Migration is a phenomenon faced by many countries, which brings a variety of effects; both in the population from which it emigrates and in the receiving population. One of the effects that worries many countries is intolerance and hostile attitudes toward immigrants. These attitudes have been the focus of many research studies, some of which are focused on individual-level psychological and socio-economic factors (Burns and Gimpel, 2000; Scheve and Slaughter, 2001), and others on the contact between immigrant population and locals (Brown et al., 2003; Hopkins, 2010; Jolly and DiGiusto, 2014). The main methods used in these studies are based on context specific surveys, which makes replication in others societies or countries difficult. The theories that explain the type of attitudes of locals interacting with immigrants can be summarized in two: the Intergroup Contact Theory (Allport et al., 1954), and the Integrated Threat Theory (Stephan and Stephan, 2013; Nelson, 2009). The former states that people support multiculturalism and integration. The latter, that people think that immigrants will bring negative effects for their society, including competition for jobs and public services, worsening of the national economy, increase in crime, and the arrival of diseases. Particularly, the attitudes explained by the threat theory can lead to acts of violence, discrimination, and abuse; thus, it is important to understand what factors enhance such attitudes.
However, measuring attitudes is costly and impractical under dynamic scenarios. The most frequent methods are surveys, which are difficult and costly to implement. In this paper, we propose to make use of the information that people publish in Twitter as a proxy of their attitudes toward immigration. It is common to find reactions and attitudes through posts in these platforms, where people express their ideas and opinions voluntarily. We propose to define a spectrum of attitudes based on the two aforementioned theories, and to classify users and tweets into that spectrum. We do so with a semi-supervised topic modeling technique named Topic-Supervised Non-Negative Matrix Factorization(MacMillan and Wilson, 2017)
. TS-NMF works in a semi-supervised way because some users can be labeled as belonging to each extreme of the spectrum, something that we do with custom-built lexicons for each theory.
We perform a descriptive case study on the Chilean society, because Chile is one of the countries in which migration has reached unprecedented volume in recent years. The statistics show that immigrant population has increased from 0.8% in 1992 to 4.35% in 2017; and where 66.7% of immigrants declare to have arrived mainly in 2016 ((INE), 2018). For this, Chileans have developed diverse perceptions regarding the number of immigrants in the country and the phenomenon itself. To measure them with our proposed method, we collected more than 206K tweets that discuss immigration in Chile, written by more than 49K users during the year 2017. After inferring user and tweets positions in the spectrum, we performed lexical and network analysis with respect to the spectrum position. In the lexical analysis, we used the “Linguistic Inquiry and Word Count” (LIWC) lexicon (Pennebaker et al., 2001), typically employed to characterize cognitive and emotional differences in discourse (Harman and Dredze, 2014; Coppersmith et al., 2014; González-Ibánez et al., 2011)
. To analyze the network structure, we estimated the polarization of the retweet and mention networks between users.
As main results, we observed that most of the discussion toward migration in Chile is targeted at Haitian migration, even though other countries have a larger share of the population. We found lexical differences in how each attitude discussed migration, and those differences were consistent with theories. For instance, social-related words were correlated with empathetic attitudes, job- and money-related words were correlated with threatening attitudes. In the network, the retweet network was polarized, in coherence as predicted by other studies regarding political discussion (Conover et al., 2011; Graells-Garrido et al., 2015). Finally, we notice that the amount and tendency of the tweets (the latter reflects the attitude towards immigration) seems to be influenced by relevant news events on national migration issues. These results can inform public policy designers to improve inter-group relations in the country, as well as increasing the understanding of how people feel regarding an important aspect of globalization.
In summary, the contribution of this paper is two-fold. We proposed a methodology to characterize local attitudes toward migration from tweets. Then, we performed a descriptive case study in Chile using this method, obtaining results that are coherent with social theory, with added depth based on the rich information that can be extracted from Twitter.
This paper is structured as follows. Section 2 discusses the related work. Section 3 describes the social theories that guided our analysis. Section 4 describes the data set we analyzed. Section 5 describes the methodology. Section 6 describes the results of applying the methodology to the data set. Section 7 discusses the implications of our work. Finally, Section 8 states our conclusions.
2. Related Work
Migration is a widely studied topic because there are many issues associated to this phenomenon. Some researchers have focused on studying the economic impacts related to migration (Scheve and Slaughter, 2001; Hainmueller and Hiscox, 2007; Hanson et al., 2007; Hainmueller and Hiscox, 2010), others on social cohesion (Sniderman et al., 2000; Hopkins, 2010; Kopstein and Wittenberg, 2009; Jolly and DiGiusto, 2014). Within these studies, those who have focused on integration (Lamanna et al., 2018) and racism/xenophobia stand out (Brown et al., 2003; Herek, 1986; Pettigrew and Tropp, 2006). Our work seeks to contribute in the latter area, mainly due to the subject of our case study, Chile, a society that in a short time has faced a massive influx of immigrants. Migration in Chile has been a national issue, causing controversy in presidential elections, news, and municipal institutions. However, measuring attitudes is not a simple problem, nor a solved one. Twitter is currently a platform widely used in studies of human behavior, since it provides a valuable source of data. Studies that have used Twitter have allowed to reveal socio-cultural characteristics of users or societies, including the level of integration of immigrants in a city (Lamanna et al., 2018), attitudes in response to triggering events, such as terrorist attacks (Darwish et al., 2017), the influence of culture in personal actions (Garcia-Gavilanes et al., 2013), political polarization (Garcia-Gavilanes et al., 2013), personality traits (Quercia et al., 2011), and personality differences between democrats and republicans (Sylwester and Purver, 2015).
Given this body of research, we propose that Twitter can be used as a proxy to understand human behavior, in our case, the attitudes of Chileans regarding immigration.
3. Social Theories
The attitudes toward immigration are varied and depend on economic, socio-cultural and psychological factors. In this context, psychology and sociology have defined theories that explain the attitudes exhibited by people, who belong to different groups, when interacting with others: the Intergroup Contact Theory (Allport et al., 1954), and the Integrated Threat Theory (Stephan and Stephan, 2013; Nelson, 2009). The attitudes toward immigration are a particular case explained by these theories.
Intergroup Contact Theory
Developed in the book “The Nature of Prejudice” by Gordon W. Allport (Allport et al., 1954), it postulates that prejudices are reduced when the interaction between different groups meets the following conditions: 1) groups are on equal terms; 2) they have common goals; 3) there is cooperation; and 4) there is support from formal and/or informal institutions. The theory states that intergroup contact reduces the fear and anxiety that exists when people interact with an unknown group (Stephan and Stephan, 1985), and that it promotes empathy and understanding towards the foreign group (Stephan and Finlay, 1999).
This theory has been used to ground several studies: contact between white and black people (Brown et al., 2003), heterosexuals and homosexuals (Herek, 1986; Herek and Capitanio, 1996), minority religious groups (Novotny and Polonsky, 2011), and locals and immigrants (Hopkins, 2010). All these studies conclude that contact improves relationships between groups.
Integrated Threat Theory
In contrast to the Intergroup Contact Theory, the Integrated Threat Theory argues that contact between disparate groups provokes perceptions of threat and contempt (Stephan and Stephan, 2013; Nelson, 2009), for instance, due to competition for work and economic resources (Ha, 2010; Esses et al., 2001). Furthermore, the threat does not have to be real, it can be subjective or fictitious (Kopstein and Wittenberg, 2009).
This theory postulates that, when the interaction conditions are not optimal, the contact between different groups will provoke conflicting and hostile relationships. The concept of “contact” is not limited only to physical contact, it can also be indirect (Dovidio et al., 2011), imagined (Crisp and Turner, 2009), and electronic (Amichai-Hamburger and McKenna, 2006; White and Abu-Rayya, 2012).
Both theories tell us what to search when we look attitudes toward immigration: attitudes motivated by empathy, in favor of immigration; and attitudes motivated by threat, against immigration. As such, we will assume that there are two attitudes, which we label as empathy and threat.
4. Data Set Description
In this section we describe our data set of posts from Twitter about migration in Chile.
Twitter is a micro-blogging platform, where users publish tweets (posts) with a maximum of 280 characters. Users may follow others, to see their tweets in their own timelines. Tweets may mention other users, quote other tweets, or retweet another tweet to share it with one’s audience. Users can report a screen name, a full name (which can be real or fictitious), a location (real, fictitious, or empty), and a small autobiography, among other attributes. To collect tweets that talk about immigration in Chile we used the Twitter Streaming API using system designed to crawl Chilean tweets (Graells-Garrido et al., 2016). The query parameters were keywords related to immigration (e.g., inmigración, inmigrante, fronteras, racismo, etc.), and origin countries with their respective demonyms (e.g., Haití–haitianos/as, Venezuela–venezolanos/as, Perú–peruanos/as, etc.). Given how generic some of these keywords are, particularly regarding the context of political issues of neighbouring countries, and the presidential elections held in Chile during November and December, we performed extensive manual clean-up of the data set.
In total, our data set is comprised by 206,353 tweets that discuss immigration in Chile during 2017, written by 49,346 users. Figure 1 shows the weekly volume of tweets. As seen on the figure, the amount of tweets has a sligth positive trend. Two peaks draw our attention: July 31th, when the news reported a case of an Haitian citizen with Leprosy; and November 19th, when an Haitian citizen rescued a woman who fell from the ninth floor of a building.
Regarding content, Figure 2 shows the most frequent words, after removing stopwords and accents. One can see that words such as Haití and Haitianos are more relevant than other countries name or demonyms, despite the fact that the largest immigrant population comes from Perú, Colombia and Venezuela ((INE), 2018). Also, Santiago and Antofagasta are two frequent keywords, the two cities with the largest immigrant population ((INE), 2018). Other relevant words that appear are: “gobierno” (government), “carabineros” (policemen), “Piñera” (current president, and presidential candidate in 2017) and “proyecto” (project); possibly because during the year an immigration reform was being discussed.
|Empathy||#todossomosmigrantes, #stopxenophobia, #chilesinbarreras, #chileterecibe, #bienvenidosmigrantes, @oimchile, bienvenidos a chile, #derribandomuros, @sjmchile, …|
|Threat||vendepatria, #nomasinmigrantes, #nomasilegales, #inmigrantesilegales, inmigrantes delincuentes, inmigracion descontrolada, indeseables, …|
We explored the data set to seek for words, phrases, and hashtags that could be mapped to the empathy and threat attitudes. In empathy we chose terms that indicated that immigrants are welcome and will be received in equal conditions (e.g., “we are all immigrants”). In threat we chose terms and words that showed that immigrants are not welcome and qualified them negatively (e.g., “illegal immigrants”). Table 1 shows some examples of the the terms we associated to both attitudes. These labeled terms are not necessarily frequent, however, the methodology that we describe in the next sections allows to propagate these labels through a topic model.
In this section we describe how to characterize users and tweets according to their attitude toward immigration. We define how to apply machine learning techniques to user profiles to derive user-attitude and term-attitude associations. Then, we define how to characterize attitudes from sentiment, lexical and network perspectives.
5.1. Attitudes and Topic Modeling
Topic models are a family of techniques used to discover the underlying semantic structure of a corpus by identifying and quantifying the importance of representative themes in all documents (Blei, 2012). Topic models assume that each text document is generated by a set of topics which have a determined distribution. At the same time, each topic is defined by a set of words, which also have a particular distribution for each topic.
A popular topic modeling technique is Non-negative Matrix Factorization (Lee and Seung, 1999). NMF works by constructing a -rank factorization of a positive document-term matrix into . Matrices and are estimated by minimizing the following objective function:
where is the Frobenius norm. In topic modeling, , and have a special interpretation: is the number of topics, quantifies the relevance of topic in document , and quantifies the relevance of term in topic .
Typical topic modeling applications select different numbers of based on metrics such as perplexity. However, the meaning of topics is not always interpretable, as the factorization may follow latent patterns not necessarily aligned with human expectations. Based on the social theories described in Section 3, we propose to guide the learning procedure to seek for two topics: one that represents empathy, and another that represents threat. In such cases, supervised methods could be employed, however, these methods require a fully labeled data set, not available in our case. Since it is possible to map specific terms (words, phrases, hashtags, URLs, etc.) into these two topics, we propose to use a semi-supervised version of NMF known as Topic-Supervised NMF (MacMillan and Wilson, 2017). TS-NMF defines the minimization problem as follows:
where is the Haddamard product operator, and is a supervision matrix, defined as if topic contributes to the document , and if the topic does not contribute to the document . Thus, TS-NMF allows to provide examples of documents labeled with known topics, and to restrict the latent representation of the corpus to align with the labeled examples.
In our context, we work with user profiles, i.e., the concatenation of tweets by a single user is one document. As terms we consider hashtags, mentions, URLs, and -grams with up to four. This allows us to define how specific phrases are mapped to each topic. The user corpus is represented as a document-term matrix weighted with TF-IDF (Baeza-Yates and Ribeiro-Neto, 2011), and then row-normalized with L2 norm. To label users in the supervision matrix, we construct a list of seed terms for each theory. Then, for each row in we estimate a preliminary attitude score for each topic, by adding the values of the cells of the corresponding seed terms. All users with a score above a certain threshold are labeled with the corresponding topic. In our experiments, we defined a threshold of 0.25, implying that only users who strongly used the seed terms of each topic were labeled.
As result, we obtain , where the rank of and is two. In our context, each topic is an attitude, the matrix contains the user-attitude associations, and the matrix
contains the term-attitude associations (transposed). We interpret these associations as probabilities.
5.2. Attitude Tendency and Polarity
To characterize attitudes, we calculate two metrics common in the sentiment analysis literature to measure the leaning and amount of sentiment: tendency and polarity(Kucuktunc et al., 2012). Tendency is defined as:
where, is the association between user and the corresponding attitude. Note that the definition is analog for terms. For tweets, tendency is defined as:
Note that tendency values close to zero do not imply a neutral attitude, as there could be non-zero contributions in both topics. To clarify this fact, we consider attitude polarity as the amount of associations to both attitudes, defined for users as:
The definition for terms is analog. For tweets, polarity is defined as:
In this way, tendency will allow us to group users/tweets (according to their attitude), while polarity will allow us to measure the intensity of the discussion (how polarized is the attitude).
5.3. Lexical Characterization
The previous metrics give an overview of user and tweet attitudes. The next step is to characterize grouped tweets belonging to each attitude according to their tendency. To do so, we use a psycho-linguistic lexicon named “Linguistic Inquiry and Word Count” (Pennebaker et al., 2001). LIWC is a lexicon used to study emotional, cognitive and structural components contained in a text. In its Spanish version, it contains 7,515 words classified in one or more of 72 categories. Categories are classified into four dimensions: 1) standard linguistic processes (e.g., articles, prepositions, pronouns, etc.); 2) psychological processes (e.g., positive and negative emotions); 3) relativity (e.g., time, verb tense, motion, space); and 4) personal matters (e.g., sex, death, home, occupation, etc.). LIWC categories are organized hierarchically, for instance, all words related to the category anger are also organized in the categories of negative emotions or affect words.
We seek to estimate the association of tweets by tendency groups to LIWC categories. After classifying tweets into groups, we estimate how associated the words in LIWC are to each group. Note that specific events may entice a more active discussion by either group, increasing the amount of tweets, thus, we need a way to control the association with these activity patterns. In previous work, this has been done to estimate gross community metrics with -scores (Kramer, 2010; Quercia et al., 2012). In our case, the definition is as follows:
where, is the association of LIWC category with the tendency , is the mean of fraction of words in in each tweet with tendency , is the mean of fraction of words in in all tweets, and
is the standard deviation of the fraction of words inin all tweets. Hence, this relative metric allows us to compare behavior between groups, by controlling for external variability.
5.4. Network Assortativity
The previous definitions capture the behavior in expression, however, the social aspect of Twitter allows to also capture network behavior. We focus on two different networks: the mention network, related to discussion, and the retweet network, related to information diffusion. In both networks, node are users, and links are weighted relations between users. Each node has as attributes its associations to each attitude. In the mention network, a directed link between users and exists if mentions in one or more tweets. The link weight is the number of times this happens. In the retweet network, a directed link between users and exists if republishes content by . The link weight is the number of times that one user retweets another. These kind of networks are commonly analyzed to understand polarization (Conover et al., 2011). To be able to analyze connectivity, we will focus on the Largest Strongly Connected Component of each network.
To analyze the networks structure, we estimate the assortativity coefficient with respect to each attitude. The assortativity coefficient is the Pearson correlation coefficient of numerical attributes between pairs of linked nodes (this numerical attributes are the attitudes given by the model). It measures the similarity of connections in the graph with respect to the given numeric attribute (Newman, 2003). Hence, the assortativity coefficient measures whether people relations are homophilic with respect to attitude. This behavior is commonly found in networks (Barberá, 2015), and it has been documented in Twitter political discussion (Conover et al., 2011), including in Chile (Graells-Garrido et al., 2015).
In the next section we apply this methodology to the data set described in Section 4, covering an entire year of discussion about immigration in Chile.
Figure 3 shows the association of words with each attitude, empathy on top, threat on bottom. One can see that words associated to empathy include “integración” (integration), “salud” (health), and “educación” (education), reflecting their empathetic attitude. Words associated to threat include “delincuentes” (delinquents), “control” (control), and “ilegales” (illegals), reflecting a feeling of threat. Also, empathy group uses the word “Migrantes” (migrants) and threat group uses “Inmigrantes” (immigrants), which can be interpreted as that the empathy group is concerned about the general phenomenon (migration includes emigration and immigration), while the threat group only for the particular phenomenon (immigration).
Tendency and Polarity
Figure 4 shows the distribution of tendency and polarity for users. One can see that the distributions are fairly symmetric, with peaks in the center of the distribution. Figure 5 shows the tendency and polarity of tweets during the year under study, estimated using LOWESS. One can see that the tendency trend exhibits two interesting periods, before and after the news about the Leprosy case of an Haitian in July 31th. In the first period, tendency is slightly negative (threat), with an arguably low variability. In the second period, variability increases, and a small negative trend appears, even though at a point in time it reaches its maximum value (i.e., maximum empathy) at the beginning of October. This could be explained by a news event reported in October 6th, about a Colombian citizen that gave birth on the street because a taxi driver expelled her from his car.
It is interesting that both news are related with the Integrated Threat theory and Intergroup Contact theory, respectively. On the one hand, the first event shows the immigrant as a threat, being a possible source of contagion of a disease (Leprosy). On the other hand, the second event shows the immigrant being a victim of violence and discrimination, which arguably makes people more empathetic. Regarding polarity, the trend exhibits a gradual increase in time, with two interesting peaks. The first one reflects the Leprosy case, and the second one reflects the presidential elections, where migration was a common topic in discussion.
Figure 6 shows the differences of cognitive and emotional categories from LIWC in tweets grouped by tendency: empathy contains all tweets with tendency 0; threat, otherwise. For each category and group, we estimated the -score for all tweets each month. As a general observation, one can see that both groups tend to have opposite behaviors. For instance, tweets in the empathy group are positively associated to the sociability, family, and positive emotions category more than tweets in the threat group. Conversely, tweets in the threat group are positively associated with money, job, and inhibition categories. This could be explained by the threat theory, as immigrants can be perceived as an economic threat and labor competition. Also, inhibition category can be interpreted by the desire to prohibit the arrival of more immigrants or to prevent them from accessing social benefits.
Mention and Retweet Networks
The largest SCC of the retweet network has nodes and edges, while the largest SCC of the mention network has nodes and links. Figure 7 visualizes both networks using Hierarchical Edge Bundling (Holten, 2006). This method allows us to make explicit the adjacency relations between users, as similar edges are bundled to decrease visual clutter. In the figure, each link is colored according to tendency of the source node (purple: empathy group, orange: threat group). Note that the visual encoding makes explicit the community structure in the retweet network and the heterogeneity of the mention network.
The assortativity coefficient for the retweet network are 0.26 (empathy) and 0.14 (threat), implying that homophilic behavior exists, but it is not as strong as in other topics (for instance, the discussion about abortion in Chile is greater (Graells-Garrido et al., 2015)), and it is not equal in both groups. As hinted by the visualization, in the mention network the results are small: 0.06 (empathy) and 0.08 (threat). Thus, the retweet network is more segregated than the mention network. This could be explained because retweets are expected to be seen by all followers, and are a key factor in information diffusion, while mentions and replies are not. For instance, one user may send tweets to another holding an opposite position, but if there is no reply, then the interaction is not meaningful.
7. Discussion and Future Work
Migration is a controversial issue in Chile, and, although there are some studies about Chileans attitudes toward immigration (Lawrence, 2015; Carvacho, 2010), they do not cover recent migration patterns. To complement knowledge about this topic, we defined a way to classify and measure attitudes, enabling to study the dynamics of perception with respect to immigration and performed a descriptive study of how immigration is perceived in Chile, according to Twitter discussion.
Our results may inform policy and intervention design, as it quantifies how people feel and communicate with respect to immigration. This is relevant, as there exists several contact strategies to improve relationships between social groups (Pettigrew and Tropp, 2006). For instance, the discussion we analyzed is mostly targeted at Haitian migration. A majority of them is from Afro-Haitian descent, an ethnicity that was almost non-present in Chile.
There are two key aspects that need further exploration, and that limit the scope of our results: the representativity of Twitter, and the validation of the TS-NMF model. In terms of representativity, Twitter is a biased sample of the population (Baeza-Yates, 2018). As such, our results only cover this sample, even though it is not know to which degree nor to which sub-populations it represents. Having these biases into account will surely improve the interpretation of results. However, one aspect that needs to be considered is that Twitter is within the most popular applications in Chile (Graells-Garrido et al., 2018), and that it reflects some cultural aspects, such as the country’s centralization (Graells-Garrido and Lalmas, 2014). In terms of validation, the lack of ground truth or approximate measures of the problem stands in the way of effectively measuring the model accuracy, leaving us only with a qualitative evaluation.
Besides working on the limitations of our approach, there are two lines of future work that we devise. On the one hand, it would be relevant to understand the relationship between attitudes and actual presence of immigrants in a place. This would provide a way to measure real and imagined threat attitudes (Kopstein and Wittenberg, 2009). On the other hand, there is a potential influence of news events in attitudes. Given the rise of fake news and post-truth media, this would provide a way to measure the effect of such phenomena on how people feel with respect to a specific issue, migration in this case.
In this paper, we have characterized attitudes toward immigration by locals in Chile. We used a semi-supervised topic modeling technique (TS-NMF (MacMillan and Wilson, 2017)) to identify attitudes grounded in two social theories, the Intergroup Contact Theory (Allport et al., 1954), and the Integrated Threat Theory (Stephan and Stephan, 2013; Nelson, 2009). Then, we measured differences in attitudes using psycho-linguistic lexicons and interaction networks. As result, we found consistent behaviour with respect to social theory. There is still work to do in the evaluation and representativeness of our model, including the definition of a suitable ground-truth perception to validate our proposal. We believe our results help to inform the design of public policy and interventions to improve relations between groups in a country.
- Allport et al. (1954) Gordon Willard Allport, Kenneth Clark, and Thomas Pettigrew. 1954. The nature of prejudice. (1954).
- Amichai-Hamburger and McKenna (2006) Yair Amichai-Hamburger and Katelyn YA McKenna. 2006. The contact hypothesis reconsidered: Interacting via the Internet. Journal of Computer-Mediated Communication 11, 3 (2006), 825–843.
- Baeza-Yates (2018) Ricardo Baeza-Yates. 2018. Bias on the web. Commun. ACM 61, 6 (2018), 54–61.
- Baeza-Yates and Ribeiro-Neto (2011) Ricardo Baeza-Yates and Berthier Ribeiro-Neto. 2011. Modern Information Retrieval: the concepts and technology behind search, 2nd. Edition. Addison-Wesley, Pearson.
- Barberá (2015) Pablo Barberá. 2015. Birds of the same feather tweet together: Bayesian ideal point estimation using Twitter data. Political Analysis 23, 1 (2015), 76–91.
- Blei (2012) David M Blei. 2012. Probabilistic topic models. Commun. ACM 55, 4 (2012), 77–84.
- Brown et al. (2003) Kendrick T Brown, Tony N Brown, James S Jackson, Robert M Sellers, and Warde J Manuel. 2003. Teammates on and off the Field? Contact with black teammates and the racial attitudes of white student athletes 1. Journal of applied social psychology 33, 7 (2003), 1379–1403.
- Burns and Gimpel (2000) Peter Burns and James G Gimpel. 2000. Economic insecurity, prejudicial stereotypes, and public opinion on immigration policy. Political science quarterly 115, 2 (2000), 201–225.
- Carvacho (2010) Héctor Carvacho. 2010. Ideological configurations and prediction of attitudes toward immigrants in Chile and Germany. International Journal of Conflict and Violence (IJCV) 4, 2 (2010), 220–233.
- Conover et al. (2011) Michael Conover, Jacob Ratkiewicz, Matthew R Francisco, Bruno Gonçalves, Filippo Menczer, and Alessandro Flammini. 2011. Political polarization on Twitter. Proc. of ICWSM 133 (2011), 89–96.
- Coppersmith et al. (2014) Glen Coppersmith, Mark Dredze, and Craig Harman. 2014. Quantifying mental health signals in Twitter. In Proceedings of the Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality. 51–60.
- Crisp and Turner (2009) Richard J Crisp and Rhiannon N Turner. 2009. Can imagined interactions produce positive perceptions?: Reducing prejudice through simulated social contact. American psychologist 64, 4 (2009), 231.
- Darwish et al. (2017) Kareem Darwish, Walid Magdy, Afshin Rahimi, Timothy Baldwin, and Norah Abokhodair. 2017. Predicting Online Islamophopic Behavior after# ParisAttacks. The Journal of Web Science 3, 1 (2017).
- Dovidio et al. (2011) John F Dovidio, Anja Eller, and Miles Hewstone. 2011. Improving intergroup relations through direct, extended and other forms of indirect contact. Group processes & intergroup relations 14, 2 (2011), 147–160.
- Esses et al. (2001) Victoria M Esses, John F Dovidio, Lynne M Jackson, and Tamara L Armstrong. 2001. The immigration dilemma: The role of perceived group competition, ethnic prejudice, and national identity. Journal of Social issues 57, 3 (2001), 389–412.
- Garcia-Gavilanes et al. (2013) Ruth Garcia-Gavilanes, Daniele Quercia, and Alejandro Jaimes. 2013. Cultural dimensions in Twitter: Time, individualism and power. Proc. of ICWSM 13 (2013).
- González-Ibánez et al. (2011) Roberto González-Ibánez, Smaranda Muresan, and Nina Wacholder. 2011. Identifying sarcasm in Twitter: a closer look. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers-Volume 2. Association for Computational Linguistics, 581–586.
- Graells-Garrido et al. (2018) Eduardo Graells-Garrido, Diego Caro, Omar Miranda, Rossano Schifanella, and Oscar F Peredo. 2018. The WWW (and an H) of Mobile Application Usage in the City: The What, Where, When, and How. In Companion of the The Web Conference 2018 on The Web Conference 2018. International World Wide Web Conferences Steering Committee, 1221–1229.
- Graells-Garrido and Lalmas (2014) Eduardo Graells-Garrido and Mounia Lalmas. 2014. Balancing diversity to counter-measure geographical centralization in microblogging platforms. In Proceedings of the 25th ACM conference on Hypertext and social media. ACM, 231–236.
- Graells-Garrido et al. (2015) Eduardo Graells-Garrido, Mounia Lalmas, and Ricardo Baeza-Yates. 2015. Finding intermediary topics between people of opposing views: a case study. In Social Personalisation & Search, Christoph Trattner, Denis Parra, Peter Brusilovsky, and Leandro Balby Marinho (Eds.). CEUR, Santiago, Chile.
- Graells-Garrido et al. (2016) Eduardo Graells-Garrido, Mounia Lalmas, and Ricardo Baeza-Yates. 2016. Encouraging diversity-and representation-awareness in geographically centralized content. In Proceedings of the 21st International Conference on Intelligent User Interfaces. ACM, 7–18.
- Ha (2010) Shang E Ha. 2010. The consequences of multiracial contexts on public attitudes toward immigration. Political Research Quarterly 63, 1 (2010), 29–42.
- Hainmueller and Hiscox (2007) Jens Hainmueller and Michael J Hiscox. 2007. Educated preferences: Explaining attitudes toward immigration in Europe. International organization 61, 2 (2007), 399–442.
- Hainmueller and Hiscox (2010) Jens Hainmueller and Michael J Hiscox. 2010. Attitudes toward highly skilled and low-skilled immigration: Evidence from a survey experiment. American political science review 104, 1 (2010), 61–84.
- Hanson et al. (2007) Gordon H Hanson, Kenneth Scheve, and Matthew J Slaughter. 2007. Public finance and individual preferences over globalization strategies. Economics & Politics 19, 1 (2007), 1–33.
- Harman and Dredze (2014) GACCT Harman and Mark H Dredze. 2014. Measuring post traumatic stress disorder in Twitter. In ICWSM (2014).
- Herek (1986) Gregory M Herek. 1986. The instrumentality of attitudes: Toward a neofunctional theory. Journal of Social Issues 42, 2 (1986), 99–114.
- Herek and Capitanio (1996) Gregory M Herek and John P Capitanio. 1996. “Some of my best friends” Intergroup Contact, concealable stigma, and heterosexuals’ attitudes toward gay men and lesbians. Personality and social psychology bulletin 22, 4 (1996), 412–424.
- Holten (2006) Danny Holten. 2006. Hierarchical edge bundles: Visualization of adjacency relations in hierarchical data. IEEE Transactions on visualization and computer graphics 12, 5 (2006), 741–748.
- Hopkins (2010) Daniel J Hopkins. 2010. Politicized places: Explaining where and when immigrants provoke local opposition. American political science review 104, 1 (2010), 40–60.
- (INE) (2018) National Statistics Institute (INE). 2018. Síntesis resultados Censo 2017. urlhttps://www.censo2017.cl/descargas/home/sintesis-de-resultados-censo2017.pdf.
- Jolly and DiGiusto (2014) Seth K Jolly and Gerald M DiGiusto. 2014. Xenophobia and immigrant contact: French public attitudes toward immigration. The Social Science Journal 51, 3 (2014), 464–473.
- Kopstein and Wittenberg (2009) Jeffrey S Kopstein and Jason Wittenberg. 2009. Does familiarity breed contempt? Inter-ethnic contact and support for illiberal parties. The Journal of Politics 71, 2 (2009), 414–428.
- Kramer (2010) Adam DI Kramer. 2010. An unobtrusive behavioral model of gross national happiness. In Proceedings of the SIGCHI conference on human factors in computing systems. ACM, 287–290.
- Kucuktunc et al. (2012) Onur Kucuktunc, B Barla Cambazoglu, Ingmar Weber, and Hakan Ferhatosmanoglu. 2012. A large-scale sentiment analysis for Yahoo! answers. In Proceedings of the fifth ACM international conference on Web search and data mining. ACM, 633–642.
- Lamanna et al. (2018) Fabio Lamanna, Maxime Lenormand, María Henar Salas-Olmedo, Gustavo Romanillos, Bruno Gonçalves, and José J Ramasco. 2018. Immigrant community integration in world cities. PloS one 13, 3 (2018), e0191612.
- Lawrence (2015) Duncan Lawrence. 2015. Crossing the Cordillera: immigrant attributes and Chilean attitudes. Latin American Research Review 50, 4 (2015), 154–177.
- Lee and Seung (1999) Daniel D Lee and H Sebastian Seung. 1999. Learning the parts of objects by non-negative matrix factorization. Nature 401, 6755 (1999), 788.
- MacMillan and Wilson (2017) Kelsey MacMillan and James D Wilson. 2017. Topic supervised non-negative matrix factorization. arXiv preprint arXiv:1706.05084 (2017).
- Nelson (2009) Todd D Nelson. 2009. Handbook of prejudice, stereotyping, and discrimination. Psychology Press.
- Newman (2003) Mark EJ Newman. 2003. Mixing patterns in networks. Physical Review E 67, 2 (2003), 026126.
- Novotny and Polonsky (2011) Josef Novotny and Filip Polonsky. 2011. The Level of Knowledge about Islam and Perception of Islam among Czech and Slovak University Students: does Ignorance Determine Subjective Attitudes? Sociologia 43, 6 (2011), 674–696.
- Pennebaker et al. (2001) James W Pennebaker, Martha E Francis, and Roger J Booth. 2001. Linguistic inquiry and word count: LIWC 2001. Mahway: Lawrence Erlbaum Associates 71, 2001 (2001), 2001.
- Pettigrew and Tropp (2006) Thomas F Pettigrew and Linda R Tropp. 2006. A meta-analytic test of intergroup contact theory. Journal of personality and social psychology 90, 5 (2006), 751.
- Quercia et al. (2012) Daniele Quercia, Jonathan Ellis, Licia Capra, and Jon Crowcroft. 2012. Tracking gross community happiness from tweets. In Proceedings of the ACM 2012 conference on computer supported cooperative work. ACM, 965–968.
- Quercia et al. (2011) Daniele Quercia, Michal Kosinski, David Stillwell, and Jon Crowcroft. 2011. Our Twitter profiles, our selves: Predicting personality with Twitter. In Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational Conference on Social Computing (SocialCom), 2011 IEEE Third International Conference on. IEEE, 180–185.
- Scheve and Slaughter (2001) Kenneth F Scheve and Matthew J Slaughter. 2001. Labor market competition and individual preferences over immigration policy. Review of Economics and Statistics 83, 1 (2001), 133–145.
- Sniderman et al. (2000) Paul M Sniderman, Pierangelo Peri, Rui JP de Figueiredo Jr, and Thomas L Piazza. 2000. The Outsider: Politics and Prejudice in Italy.
- Stephan and Stephan (2013) Cookie White Stephan and Walter S Stephan. 2013. An integrated threat theory of prejudice. In Reducing prejudice and discrimination. Psychology Press, 33–56.
- Stephan and Finlay (1999) Walter G Stephan and Krystina Finlay. 1999. The role of empathy in improving intergroup relations. Journal of Social issues 55, 4 (1999), 729–743.
- Stephan and Stephan (1985) Walter G Stephan and Cookie White Stephan. 1985. Intergroup anxiety. Journal of social issues 41, 3 (1985), 157–175.
- Sylwester and Purver (2015) Karolina Sylwester and Matthew Purver. 2015. Twitter language use reflects psychological differences between democrats and republicans. PloS one 10, 9 (2015), e0137422.
- White and Abu-Rayya (2012) Fiona A White and Hisham M Abu-Rayya. 2012. A dual identity-electronic contact (DIEC) experiment promoting short-and long-term intergroup harmony. Journal of Experimental Social Psychology 48, 3 (2012), 597–608.