Emotional distributions across Countries: highlighting biases in LLM

This thesis work aimed to understand whether the nationality of a person influences the answers of Large Language Models when they are asked to assign an emotion to an explicit situation. Furthermore, a study about the impact of differentiating the prompt and the analysis of the benefits of using country information in the emotion assignment task was conducted. In order to use a different strategy and improve the classification metrics, clustering techniques were performed. The final results show that country information is useful in helping the model assign an emotion and solve the indefiniteness problems when a generative model is used. Nonetheless, removing the country datum from the prompt, specifically for those countries for which the model struggles the most in assigning the correct emotion, improves the results. Regarding the prompt analysis, outcomes show that the biases depicted in the emotional distributions do not change; instead, the efficiency in assigning an emotion does it .