Deep neural networks are hegemonic approaches to many machine learning areas, including natural language processing (NLP). Thanks to the availability of large corpora collections and the capability of deep architectures to shape internal language mechanisms in self-supervised learning processes (also known as “pre-training”), versatile and performing models are released continuously for every new network design. These networks, somehow, learn a probability distribution of words and relations across the training collection used, inheriting the potential flaws, inconsistencies and biases contained in such a collection. As pre-trained models have been found to be very useful approaches to transfer learning, dealing with bias has become a relevant issue in this new scenario. We introduce bias in a formal way and explore how it has been treated in several networks, in terms of detection and correction. In addition, available resources are identified and a strategy to deal with bias in deep NLP is proposed.

Deep neural networks are hegemonic approaches to many machine learning areas, including natural language processing (NLP). Thanks to the availability of large corpora collections and the capability of deep architectures to shape internal language mechanisms in self-supervised learning processes (also known as "pre-training"), versatile and performing models are released continuously for every new network design. But these networks, somehow, learn a probability distribution of words and relations across the training collection used, inheriting the potential flaws, inconsistencies and biases contained in such a collection. As pre-trained models have found to be very useful approaches to transfer learning, dealing with bias has become a relevant issue in this new scenario. We introduce bias in a formal way and explore how it has been treated in several networks, in terms of detection and correction. Also, available resources are identified and a strategy to deal with bias in deep NLP is proposed.

The study of bias in language models is a growing area of work, however, both research and resources are focused on English. In this paper, we make a first approach focusing on gender bias in some freely available Spanish language models trained using popular deep neural networks, like BERT or RoBERTa. Some of these models are known for achieving state-of-the-art results on downstream tasks. These promising results have promoted such models’ integration in many real-world applications and production environments, which could be detrimental to people affected for those systems. This work proposes an evaluation framework to identify gender bias in masked language models, with explainability in mind to ease the interpretation of the evaluation results. We have evaluated 20 different models for Spanish, including some of the most popular pretrained ones in the research community. Our findings state that varying levels of gender bias are present across these models.This approach compares the adjectives proposed by the model for a set of templates. We classify the given adjectives into understandable categories and compute two new metrics from model predictions, one based on the internal state (probability) and the other one on the external state (rank). Those metrics are used to reveal biased models according to the given categories and quantify the degree of bias of the models under study.

The study of bias in language models is a growing area of work, however, both research and resources are focused on English. In this paper, we make a first approach focusing on gender bias in some freely available Spanish language models trained using popular deep neural networks, like BERT or RoBERTa. Some of these models are known for achieving state-of-the-art results on downstream tasks. These promising results have promoted such models' integration in many real-world applications and production environments, which could be detrimental to people affected for those systems. This work proposes an evaluation framework to identify gender bias in masked language models, with explainability in mind to ease the interpretation of the evaluation results. We have evaluated 20 different models for Spanish, including some of the most popular pretrained ones in the research community. Our findings state that varying levels of gender bias are present across these models. This approach compares the adjectives proposed by the model for a set of templates. We classify the given adjectives into understandable categories and compute two new metrics from model predictions, one based on the internal state (probability) and the other one on the external state (rank). Those metrics are used to reveal biased models according to the given categories and quantify the degree of bias of the models under study.