In this paper, we apply a method to quantify biases associated with named entities from various countries. We create counterfactual examples with small perturbations on target-domain data instead of relying on templates or specific datasets for bias detection. On widely used classifiers for subjectivity analysis, including sentiment, emotion, hate speech, and offensive text using Twitter data, our results demonstrate positive biases related to the language spoken in a country across all classifiers studied. Notably, the presence of certain country names in a sentence can strongly influence predictions, up to a 23% change in hate speech detection and up to a 60% change in the prediction of negative emotions such as anger. We hypothesize that these biases stem from the training data of pre-trained language models (PLMs) and find correlations between affect predictions and PLMs likelihood in English and unknown languages like Basque and Maori, revealing distinct patterns with exacerbate correlations. Further, we followed these correlations in-between counterfactual examples from a same sentence to remove the syntactical component, uncovering interesting results suggesting the impact of the pre-training data was more important for English-speaking-country names.

Classical bias detection methods used in Machine Learning are themselves biased because of the different confounding variables implied in the assessment of the initial biases. First they are using templates that are syntactically simple and distant from the target data on which the model will deployed. Second, current methods are assessing biases in pre-trained language models or in dataset, but not directly on the fine-tuned classifier that can actually produce harms. We propose a simple method to detect the biases of a specific fine-tuned classifier on any type of unlabeled data. The idea is to study the classifier behavior by creating counterfactual examples directly on the target data distribution and quantify the amount of changes. In this work, we focus on named entity perturbations by applying a Named Entity Recognition on target-domain data and modifying them accordingly to most common names or location of a target group (gender and country), and this for several morphosynctactically different languages spoken in relation with the countries of the target groups. We used our method on two models available open-source that are likely to be deployed by industry, and on two tasks and domains. We first assess the bias of a multilingual sentiment analysis model trained over multiple-languages tweets and available open-source, and then a multilingual stance recognition model trained over several languages and assessed over English language. Finally we propose to link the perplexity of each example with the bias of the model, by looking at the change in label distribution with respect to the language of the target group. Our work offers a fine-grained analysis of the interactions between names and languages, revealing significant biases in multilingual models.