Add Client Gender with airSlate SignNow

Add client gender

Thanks to breakthroughs in artificial intelligence and natural language processing, computers are able to infer more and more about the world we live in. Like Messi's position. Or the company Steve Ballmer ran. But some parts of our world are better left out. (buzzer ringing) Both researchers and everyday internet users have begun to notice the very human gender bias that snuck it's way into AI systems, which can actually exhibit their own form of bias. And when human bias joins forces with machine learning bias, it can grow exponentially. To better understand how this happens, we need to talk about word embeddings. Word embeddings are a list of numbers that encode information about a word's meaning and usage. We can compare these values to determine how related two or more terms are. But if language models are trained on data that makes a lot of assumptions about the roles of men and women, they can start to get the wrong idea and produce pure MT translations that project gender, even when the original language doesn't specify one. Like this English translation of a gender neutral text in Hungarian. Machine learning systems are what they eat and sometimes, we end up feeding them sexist garbage. Some people might ask whether it's AI's role to describe or prescribe human behavior. But human bias creates AI systems that are straight up inaccurate. A man who writes code isn't analogous to a female homemaker, he's analogous to a woman who writes code. The good news is that a number of AI scientists are working to root out this distortion. In translation, one way to do this is by augmenting training data so that for every gender sentence, an additional sentence is generated, replacing pronouns and gendered words with the opposite genders and blanking out names. This way, the training data becomes gender balanced and doesn't learn gender characteristics associated with certain names. Another way to reduce gender-accurate translations is to allow users to add metadata to text that stores information about genders. The road to unbiased machine learning is a long one because as a society, we're still a long way from equality. But whether we're talking about data or people, the key to overcoming bias is the same, thoughtful training.