Analysing social biases toward migrant groups encoded in language models

Embedding models are powerful machine-learning-based representations of human language used in a myriad of Natural Language Processing tasks. Due to their ability to learn underlying word association patterns present in large volumes of data, it is possible to observe various sociolinguistic phenomena encoded in the distributional vector spaces, among them, social stereotypes. Even if such models must be carefully tested for social biases and not blindly employed in downstream applications due to ethically concerning outcomes, they can be useful for discourse analysis of large volumes of textual data, for instance. In this thesis, we explore the use of language models to analyze and quantify biases towards migrant groups. We start by conducting a monolingual diachronic study of articles published in the Spanish newspaper 20 Minutos between 2007 and 2018. Then, we analyze the Danish, Dutch, English, and Spanish portions of four different multilingual corpora of political discourse, covering the 1997-2018 period. For both the aforementioned studies, we examined the effect of sociopolitical variables such as unemployment and criminality numbers on our bias measurements using statistical models. Finally, we contribute to the creation of linguistic resources for investigating biases against migrants by releasing a multilingual dataset for the Catalan, Portuguese, and Spanish languages inspired by social surveys that measure perceptions and attitudes towards immigration in European countries.