Detecting Anti-Immigrant Bias of U.S. Politicians by Region via Word Embeddings

Abstract

While the population of immigrants in the United States of America has grown over the years, so too as anti-immigrant sentiment. It is important, then, that we understand the areas in which immigrants face the most stereotypes, as well as the ways in which such stereotypes are expressed. Specifically, this work aims to study the anti-immigrant bias present among politicians in four regions of the United States through the use of word embeddings trained on Tweets that they have posted. Word embeddings have been proven to accurately reflect relationships between words based on the context in which they appear in the text, and as a result, reflect the bias of the authors of text as well. Therefore, we quantify anti-immigrant bias using two measures of bias in word embeddings proposed by previous work and explore whether human stereotypes towards immigrants in relation to occupations, crimes, countries, and general sentiment are present in the word embeddings. We find that as expected, the measures of bias computed using the word embedding models trained on the Tweet text generally match the stereotypes that are present in society, and that such bias appears across all four regions of the United States.