twitter users in Belgium and Turkey. Offering a qualitative study of tweets with a “small-data” approach, this research highlights the significance of national context and national migration history. Islam has been a prominent theme in both Belgium and Turkey as either a marker of otherness or cultural proximity (Bozdag & Smets, 2017). Social media platforms have also been studied regarding how they influence content on Syrian refugees. A comparison of visual media platforms—Instagram and Pinterest—revealed that Pinterest included more security-concern sentiments while Instagram predominantly included humanitarian-concern posts (Guidry et al., 2018).

Opinion polls are commonly used to track and analyze the trends in the public response to immigrants. A cross-national comparison of attitudes towards Syrian refugees reveals striking findings. Face-to-face interviews of a nationally representative survey on attitudes towards Syrian refugees conducted in 2016 suggest that the Turkish public has the highest intergroup contact with refugees in comparison to other studies conducted in Belgium, France, the Netherlands, and Sweden. Nevertheless, the Turkish public ranks first among those countries in terms of reporting refugee-related economic and quality-of-life concerns. Turkish women, especially, reported higher quality-of-life related threats from immigrants. According to this cross-national comparison, Turks face the most strict conditions—tied with French people in this measure—regarding public acceptance of refugees and refugee settlement (Coninck et al., 2020). A recent opinion poll (Doğan et al., 2021) on immigration reports that the overwhelming majority of the Turkish public in Istanbul (86%) sees the immigrant culture as the reason why Syrian immigrants do not integrate into the Turkish society, while the linguistic barrier follows culture with 82%. Sociocultural reasons highlighted by this thorough research report illustrate another important finding. More than four of five (81%) respondents said that the reason behind the failure of integration lies in the fact that immigrants do not make enough effort to integrate (Doğan et al., 2021).

The social determinants of the attitude towards immigrants and immigration have been the subject of sociological research. A review of public opinion scholarship finds that perception about the consequences of immigration and perception about the size of the immigrant population are among the micro-level attitudinal predictors with the political orientation and identity. The literature of public opinion scholarship demonstrates that the minority presence, economic condition, and political-ideological climate are contextual, structural, and macro-level determinants of public attitude towards immigrants and immigration (Ceobanu & Escandell, 2010).

This research builds upon the existing body of literature on the historical and cultural climate of Turkey as the macro-level determinants of attitude towards immigrants. The textual analysis of tweet corpus sheds light on the micro-level components of immigrant attitude in the Turkish Twittersphere. Hence, we classified a great number of tweets by their attitude towards immigrants. In addition to the training of a classifier model, we investigated major themes within the corpus by using deep learning-based topic modeling techniques.

Recent advancements in computational fields enable researchers to conduct sociologically informed textual analysis. The utilization of text mining (natural language processing) is essential for processing a large corpus of text data that cannot be processed or summarized by an expert. Traditional natural language processing (NLP) methods use a discrete representation of the words in the input text such as the bag of word model. Then, probabilistic models such as naive Bayes for sentiment analysis or LDA for topic modeling are performed on those word counts (Sun et al., 2017). However, words in a text are not independent discrete entities; rather a text as a whole has a meaning. Moreover, traditional NLP methods require preprocessing of the text based on the syntax features of a specific language (Straková et al., 2014), yet software packages besides English either lack those packages or suffer from poor performance. Thus, traditional NLP has limited performance for text mining. There are recent advancements in natural language processing with the introduction of deep learning, which is a machine learning method inspired by the brain and contains multiple layers of artificial neurons, meaning that deep learning can learn a higher level of abstraction of raw input data such as text. Deep learning-based language models lead to huge performance improvement in text mining tasks such as machine translation, question answering, reading comprehension, and summarization (Radford et al., 2019; Vaswani et al., 2017). BERT is a type of deep learning-based language modeling method to represent text numerically (Devlin et al., 2018). BERT takes any raw text in the trained language as input and provides a numerical representation (embedding) of the text as a vector. Those sets of numbers can represent the meaning of the text because BERT models are trained with a big textual dataset to learn a semantic representation of the text. Unlike traditional NLP methods, BERT provides a holistic representation of the text without language-specific preprocessing. Thus, text mining methods, such as text classification and topic modeling trained with BERT embedding, outperform traditional methods, although those methods are not yet diffused to social science research from theoretical computer studies.

Earlier research on social media representation of Syrians in Turkey has been either weak in methodology or was unable to explain research findings through a sociological lens. Attitude towards immigrants in the largest refugee-hosting society in the world remains an under-researched area. The purpose of this study is to understand the public response to Syrian refugees in Turkey as it has been a great political limitation for public policy and immigrant incorporation. This study used novel deep learning-based language models to extract meaningful information from vast chunks of textual data from Twitter on the framing and competing narratives of immigration of the recent Syrian immigration to Turkey.

Methods and Results

We use deep learning-based text mining methods (Bianchi et al., 2020; Schweter, 2020; Yildirim, 2020) to extract insights from the tweets on Syrian immigration in Turkish Twitter. Text mining is a quantitative text analysis method used to analyze textual data that is too large to be read and interpreted qualitatively by humans. We used the leading edge deep learning method to mine Twitter data as presented in Figure 1.OPEN IN VIEWER

Data Collection and Annotation

Data was collected from publicly accessible tweets posted between February 1 and July 15, 2020, creating a nearly six-month period that encompasses major events of the Turkish public response to Syrian immigration. The Turkish offensive into northeastern Syria is hypothesized to be associated with a change in the contents and the volume of social media discussions. The Turkish Armed Forces initiated a cross-border military operation called “Operation Spring Shield” against the Syrian regime forces on February 27, 2020 as a retaliation to a deadly airstrike that killed more than 30 Turkish soldiers on the same day. ¹¹ The operation ended on March 6, 2020 when the Turkish and Russian presidents jointly announced a ceasefire. ¹²

We hypothesized that the unilateral Turkish decision to open the European borders for Syrian immigrants is another key event in this period. Turkey opened the gate to Europe just a day after Operation Spring Shield to exert pressure on the EU to take action and support the Turkish side (Deutsche Welle, 2020). Thousands of migrants gathered at the border of Greece and we hypothesized that Turkish Twittersphere responded strongly to this event. Turkish authorities later announced that the border had been closed due to coronavirus-related risks on March 6, 2020.

The last set of landmark events is related to the global pandemic and Turkish response to combat the coronavirus. The World Health Organization declared a global emergency on the coronavirus outbreak on January 30, 2020. First confirmed COVID-19 related deaths in Turkey occurred on March 17, 2020, while the first-weekend lockdown was enforced on April 11, 2020.

We take these three dates as turning points or landmark events that are relevant to the timeframe of our research as Figure 2 summarizes.OPEN IN VIEWER

The content of this Twitter data comprises 295,775 tweets. Tweets containing the keyword “Suriyeli,” which is the Turkish word for “Syrian,” and all possible suffixes to that word made up the textual corpus of the study (extended method). Total number of COVID-19 cases and deaths in Turkey are obtained from OurWorldinData ¹ (Roser et al., 2020).

We annotated 5250 tweets to classify attitudes towards Syrian immigration with the following annotation criteria:

Themes associated with pro-immigrant attitude

○ “Immigrants make our country (economically, politically or morally) stronger”

Themes associated with anti-immigrant attitude

○ “Immigrants are a burden (economically, politically or morally) on our country”

A small portion of tweets that neither had the pro-immigrant nor the anti-immigrant themes remained unclassified and have been annotated as rest (or neutral). For this classification task, we annotated approximately 250 tweets for each day between March 3, 2020 and March 21, 2020 with the labels of pro-immigration, anti-immigration, and neutral.

Text Embedding and Attitude Analysis

In this study, we employed BERT (Devlin et al., 2018), a recent Natural Language Processing (NLP) method for numerical vector representation of the text. More specifically, we used Turkish-BERT (Schweter, 2020), which is trained on 44 × 10⁹ sentences sourced from the Turkish Wikipedia. Embeddings of each tweet in the dataset is obtained with the Turkish BERT model (extended method). We utilized those text embeddings as the input of text mining methods in this study.

We first investigated attitudes towards immigrants. The expert annotator could only manually annotate a small subset of the tweets. Thus, we trained a classifier model to classify attitudes towards immigrants with the following labels: “pro-immigration,” “anti-immigration,” and lastly “rest.” Classifier takes BERT embeddings of tweets as input and predicts their attitude. We finetuned ¹⁰ the BERT model with those manually annotated tweets. Finetuning a model based on BERT is more powerful than training a model from scratch since BERT already represents the meaning of the text as a vector (embedding). Thus, a finetuned model learns to classify text based on this numerical representation (extended method). 80% (n = 4200) of annotated tweets are used as a training set and 20% of annotated tweets (n = 1050) are used as a test set. Although 5250 tweets are a limited amount of training data, the model achieves the recall of 73%, 73%, 60%, and precision of 67%, 73%, 65% for each class of pro, anti, rest labels of tweets thanks to the finetuning of the BERT model (see Supplementary Figure 2; extended method). Next, we annotated all 295,775 tweets based on the predictions of the classification model. Attitude classification models have a lower performance on rest labels and this label is not relevant for the analysis. Thus, we excluded tweets predicted with the “rest” label from the analysis in this study. Excluding the rest label increases the precision of the model to 82% for pro-immigrant and 92% for anti-immigrant labels.

Anti-immigrant attitudes are significantly more prevalent in the Turkish Twittersphere during the scope of this research (Wilcoxon signed-rank test, p = 0, supplementary method). Attitudes of 50% (n = 148,682) and 20% (n = 60,371) of tweets are predicted as anti-immigrant and pro-immigrant, respectively. The classification of the immigrant attitude of tweets revealed the deviations of immigrant attitude over time and by landmark events (see Figure 3 and Supplementary Figure 1).OPEN IN VIEWER

We further investigated aberrations in the attitude trends by employing attitude trend outlier analysis (extended method; see Supplementary Figure 2, Supplementary Table-1). Each day has a discrete number of positive and negative tweets; thus, binomial distribution can be employed to calculate likelihood of each day against overall behavior. We further parameterized binomial distribution with conjugate beta prior and used beta-binomial distribution to avoid outliers caused by random fluctuations (dispersion). Firstly, most of the tweets were posted during the military campaign of the Turkish army (Operation Spring Shield) between February 26 and March 6, 2020. On the first day of the operation, the attitudes in tweets are slightly negative (no significance, p = 0.130) but they get more positive with the progress of the operation (no significance, p = 0.150). In addition, significant (p < 0.030) anti-immigrant attitude outliers observed between March 27 and March 31 parallels the exponential increase of COVID-19 pandemic cases. Lastly, the pro-immigrant attitude outlier observed on April 28, 2020, the day when a Turkish police officer shot and killed a Syrian immigrant (p < 0.0001). ⁷

Sentiment analysis is another commonly used method to obtain opinion polarity on a subject (Mäntylä et al., 2018). In this study, we utilized a deep learning based sentiment analysis model (Yildirim, 2020) to obtain sentiment of the tweets from BERT embeddings. Sentiment analysis of the classified tweets shows that both pro-immigrant and anti-immigrants tweets predominantly have negative sentiments. After anecdotal investigation by reading tweets with the most negative sentiment, we suggest two reasons for this observation. Although Twitter users of both of the attitudes advocate for very different policies, they express their dislike of the status quo. Moreover, Twitter users in both camps harshly criticize the opposite camp. Overall, tweets with an anti-immigrant attitude have greater negative percentages with a p-value of ∼0 with the Fisher-exact test, as Figure 4 shows. These results show that sentiment alone may not predict the attitude toward immigrants. The fact that most of the tweets have negative sentiments regardless of the attitude indicates that sentiment is not a proxy of attitude in this corpus of tweets, hence why we went beyond a sentiment analysis, unlike many social media analyses in the current literature. We instead trained the attitude classifier model in addition to the sentiment analysis. Methodologically, this enabled us to distinguish the content of the sentimentally similar tweets with very different attitudes in the corpus.OPEN IN VIEWER

Topic Modeling

We explored major topics in the corpus that influence the public attitude toward immigrants. We also investigated how those topics change over time while identifying landmark events. For this purpose, we trained another model, the Contextualized Topic Models (CTM) (Bianchi et al., 2020) based on BERT embeddings of tweets. CTM is a type of deep variational autoencoder model which works in two steps: encoder and decoder. The encoder step takes embeddings of the tweets as input and then learns the distribution of the embedding (meaning) space in a lower dimension. Then, the decoder step samples the vectors from the lower-dimensional space using parameters learned in the encoding step, and attempts to predict original tweets from the lower-dimensional sampled vectors (Hou et al., 2017). The purpose of the variational autoencoder is to learn non-linear latent distribution (common characteristics) in the high dimensional data. The variational autoencoder is useful for topic modeling since latent distributions in the embedding space of text data are the topics. Using this type of a deep variational autoencoder, CTM, we grouped all tweets into five topics. Our CTM also provided the top keywords related to each of these topics.

The CTM of the large corpus of tweets revealed five major topics with regard to Syrian immigrants in Turkey. Topics are listed with keywords in Table 1 and we named each topic according to these keywords by reading the most representative tweets predicted by the model. Among those five topics “Patriotism,” “Syrian integration,” and “Social benefit demands” are three key topics for this analysis and 79% of the tweets map to those three topics (Figure 5).

OPEN IN VIEWER

Table 1.

(Summary): Topic Keyword Table.

Topic	Keywords	Keywords translated in English
Patriotism	askerimiz, irkcilik, değildir, vatanini, savasa, isimiz, kurban, kani, can, turkler, sehitler, serefsiz, arap, kacmis, korumak	Our soldier, racism, not, homeland, war, our duty, victim, blood, life, turks, martyrs, dishonest, arab, fled, protect
Syrian integration	bm, entegre, proje, gorev, cocuklarimiza, koruma, egitimi, vekilim, gaziantep, engelli, turkce, uygulama, ogretmen, ogrenci	The UN, integrated, project, task, for our children, protection, education, member of parliament, gaziantep, disabled, turkish, practice, teacher, student
News	ayagini, ot, babalarinin, cevirdiler, anlasiliyor, saniye, uz, yolluyor, soruya	Their feet, grass, their fathers, they turned, understood, second, long, sending, to the question
Misc	kilini, birakacak, secti, bugunun, ikiye, vermemis, basiyor, ucaklarla	Their hair, left, chose, today, to two, he has not given it, he is printing, with airplanes
Social benefit demands	lira, kizilay, vatandasin, parasini, eyt, emekli, halka, vatandasa, parayi, kira, harcandi, kampanyasi, maasi, deprem, vergiler	Lira, Kizilay (red crescent), citizen, money, EYT (retired pension related), retired, public, citizen, money, rent, spent, public campaign, salary, taxes

OPEN IN VIEWER

Figure 5.

Topics by the immigrant attitude.

The first topic includes patriotic, nationalist as well as anti-Arab themes. The topic has a few Turkish nationalist keywords such as “our soldier,” “homeland,” “our duty,” and “to protect.” Yet, some keywords in the topic have a condescending tone targeting Syrian immigrants. Among those keywords, the fled,” depicts Syrian immigrants as cowards leaving their country without fighting, and “dishonest,” an example of offensive language targeting immigrants and refugees.

The second key topic, “Syrian integration,” includes demands for Syrian integration as well as social benefits. Education-related themes are evident in this topic. “Teacher,” “student,” “education,” and “for our children” keywords indicate the demand for Syrian integration through education in Turkey. Because we only analyzed tweets in the Turkish language, it is unlikely that tweets that fall into this topic are exclusively written by Syrian immigrants and refugees. Rather, it is very possible that the topic lies at the intersection of Twitter activism of Turkish people and NGOs due to other related keywords in the topic such as the “disabled” and “project.”

The third and last key topic of the corpus is the social benefit demands of the Turkish people. This theme includes pensions, insurance, money, and tax-related keywords. “Lira” the official currency of Turkey, “money,” “rent,” and “salary” highlight the monetary content of the benefit demands. The topic also includes keywords about more extensive welfare benefits such as retiree pensions as well as taxes. The prevalence of money-related keywords indicates a strong link between discussions on immigrants and eligibility to the welfare benefits. This topic can generally be understood as natives’ demand for priority in welfare benefits in the context of immigration debates, which is conceptually defined as welfare chauvinism (Betz, 2019).

The relationship between analysis of aforementioned topics and attitude towards immigrants is summarized in Figure 5. The largest topic (33%) is “social benefit demands” (Topic 4) and this analysis suggests that most of the tweets share anti-immigrant content whereas Syrian integration (Topic 1) have the largest pro-immigrant content. Surprisingly, patriotic content (Topic 0) is roughly equally divided between pro-immigrant and anti-immigrant attitudes. The other two topics, news, and disorganized miscellaneous keywords have overwhelmingly anti-immigrant content yet those topics contain a minority of tweets (21%). Although context topic modeling and attitude analysis are independently trained, findings from two independent models valid with each other given that the “social benefit demands” topic has a dominant anti-immigrant attitude while the “Syrian integration” topic has a pro-immigrant attitude.

The landmark events influence topic trends. Figure 6 demonstrates the prevalence of five identified topics and their change over time in tweet corpus. The relative share of “Patriotism (Topic 0),” “Syrian integration (Topic 1),” and “Social benefit demands (Topic 4)” topics fluctuated across the scope of this study. The juxtaposition of Figure 6 and the landmark events summarized in Figure 2 links the macro environment of social media users and social media trends. The timing of the Turkish cross-border military operation in Syria and the opening of the European border for immigrants following the military operation correspond to a relative increase in the share of tweets with patriotic themes. The high volume of tweets (Figure 2) during the campaign of the Turkish army (Operation Spring Shield) between February 26 and March 06, 2020 shows how responsive the Turkish Twittersphere is to patriotic and nationalists themes. Patriotism (Topic 0) lost prominence by the end of Turkish cross-border military operation and social benefit demands were echoed more strongly thereafter in the Turkish Twittersphere. Social benefit demands of Turkish citizens (Topic 4) peaked after the announcement of the first COVID-19 related death in Turkey. In the meantime, Syrian demands for welfare benefits (Topic 1) gained prominence across the timeline. The other two topics—news (Topic 2) and miscellaneous keywords (Topic 3)—have a relatively more stable trend.OPEN IN VIEWER

Figure 6.

Proportional trends in topics of the analyzed tweets.

Discussion

This study exposes immigrant attitudes in social media content in the largest refugee-hosting society. Our computational analysis of textual data suggests two main findings in the larger migration context. The COVID-19 pandemic was influential in shaping social media discussions on who is perceived to “deserve” to be a priority in receiving welfare benefits. In Turkey, a sharp increase in the public discussion on “social benefit demands” coincided with the first COVID-19 related death. Specifically, a high volume of Twitter activity (i.e., tweets) with the common theme of welfare chauvinism directed against immigrants. Meanwhile, a surge appeared in the number of tweets on the topic of immigrant demands for welfare benefits as well as their integration into Turkish society. Tweets on this topic included keywords such as “education” and “our children” (keywords translated from Turkish to English), which indicate a demand for more extensive immigrant integration into Turkish society by second generation immigrants. The ongoing pandemic in Turkey during 2020 corresponded with increased discussions on welfare benefit demands on social media.

Patriotic themes fueled by the interventionist foreign policy may create a sentiment of “patronizing tolerance” among host-society members, who would condone the presence of immigrants as long as it reinforces their claims of moral superiority. This finding is supported by a temporal correlation in our data. The share of the theme “patriotism” in the sample corpus increased during the Turkish cross-border military offensive against the Syrian regime. The topic “patriotism” had a few nationalist keywords such as “our soldier,” “homeland,” “our duty,” and “to protect,” while other keywords such as “fled” and “dishonest” were targeting Syrian immigrants and refugees. This topic has a large share of pro-immigrant tweets that are essentially nationalist but sympathetic to immigrants through coding Turkish soldiers as the saviors.

The first finding of our analysis suggests that the concurrence of the pandemic and nativism creates a hostile environment for immigrant communities. Earlier analyses classified the main facets of the recent nativist populism in Europe and North America in three categories: economic nativism, welfare chauvinism, and symbolic nativism (Betz, 2019). Immigration attitudes in social media fall largely under the category of “welfare chauvinism,” according to which Turkish citizens deserve a top priority in social benefits offered by the government. The COVID-19 pandemic has fired up welfare chauvinism in the Turkish Twittersphere at the expense of immigrant exclusion. The public statements of policy makers about generous public spending on immigrants evoked negative attitudes among social media users. “Economic nativism” is not as strong as in Europe or North America since taking the jobs of native workers is irrelevant in the Turkish society where refugees and immigrants face unfair labor market conditions, hiring quotas, and exclusionary policies (International Crisis Group, 2016, p. 7–8). “Symbolic nativism,” which is the central facet of European right-wing populist mobilization, is barely present in the data collected in our analysis (Betz, 2019). Cultural markers are not strongly accentuated on the contrary of the previous studies in immigrant-host society relationships (Bhatia & Jenks, 2018; Bozdag & Smets, 2017). Despite the evident anti-immigrant attitude, immigrants are seldom framed as a threat to the “Turkish way of life.” Turkish Twitter users did not extensively dwell on the common security themes identified by earlier research in the immigrant-hosting European societies such as terror, violence, and rape. The cowardice of the immigrants, however, was a frequently recurring concept in tweets with patriotic content. This theme has also been identified by previous literature (Onay-Coker, 2019; Rettberg & Gajjala, 2016). Hence, our results are consistent with prior research in this regard.

The second finding can best be understood within the context of Turkish society. Nativist sentiments characterized by patronizing tolerance are of a very similar nature to social relationships in ethnically or racially stratified societies. The supremacist ethos offers partial inclusion to migrants under rigid conditions. Before everything, docility is expected from immigrants to such an extent that immigrants must be silent around political matters that directly affect their lives and futures. In addition to docility, immigrants must openly express and iterate their thankfulness. The public sentiment on Syrian immigrants in social media has gradually moved to a more positive outlook, as a juxtaposition of Figure 8 and the landmark events summarized in Figure 2 suggests, right at the week when two major developments happened: Turkish forces launched a cross-border operation in the Syrian soil and following the operation Turkish authorities opened the European border for Syrian immigrants. This gives us enough reason to think that refugees and immigrants are seen as tools to embarrass the rhetorical enemy, the Western civilization or Europe, in the Turkish version of civilizationism. Refugees and immigrants are useful in this regard, yet the ultimate decision to stay or leave does not belong to them. In this sense, patronizing tolerance is a strong theme regarding the context of immigrant-host society relationships.OPEN IN VIEWER

Our observation on the alleviative effect of the military camaraderie on ethnic and racial discrimination concurs well with Stouffer et al.’s (1949) seminal social psychological work on the American soldier, although in our case military intervention provides a symbolic sense of camaraderie manufactured by nationalist ideology. Served alongside yet segregated from their fellow Blacks soldiers in the military, fewer White soldiers expressed hostility towards Black people in the US (Stouffer et al., 1949, p. 592). However, many White soldiers were still holding a belief of racial superiority that exhibited itself in the reasoning that Black soldiers deserve more strenuous work since their lives were perceived as less valuable when compared to the lives of Whites (Stouffer et al., 1949, p. 590). Turkish public attitude towards Syrian immigrants as well as the armed Syrian groups fighting alongside the Turkish army shows similar characteristics in our research. Militaristic nationalism offers Syrian immigrants a place within the imagined social order. In that designated place, Syrians can be easily disposed of when they do not fulfill their duty as people of inferior and lower rank.

Conclusion

This novel social media study reveals a possible link between a rising public sentiment of patronizing tolerance towards immigrants and the implementation of significant foreign policies in Turkey. Further, this article presents possible future developments of populist movements in a post-pandemic global political setting. Using the case of Turkey as a current example, rising public social expenditure in response to COVID-19 may lead to stronger welfare chauvinism in immigrant-hosting societies. Future research will need to explore a more systematic and theoretical analysis to make a valid and reliable connection between far-right populist movements and immigrant attitudes in the political—and socioeconomic—aftermath of COVID-19. Additionally, similar movements in other countries besides Turkey would need to be investigated to continue the work started in this initial paper.

Supplemental Material