A tool to overcome technical barriers for bias assessment in human language technologies

Abstract

Automatic processing of language is becoming pervasive in our lives, oftentaking central roles in our decision making, like choosing the wording for ourmessages and mails, translating our readings, or even having full conversationswith us. Word embeddings are a key component of modern natural languageprocessing systems. They provide a representation of words that has boosted theperformance of many applications, working as a semblance of meaning. Wordembeddings seem to capture a semblance of the meaning of words from raw text,but, at the same time, they also distill stereotypes and societal biases whichare subsequently relayed to the final applications. Such biases can bediscriminatory. It is very important to detect and mitigate those biases, toprevent discriminatory behaviors of automated processes, which can be much moreharmful than in the case of humans because their of their scale. There arecurrently many tools and techniques to detect and mitigate biases in wordembeddings, but they present many barriers for the engagement of people withouttechnical skills. As it happens, most of the experts in bias, either socialscientists or people with deep knowledge of the context where bias is harmful,do not have such skills, and they cannot engage in the processes of biasdetection because of the technical barriers. We have studied the barriers inexisting tools and have explored their possibilities and limitations withdifferent kinds of users. With this exploration, we propose to develop a toolthat is specially aimed to lower the technical barriers and provide theexploration power to address the requirements of experts, scientists and peoplein general who are willing to audit these technologies.