Immigration and integration

The theoretical framework surrounding immigration and the subsequent process of integration has been established: Social identity theory [40] describes the mechanisms in which people identify with a cohort or group, including the perception of other groups (i.e., Ausländer). The role of political alignment in social identity formation can be observed in Germany´s internal struggle between nationalism and “welcome culture”, in which people identify themselves socially according to political belief [41].

Intergroup threat theory [42] also relates to Germany´s evolving social landscape and the hostile reception of foreigners by natives, particularly over the perceived competition for resources and jobs [43]. Literature supporting threat theory finds that right-wing extremism and prejudice arise by feeling threatened by the “outgroup” (minority group) [44–47], which can be an unfortunate result of “acculturative stress” on the native population [48, p. 378] due to increased foreigner migration into the country. Thus, the connection between political and social identity is made stronger.

Despite the social clashes which arise from mass migration (the perception of threat, the likening to cliques and cohorts), social integration is the result of social participation. Shadid [49, p. 362] further defines social integration as “the participation of ethnic and religious minorities, individually and as groups, in the social structure of the host society [... ]”; whereas Mogahed and Nyiri [50] de-emphasized the cultural conformity aspect of integration and focused on the shared goals and commitment towards the host society. As such, integration itself is a process which requires both consent from the host society, and the willingness of the migrant to integrate. Budyta-Budzyńska [51] describes the spectrum as beginning with segregation (remaining separate from the host population), which leads into adaptation (the gradual and personal change one experiences over time through the acculturation process), which leads to integration (the blending into society more so than before by robust relationships but keeping a distinct national identity), and transforms into the final assimilation, in which the individual or group is similar with the host (in behavior, customs, beliefs, for example). As such, immigration and integration exist as objective events, which are connected to nuanced social theories. The crux is to determine how discourse surrounding these themes change over time, as mass media has been regarded and investigated as a reflection of public discussion [52].

Agenda setting, salience, and framing

The general assertion is that what becomes a hot topic in the media, becomes a topic in society [53]. Agenda setting includes the media actor’s decision over which stories to cover, how they are framed, including what their intentions are by covering the news through this framing. This access of topics raised by media in the individuals’ minds becomes the state for “issue salience”; the extent to which people cognitively engage with a topic [54].

Goffman [55] had initially theorized that the organization of discourse will affect an individual´s thoughts and actions. As such, framing involves the organization of information. Labeled frames enable people to make necessary interpretations. Minsky [56] defined a frame as a template or data structure that organizes various pieces of information. Gamson and Modigliani [57, p. 143] had explained that frames are the “central organizing idea or storyline that provides meaning”. Over the decades, the intentional negative framing of media, termed a “strategy frame” [58] has gained attention in media discourse research. It has been theorized that negative media bias emerged through press freedom: to criticize politicians and political parties, and political affairs at large [59]. These theories have been applied in literature to enhance media analysis.

Mass media data collection

We use one GGSS item (originally [MA01]), which we label as “Adaption” as a reference point (Fig 1). This guided our decisions according to which years of media to collect, whether the media data should be national or regional, and serves as further justification as to why we make the term “Ausländer” central to our text analysis. Moreover, the GGSS is administered during key “time frames” of each year, which we adapt as well. Therefore, the media corpora for each year contain articles which correspond specifically to the GGSS survey administration periods. We selected the years 2006, 2010, 2012, 2016, and 2021 and collected our media according to the specific annual windows in which the survey was administered. The combined corpora amount to 10,063 articles. S2 Fig provides a full table describing corpus size per year, media outlets, and time frame for corpus collection.

We used FID Media (formerly called adlr.link) as a resource for our media corpora. It is an archival database provided by the University of Leipzig. FID Media is an academic information service for communication and media studies [68]. Unfortunately, FID Media does not allow for the sharing or republishing of articles, therefore the articles from the corpora can be only accessed directly from the archival database itself; the analysis method itself complied with the terms and conditions of FID Media.

For the term selection strategy, we focused exclusively on the term “Ausländer”. The advantage of this term is that it is rich in etymological history and controversy; despite the stance that it is politically “incorrect”, there are several Ausländer-compound terms available in the German language, which are widely used in media discourse, and serve as the basis for our text analysis. This term is also used in many GGSS items; we would like to maintain the connection between public opinion data and media discourse regarding foreigners by keeping terms consistent; this is especially important as our corpora of published media articles coincides with the survey collection phases. Therefore, our media discourse search was restricted to the keyword “Ausländer” (foreigner) and all articles in the corpus containing this term were selected for the predefined time periods.

Text cleaning

Once the corpora have been compiled, the next step is to clean the text. Cleaning the text typically refers to removing any unnecessary words or symbols from the body of text. Specifically, this involves removing stop words, numbers, punctuation, as well as converting all words to lowercase letters [69].

The cleaning of our corpus was conducted using nltk (Natural Language Toolkit), which is a Python package commonly used for removing stop words. A list of 155 stop words were effectively removed from each corpus. Once each corpus was collected, each corpus was then cleaned individually. To identify relevant topics, python libraries re, defaultdict, Counter were utilized to return the top 100 most frequently used terms in the corpus. However, the terms “Ausländer” and the conjugation “Ausländern” were excluded as they are not descriptive. In German, compound words are incredibly common, therefore many terms used to describe situations or people are often confounded into one simple term (for example, “Ausländerproblem”, which means foreigner problem, or “Ausländerintegration”, meaning foreigner integration).

NLTK and WordNetLemmatizer were used for the purposes of lemmatizing the terms (combining singular and plural tenses and words of the same meaning). As a result, the frequencies of one term are added to the frequency of another terms (if they are the same in meaning and/or the singular/plural or some conjugated form of the same word, such as Ausländerbehörde/n (a government agency for foreigners) or Ausländerfeind and Ausländerfeindlichkeit (xenophobe and xenophobia). Once the terms for each category (and all their consolidated forms) have been identified and consolidated into their groups, the categories can finally be compared for each year.

Categorizing media with Latent Semantic Analysis (LSA)

We used Latent Semantic Analysis (LSA) to evaluate the term “Ausländer” in mass media discourse. LSA was particularly suitable, because its categorization is context dependent; semantic “meaning” emerges through the relationships between terms in a context-based subspace [70]. LSA relies mainly on the co-occurrence of terms in order to categorize them [71,72] and operates on the premise that similar words are used in similar contexts. LSA exploits the frequency count to derive a meaning that is purely based on the text (the language itself) [73]. In regard to the lists of consolidated terms, this means that terms which are frequently co-occurring or exist within the same contexts as one another are most likely to be sorted within one category. As terms appear across different categories (as they are discussed in various contexts), overlap in terms across categories is expected.

LSA algorithms commonly perform with td-idf (term frequency—inverse document frequency), which helps LSA to determine the significance of words based on how often (or how infrequently) they appear in the text [74]. It is often used as a tool to reduce noise within a document and can be particularly useful when researchers are attempting to investigate the contents of the document without prior knowledge of its topics. In the current case, although td-idf is an inherent part of the algorithm, its purposes are mediated by the deliberate use of pre-determined terms of interest (the Ausländer compound terms). Furthermore, additional text processing steps –such as the lemmatization of the compound terms – further assisted in the cleaning of text and reduction of noise. In the lemmatization process, terms of the same meaning (for example, ausländerfeind, ausländerfeindlich, and ausländerfeindlichkeit) are combined, along with their frequencies. Therefore, although td-idf is elemental in reducing noise, reducing the analysis to specific terms also effectively reduces noise.

One caveat to LSA is that the number of categories must be predetermined. Therefore, we measured coherence for each corpus to determine what number of categories would be best. For each year (except 2016), three categories showed the highest coherence (with relatively high values ranging from 0.74-0.78). For 2016, four were considered ideal with a score of 0.85. For years 2006, 2010, 2012, and 2021, four categories still achieved satisfactory coherence values, despite that three topics were most recommended. Coherence measurement charts can be found under S3 Fig. Ultimately, we can achieve more nuanced categories with a manageable amount of overlap with four categories. It is important to keep some overlap [75], as it is still a fair indicator of the different contexts in which these terms are used.

Although LSA organizes these terms into respective categories, it does not label the categories themselves. To appropriately label the categories, two coders (Coders 1 and 2) labeled the categorized term cluster for each year. Therefore, the labels should reflect the prevalent terminology within each category (i.e. the category labeled “Xenophobia” has a disproportionately high amount of discourse using the term “Ausländerfeind” whereas the discourse category “Social Integration” more so contains terms such as “Ausländerintegration” or “Ausländerrecht”). The labels therefore concern themselves with the general category theme and is not concerned with the sentiment of the discourse.

The coders received an Excel chart containing four clusters of terms – there was one Excel chart per year, to ensure that each year would have four respectively labeled categories, and to prevent the intermixing of discourse terms from the corpora. The prompt was: “Based on the terms within each group, label the categories as either Administration& Policy Social Integration, Limiting Migration, and Xenophobia. Each group can only have one label, and each category should be represented.”

This framework was intended to reflect a spectrum of discourse, ranging from the bureaucratic (potentially objective), to discourse discussing alternative stances (from pro- to anti-foreigner). The intercoder agreement labels had a Cohen´s K agreement score of 0.64, which is considered substantial [76]. There was a total of twenty clusters (n=20), which are four categories for five years. There was consistent agreement on 14/20 labels. The only disagreement was found between Social Integration and Administration, which is likely due to the nuanced discourse surrounding rights, which can concern bureaucracy and integration alike; moreover, there is a distinction between discourse terms referencing government agencies (such as Ausländerbehörde) and the applicants. In the six instances of disagreement, a third coder (Coder 3) was used to decide and provide rationale.

Sentiment analysis

Sentiment analysis is a field at the intersection of linguistics and computer science that attempts to determine the sentiment contained in text [77]. It is not purely psychological or emotional in nature, but rather a nuanced approach to how we can classify linguistic expressions according to the type of opinion that they convey [78]. Human beings are inherently emotional beings, which often influences our linguistic expression and interpretation of text, however sentiment analyses are also occasionally a matter of subjective or objective interpretations [77,78]. Therefore, we naturally interpret texts based on fact, our innate emotional reasoning, and opinion. Sentiment analysis, as being intertwined with computer science, has evolved to include non-human analytics. For example, Chat-GPT has been extensively utilized in sentiment analysis for media and social media text [79, e.g.,], including on multilingual texts [80]. Moreover, the comparison between Chat-GPT and human annotation has been extensively explored with the consensus that Chat-GPT provides competitive accuracy [81] across a wide range of text topics, including media text [82]. We adapt this method for our sentiment analysis.

At this point, the number of categories has been established and the categories have been labeled but the discourse within each category still require a sentiment analysis. This step helps us to decipher whether the discourse in a given category (i.e. Social Integration) contains positive, negative, or neutral discourse. The sentiment analysis provides valuable insight into each category, as (for example) it may be the case the category “Xenophobia” contains discourse against Xenophobia, but it nonetheless categorized with “Xenophobia” discourse for the terminology provided and the fundamental topic being discussed. Therefore, it could be that an increase in categorical prevalence for Xenophobia does not necessarily mean growing disdain for migrants in media. For this reason, a sentiment analysis is in order. Given the bank of prior literature using Chat-GPT for inter-reliability, we compare the Chat-GPT annotations to human (German native speaker). Both Chat-GPT (Coder 1) and the human coder (Coder 2) were provided sentences (n=296); markers such as year and category labels were removed to prevent bias. Coders were given the same prompt: “Based on the “Ausländer” compound terms, please label each sentence as either positive, negative, or neutral in sentiment.”A category which achieves the majority consensus of any sentiment is deemed to be positive, negative, or neutral overall (i.e., 50% positive, 25% negative, 25% neutral). We calculated the agreement of human annotation and Chat-GPT with Cohen´s Kappa, which is a robust measurement of agreement that accounts for degrees of randomness in the sample and follows this scale: ≤ 0 indicates no agreement, 0.01–0.20 as none to slight, 0.21–0.40 as fair, 0.41– 0.60 as moderate, 0.61–0.80 as substantial, and 0.81–1.00 as almost perfect agreement [83]. Confidence intervals are also reported with the results for each sentiment analysis per category.

Categories extracted

Addressing RQ1: Given the competing “pro- and anti-immigrant” themes in media discourse, we had hypothesized a minimum of two categories. When using the term “foreigner” (Ausländer) as a reference, we found the following four categories that described the media discourse when using the term Ausländer as a reference:

Administration and Policy emphasizes government agencies or contains bureaucratic terms. This category was labeled for the following terms having highest weights: Ausländerbehörde and Ausländerbeirat, which are organizations that typically assist foreigners in a wide spectrum of tasks (applying for residency permits, visas, asylum, etc.). Therefore, this context represents the legalities surrounding migration. One example from this, direct from the corpus (translated from German) is: “Now, a politically controversial question that has long been a hot topic of debate is being passionately debated again: Are we deporting the wrong people? And is what our immigration authorities (Ausländerbehörden) are doing right – is it humane?” (Die Zeit, 2021). Another such context of administrative terminology includes: “In response to a BamS (Agency for Migration Services) inquiry, the responsible immigration office (Ausländerbehörde) explained that [a] “family reunification visa” for permanent residence in Germany [... ]” (Frankfurter Allgemeine, 2021).

Limiting Migration includes anti-immigrant discourse combined with policy discourse. It was so named for the weighted emphasis on Ausländerbeschränkung (to limit foreigners), Ausländerfeind (xenophobia), or “Ausländerkriminalität” (to the effect of foreigner criminality). One example direct from the corpus (translated from German) is: “Plans to restrict foreign players (Ausländerbeschränkung) and increase playing time for German talents in the league should be implemented quickly. Now seems like the right time” (Die Welt, 2010). Another example being: “Foreigner criminality (Ausländerkriminalität) and Islam is not a winning theme in the Union” (Bild, 2021). The intermingling of anti-immigrant discourse with legal contexts causes a lot of nuances, which on the one hand enhances our analysis by uncovering variety among a dichotomy of polarized discourse. On the other hand, it can complicate the analysis (for instance, in the case of coherence values straddling between three and four topics; moreover, the nuance can also be construed as overlap).

Xenophobic discourse has some overlap with “Limiting Migration” however contained less content surrounding policy and has more emphasis on the anti-immigrant sentiment alone, including terms such as: Ausländerhass (hatred of foreigners), Ausländerproblem (the foreigner problem), Ausländerfeindlichkeit (xenophobia), etc. Therefore, discourse which use these terms categorically discuss themes of xenophobia, which includes discourse which may include both discourse which discusses anti-immigrant sentiment, and also discourse which is critical of xenophobia, but is nonetheless discussing xenophobia as a topic of discourse. One example is as follows: “Following a xenophobic attack on a Turkish woman in Giesing, police are seeking witnesses whose information could help in the search for the perpetrator” (Süddeutsche Zeitung, 2006). Another example includes: “So far, the only party I can really rule out is the AfD. They’re a xenophobic party. What they’re doing isn’t right.” (Die Welt, 2021).

Social Integration contained discourse including terms Ausländerintergration (integration of foreigners), Ausländerhilfe (helping foreigners), Ausländerrecht (law to regulate foreigner rights), and Ausländerwahlrecht (the right to vote for foreigners). It should be noted that Social Integration should be construed objectively; despite that the discourse pertains to the inclusion of foreigners into society, it cannot and should not be assumed that the discourse itself is “pro-immigrant” per se, but in any case, a discussion about integration. This discourse can be categorized as integration issue in accordance with Shadids’ [49] participation focus. One translated example is: ““We will fight for the right to vote,” said Raed Saleh. The group is open not only to SPD members, but also to fellow activists from other political camps and from associations of foreign communities. Mohamad Beidoun from the German-Arab umbrella organization supports the initiative” (Die Welt, 2010). Another example of context in which terms related to social integration is discussed includes this excerpt: “It’s about two splinter parties that only entered the Römer because there is no five percent threshold in Frankfurt: on the one hand, Yildiz’s BIG party, which advocates for foreigners’ voting rights and for people with a migration background to retain their cultural identity [... ]” (Die Zeit, 2021).

Change in discourse

Given the previous overview of what each category represents, we now address RQ2 with Fig 2, which represents the side-by-side categorical comparison for all four categories for 2006, 2010, 2012, 2016, and 2021. This ultimately shows the average prevalence of each category per year, based on the categorized terms from each unique corpus. These results effectively compare the discourse of each category relative to the other categories. ANOVA and Tukey HSD analyses were used to determine whether the discourse changes significantly for each individual category by conducting pairwise comparison of the years (each category was the dependent variable and the year was independent variable). Results are described for each category below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Categorical comparison over time.

https://doi.org/10.1371/journal.pone.0340164.g002

Limiting migration

The ANOVA results in Table 1 indicate that there are significant differences in attitudes toward Limiting Migration across the years examined (F(4, 70) = 3.72, p = 0.008), suggesting that public opinion or policy emphasis has shifted over time. As shown in Table 2, post-hoc analysis using Tukey’s HSD test reveals that 2012, 2016, and 2021 all differ significantly from 2006, with each showing a substantial change in limiting migration discourse (p < 0.05). However, no significant differences were found between the more recent years (2012–2021), indicating that the changes largely occurred between 2006 and 2012, and then stabilized afterward. Compellingly, the discourse on Limiting Migration was most prevalent during the more “conservative” years of opinion (such as 2006 and 2016), according to our GGSS item on adaptation attitudes Fig 1. Although a causal link between this discourse category and change of opinion cannot be directly made, the results show a correlation between a rise in this discourse and a more conservative opinion at corresponding points of time.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. ANOVA for limiting migration across years.

https://doi.org/10.1371/journal.pone.0340164.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Tukey HSD Post-Hoc comparisons for limiting migration across years.

https://doi.org/10.1371/journal.pone.0340164.t002

Considering Germany had experienced its largest mass-migration in 2015, it perhaps makes sense that “limiting migration” discourse is most prevalent in 2016 (compared to 2006 and 2021). The relationship between Xenophobic discourse and Limiting Migration is important to consider: Whereas 2006 held the most Xenophobic discourse within its own category, it was 2016 which had the most discourse on “limiting migration”, however the difference in discourse in the surrounding years does not appear to be statistically significant. Historically, the lowest number of immigrants since reunification was recorded in 2005 — the number fell further in 2006 [84].

Administration & policy

As shown in Table 3, ANOVA indicated a statistically significant difference in Administration & Policy scores across years, F(4, 45) = 3.44, p = .016. Post-hoc comparisons using Tukey’s HSD test, as shown in Table 4, revealed that 2006 had significantly higher scores than 2012, 2016, and 2021, suggesting a decrease in policy administration-related measures over time. No other comparisons between years reached statistical significance. This is consistent with Fig 2, which shows Administrative discourse is more prevalent in 2006 than in other years. One potential explanation for the lack of prevalence in 2016 relative to other years is that the discourse surrounding the “refugee crisis” of 2016 placed more emphasis on the critical social reception rather than the administrative aspects of this. In 2006, Administration & Policy constitutes 39.4% of the discourse using “Ausländer”, almost as much as Xenophobic discourse. For 2021, most of the discourse was Administrative in nature (42.7%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. ANOVA for administration & policy across years.

https://doi.org/10.1371/journal.pone.0340164.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Tukey HSD Post-Hoc comparisons for administration & policy across years.

https://doi.org/10.1371/journal.pone.0340164.t004

Social integration

As shown in Table 5, ANOVA for Social Integration revealed a statistically significant difference in discourse across years, F(4, 55) = 6.61, p < .001. Post-hoc comparisons using Tukey’s HSD test showed that 2021 significantly differed from all other years (2006, 2010, 2012, 2016), with higher mean levels of social integration, as shown in Table 6. No significant differences were found among the years prior to 2021, suggesting a marked increase in social integration specifically in that year. This is consistent with Fig 2, which depicts 2021 being something of a “hallmark year” for social integration discourse. Considering what is known about integration discourse and “Willkommenskultur” in Germany in during the 2015 refugee crisis and the time immediately thereafter, it is perhaps surprising that the years prior to 2021 were not statistically significant. This could be due to a lack of GGSS data availability (as there was no surveyed year between 2016 and 2021), and/or due to the complicated and jaded history Germany has with integration discourse.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. ANOVA for social integration across years.

https://doi.org/10.1371/journal.pone.0340164.t005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Tukey HSD Post-Hoc comparisons for social integration across years.

https://doi.org/10.1371/journal.pone.0340164.t006

Xenophobic discourse

As shown in Table 7, ANOVA showed a statistically significant effect of year on xenophobia scores, F(4, 100) = 4.36, p = .003, indicating that levels of xenophobic discourse varied across the sampled years. Post-hoc Tukey HSD comparisons revealed, as shown in Table 8 that 2016 differed significantly from 2010, 2012, and 2021. Specifically, xenophobic discourse was significantly higher in 2016 than in both 2010 and 2012, and significantly lower in 2021 compared to 2016. No other year pairs showed significant differences.This fits to Fig 2, in which Xenophobia was most prevalent in 2016 relative to other categories. Compellingly, discourse on Xenophobia was most prevalent during the most “conservative” years of opinion on our GGSS item (also 2016), which draws another interesting parallel between media discourse and opinion trends from corresponding points of time.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. ANOVA for Xenophobia across years.

https://doi.org/10.1371/journal.pone.0340164.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Tukey HSD Post-Hoc comparisons for Xenophobia across years.

https://doi.org/10.1371/journal.pone.0340164.t008

Sentiment analysis results

In regard to RQ3, the sentiment analysis revealed that none of the four categories were positive. Although “Social Integration” may imply a positive shift in societal acceptance, the sentiment analysis results reveal that the discourse is neutral at best – more neutral than any other category. This fits our hypothesis, in which we assumed based on the historical pessimism surrounding integration discourse, that this category would also not be positive in sentiment. In fact, there was a practically equal amount of neutral and negative sentiment. A set of 59 consensus-labeled sentences (the sentences which yielded total agreement between coders) indicated that 40.7% of sentences were coded as neutral (95% CI = 29.1% – 53.8%), 37.3% as negative (95% CI = 25.7% – 50.5%), and 22.0% as positive (95% CI = 12.9% – 35.7%).

The negative sentences primarily highlighted challenges or shortcomings in integration policies, whereas positive sentences emphasized efforts to improve integration, and promote inclusion, particularly for families and children. Neutral sentences often provided descriptive or legalistic information without evaluative judgment. The confidence intervals reflect the uncertainty in these estimates due to the modest sample size of sentences analyzed: the challenge here is twofold, as social integration discourse (as shown in Fig 2), was not so prevalent prior to 2021, therefore sample sentences are unevenly distributed across the years. Moreover, (as previously stated), we collected the corpora according to GGSS´s data collection periods: year 2021 had the smallest collection frame, which means that the 2021 corpus was also the smallest. These limitations influence corpus size, and therefore sentence sampling for the sentiment analysis.

Nevertheless, the neutral (borderline negative) sentiment of Social Integration is consistent with prior literature which had established social integration discourse in Germany as being latent with criticism towards foreigners, rather than an optimistic (or positive) welcome culture. Additionally, this category achieved the most intercoder agreement of all four categories, with a Cohen´s Kappa score of 0.642. Table 9 demonstrates the percentage of agreement for each category.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Cohen’s Kappa and intercoder agreement.

https://doi.org/10.1371/journal.pone.0340164.t009

By stark contrast, the most disagreement was found on the category Limiting Migration (0.41). The most likely culprit of this disagreement is: of all four categories, Limiting Migration had the most “interdisciplinary” discourse; it contained bureaucratic or policy-oriented discourse on limiting migration, (including crimes committed by foreigners and border control), which is easily construed as anti-immigrant. If we had used three categories rather than four, it is likely that there would have been higher intercoder agreement. Most sentences conveyed negative sentiment toward migrants or migration-related issues (17 out of 36, 47%). Positive sentiment, emphasizing legal fairness, integration, or recognition of migrant contributions, was less frequent (9 sentences, 25%). Neutral or informational statements accounted for the remaining sentences (10 sentences, 28%).

This pattern suggests that discourse in this category is predominantly critical, reflecting concerns over migration restrictions, social tensions, or enforcement policies, while only a minority of texts highlight positive or neutral perspectives on migration management. This also affirms the appropriate label assigned to the category based on the most prevalent terminology that had been clustered with LSA.

Unsurprisingly, the category Xenophobia contained predominantly negative discourse with a Cohen´s agreement of 0.61, which is considered a sufficient degree of agreement. It had the second-highest agreement (after Social Integration). This supports prior literature which had reported on the use of key “Ausländer” terms (Ausländerfeind in particular) as being historically negative in connotation, although not exclusively so. In the Xenophobia discourse, around 59% of sentences are negative, 26% positive, and 15% neutral. These proportions are estimated with 95% confidence intervals of 0.49–0.68 (negative), 0.17–0.37 (positive), and 0.09–0.25 (neutral), indicating that the discourse is primarily negative, with some positive content highlighting anti-xenophobia efforts, and relatively few neutral statements. It is interesting to note that the Xenophobia category contains more “positive” discourse than the Social Integration discourse, presumably because the category discussing Xenophobia includes discourse combating Xenophobia. It is important to see the distinction between discourse which is critical of Xenophobia and pro-integration discourse: Discourse which discusses Xenophobia in a critical way is often addressing a problem with intolerant German natives, which is entirely different from a discussion surrounding the migrant population.

Regarding Administration and Policy – both human and Chat-GPT coders found a negative bias in this discourse: Cohen´s agreement for the sentiment analysis was 0.485, which is considered moderate. In this category, sentiment analysis of 62 consensus-labeled sentences revealed a predominance of negative sentiment (57.4%, 95% CI = 44.2% – 69.7%), followed by neutral (26.2%, 95% CI = 15.8% – 38.9%) and positive sentences (18.0%, 95% CI = 9.4% – 30.0%). Negative sentences highlighted bureaucratic obstacles, restrictive policies, and challenges faced by immigrants when interacting with local or federal authorities. Positive sentences focused on administrative initiatives, support programs, and constructive actions by Ausländerbeauftragte or advisory boards. Neutral sentences largely contained factual or statistical information regarding registries, legal provisions, or the role of administrative actors. The confidence intervals reflect some uncertainty due to the sample size of sentences analyzed.

Therefore, the sentiment analysis results suggest that it was easier and more reliable to label Social Integration and Xenophobic discourse. One potential reason for the higher agreement towards “integration” discourse is that matters of anti-immigrant are more nuanced in nature: discourse may be critical of migration politics but not anti-immigrant, per se. Each category (except Social Integration) was found to be predominantly negative, with varying degrees of agreement.

Summary of results and discussion

In response to our first research question: we found four main topics as frames of the media discourse with a reference to the overacting topic “Ausländer”. However, these theoretically could have been reduced to three. Our second research question addresses the change of discourse over time: The media discourse trend shows that the prevalence of xenophobic discourse spiked in 2006 and again in 2016. In regard to the categories Xenophobia and Limiting Migration, our hypothesis was correct, although the difference between discourse was significant for Xenophobic discourse relative to other years of discourse within that category, whereas 2016 Limiting Migration discourse did increase, the difference relative to other years in that category were not significant. The hypothesis expecting anti-immigrant discourse to rise in the mid 2010s was therefore correct, albeit with caveats. The second half of the hypothesis (pertaining to social integration) was accurate, in that social integration discourse did increase after mid 2010s; despite that integration discourse had been slowly rising over the years from a descriptive standpoint, only the year remains significantly different from the others. Lastly, in regard to our third question regarding sentiment: Results reflect a negative sentiment for all discourse categories except Social Integration, which was mostly neutral (by a narrow margin). However, results show that Social Integration discourse is still more negative than positive, upholding the theoretically jaded framework surrounding Germany´s “welcome culture”. In this regard, our hypothesis based on integration skepticism was accurate: integration discourse is not inherently positive, despite the increase over time. The sentiment analysis for Social Integration had the highest agreement, followed by Xenophobia. The polarized discourse is likely harder to define, whereas Administration and Limiting Migration had lower agreement. Potential options for improving agreement include dichotomizing the sentiment (Positive and Negative, no Neutral), testing if agreement is better between two native speakers (as opposed to one native speaker and ChatGPT). Another option would be to introduce alternative forms of sentiment analysis, such as a pre-trained German lexicon approach. In addition to the triangulation of results, here are a number of limitations/challenges to consider:

1. Our analysis was restricted to the term “foreigner”. Therefore, these results only attest to how discourse pertaining to foreigner compound buzzwords has changed over time. Although the term “Ausländer” (foreigner) has historical relevant and controversy, the term become antiquated(and politically incorrect); alternative terms such as migrant and immigrant may have grown in prevalence. Therefore, it is of crucial interest to extend this analysis, to compare the prevalence and usage other terms, and to understand the contexts in which these alternative terms are used. Hence, as the term “foreigner” is a relevant reference point in the German language and can be used to examine the etymology of synonyms (such as “migrant” or “immigrant”).
2. The success of this application on our corpus may be largely due to the nature of the German language itself; the German language uses and creates compound words so often, which allowed us to streamline the search for discourse on foreigners, and terms describing foreigners (Ausländer compound terms). This is not so much a limitation for the current analysis but poses a challenge to implement this method with text in other languages (such as English).
3. It would be remiss to not discuss the differences between prevalence and salience. The present paper is concerned with the former, as we cannot attest to the effects of these terms on media consumers (readers). However, we can and do attest to the changes in prevalence of key words and categorically overall.