Multidimensional analysis of social discourse on Twitter (X) around Venezuelan migration in Colombia

A computational study of Twitter, now X, found that discourse about Venezuelan migration in Colombia became more negative in original posts from 2017 to 2018, while reposts disproportionately amplified neutral and positive content. The study analyzed 7.4 million posts collected from January 2014 through June 2022, including 1,664,903 posts classified as relevant and originating in Colombia, and examined changes over time, across regions and in the use of derogatory terms.

The researchers used the former Twitter Streaming API to retrieve original posts, reposts and mentions containing terms associated with Venezuelan nationality or migration. They excluded quote posts because their commentary could endorse, criticize or contradict the quoted material, making sentiment attribution ambiguous. From the initial 7.4 million posts, the researchers inferred Colombian origin for about 2 million posts by combining user-provided location fields with a text-based classification model.

Geolocation inference carried a measured error rate. The selected spaCy model, trained on 270,000 geographically labeled posts, achieved 0.79 accuracy and a macro-averaged F1 score of 0.794 on its held-out test set. The researchers separately checked national changepoints using 167,986 directly geolocated posts from January 2014 through April 2020 and obtained the same break dates found in the full analysis, although this check could not cover 2020–2022. The department-level analysis also used directly geolocated posts.

A second classifier was used to distinguish posts relevant to Venezuelan migration from unrelated posts. Human reviewers labeled 4,684 posts, and the selected spaCy relevance model achieved 0.90 accuracy on a held-out test set. For sentiment classification, two coders assigned 2,990 relevant posts scores on a seven-point scale that was then grouped as negative, neutral or positive. Intercoder agreement was 63.5%, with Cohen’s kappa of 0.56. A fine-tuned Spanish BERT model was subsequently selected to label migration-related posts; its held-out accuracy was 0.699, with class-specific F1 scores of 0.720 for negative sentiment, 0.668 for neutral sentiment and 0.712 for positive sentiment.

Changepoint detection identified statistically significant breaks in monthly sentiment counts during June 2017 through January 2018, with all reported p-values below 10−30. Before the transition, volumes were relatively low and sentiment more balanced. Afterward, negative posts became more prominent, while monthly volumes exceeded pre-break means by 797% to 1,843%, depending on sentiment category and whether reposts were included. The analysis also recorded a negativity spike in November 2019, around the expulsion of 59 Venezuelans accused of inciting unrest, another in early 2020 during the onset of the COVID-19 pandemic, and increased positive sentiment in February 2021 after the announcement of a temporary protection statute for Venezuelan migrants.

Reposting substantially changed the apparent sentiment distribution. Approximately 67% of reposts were neutral or positive, whereas original posts consistently contained a higher proportion of negative sentiment across the study period. The authors distinguish this measured difference between original and reposted content from a claim about individual users’ attitudes: a negative post may express hostility, concern for migrants, criticism of government policy or reporting about adverse events, and the classifier did not identify the motivation behind its sentiment label.

Regional analysis covered six territorial units with high concentrations of Venezuelan migrants. Negative sentiment was dominant in five, with La Guajira the exception, where neutral posts consistently outnumbered negative ones during the observation period. Structural breaks occurred at different times across departments: Valle del Cauca had its earliest detected break in May 2015, while Norte de Santander and La Guajira shifted beginning in late 2016. These are temporal associations within online discourse and do not establish causal effects of migration or local conditions. Department-level correlations were mostly low in magnitude; the Gini coefficient and unemployment rate were negatively correlated with the proportion of negative posts, at r = −0.454 and r = −0.404 respectively, but their underlying mechanisms remain uncertain.

Manual review of 1,175 posts containing the derogatory terms “veneco” and “veneca” found increasingly negative usage over time. The negative proportion rose from an average of 54% to 82% for the masculine term and from 72% to 95% for the feminine term between 2015 and mid-2018, reaching peaks in 2018–2019. After 2019, the averages declined to 73% for the masculine term and 79% for the feminine term but remained above pre-2018 levels. Agreement was modest, with Cohen’s kappa of 0.48 for each term, so the estimates remain subject to coding uncertainty.

Topic modeling identified eight broad themes, including derogatory language, migration policy, border and national policy, xenophobia and cities, migration crisis, Maduro and security, and hunger and leaders. Relative prominence shifted from hostile language in the earlier years toward discussions of migration’s scale, visibility and policy responses around 2017–2018. The model had a coherence score of 0.327, which the authors described as modest for short social media posts, making its categories indicative thematic groupings rather than precise, mutually exclusive classes.

The findings describe the discourse of an active online community rather than the views of Colombia’s general population. X had an estimated 4.3 million Colombian users in early 2022, about 8.4% of the national population, and users tend to be younger and more educated than the broader public. The pipeline also introduced classification error at geolocation, relevance and sentiment stages, and the study did not formally model how that error propagated. Computational methods may not capture irony or sarcasm, classifier performance was not assessed year by year, and bots or coordinated inauthentic activity were not assessed. The results therefore characterize aggregated patterns in collected posts and reposts, not population attitudes or causal changes in social conditions.

Subject of Research: Computational analysis of X discourse about Venezuelan migration in Colombia from 2014 to 2022

Article Title: Multidimensional analysis of social discourse on Twitter (X) around Venezuelan migration in Colombia

Article References: Martínez, J., Miller-Felton, M., Padilla, J. J., Frydenlund, E., & Palacio, K. (2026). Multidimensional analysis of social discourse on Twitter (X) around Venezuelan migration in Colombia. PLOS One, 21(10), e0360050. https://doi.org/10.1371/journal.pone.0360050

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360050

Keywords: Venezuelan migration, Colombia, Twitter, X social media, sentiment analysis, geolocation inference, topic modeling, derogatory language, changepoint detection, Multidimensional, analysis, social