AI researchers seek method for removing gender bias in natural language

Technology Canada CU- As AI further advances to a variety of different fields and becomes more and more human like, researchers are attempting to make sure that AI does not pick all human traits.

A new study is focusing on effective methods of lowering gender bias in natural language processing models while keeping necessary details such as the definition of words.

In spite of computers showing no particular bias the researchers point out that the individuals designing or using the AI are likely to influence certain aspects of AI responses.

Bias to reflect a particular group or favor a certain point of view or a certain political entity has been part of a global debate to which free speech activists have raised concerns that the internet is becoming less and less open and a place certain point of view have been magnified and opposing points of views suppressed.

Lei Ding, 1^st author of the study and graduate student in the Department of Mathematical and Statistical Sciences stated the computer is unable to understand text and embedding where words are changed into numbers.

Bei Jiang, associate professor in the Department of Mathematical and Statistical Sciences stated that Natural language processing is making computers grasp texts and languages.

Researchers can then plot words as numerals on a 2D chart and visualize the words with comparisons to one another. This permits them to effectively gage the level of gender bias, that can be followed by an evaluation if the bias was removed.

The researchers however indicated that taking out the bias can lead to other important features being eliminated.

“In many gender debiasing methods, when they reduce the bias in a word vector, they also reduce or eliminate important information about the word,” said Jiang and further explained that these types of details are referred to as semantic information, and it offers significant contextual data that could be required in further tasks involving those word embeddings.

The researcher were looking at a new methodology to further optimize the results.