An analysis of unconscious gender bias in academic texts by means of a decision algorithm

Inclusive language focuses on using the vocabulary to avoid exclusion or discrimination, specially referred to gender. The task of finding gender bias in written documents must be performed manually, and it is a time-consuming process. Consequently, studying the usage of non-inclusive language on a document, and the impact of different document properties (such as author gender, date of presentation, etc.) on how many non-inclusive instances are found, is quite difficult or even impossible for big datasets. This research analyzes the gender bias in academic texts by analyzing a study corpus of more than 12,000 million words obtained from more than one hundred thousand doctoral theses from Spanish universities. For this purpose, an automated algorithm was developed to evaluate the different characteristics of the document and look for interactions between age, year of publication, gender or the field of knowledge in which the doctoral thesis is framed. The algorithm identified information patterns using a CNN (convolutional neural network) by the creation of a vector representation of the sentences. The results showed evidence that there was a greater bias as the age of the authors increased, who were more likely to use non-inclusive terms; it was concluded that there is a greater awareness of inclusiveness in women than in men, and also that this awareness grows as the candidate is younger. The results showed evidence that the age of the authors increased discrimination, with men being more likely to use non-inclusive terms (up to an index of 23.12), showing that there is a greater awareness of inclusiveness in women than in men in all age ranges (with an average of 14.99), and also that this awareness grows as the candidate is younger (falling down to 13.07). In terms of field of knowledge, the humanities are the most biased (20.97), discarding the subgroup of Linguistics, which has the least bias at all levels (9.90), and the field of science and engineering, which also have the least influence (13.46). Those results support the assumption that the bias in academic texts (doctoral theses) is due to unconscious issues: otherwise, it would not depend on the field, age, gender, and would occur in any field in the same proportion. The innovation provided by this research lies mainly in the ability to detect, within a textual document in Spanish, whether the use of language can be considered non-inclusive, based on a CNN that has been trained in the context of the doctoral thesis. A significant number of documents have been used, using all accessible doctoral theses from Spanish universities of the last 40 years; this dataset is only manageable by data mining systems, so that the training allows identifying the terms within the context effectively and compiling them in a novel dictionary of non-inclusive terms.