Bias in language: Is gender bias a factor in industrial speech processing?

You are a senior scientist in the field of machine speech processing. What exactly are you working on?

I’m responsible for data acquisition and management. My team’s job is to make sure that the AI modules for the voice assistants have high-quality data to learn from. Our work typically revolves around voice assistants for industrial use cases in German.

It’s crucial that the voice assistants process user requests correctly. And to do that, they need what’s known as training data to show them what sentences, commands or questions to expect. The performance of voice assistance modules, machine translation modules, or natural language processing modules in general can only be as good as the training examples they were trained on.

How did you get into the topic of gender bias in natural language processing?

Gender bias in natural language processing has been an issue for several years now. It’s not just about differences between women and men but about diversity in general. I personally think diversity is incredibly important and that it needs to be reflected in the environment we work in every day. After all, if you’re only ever exposed to one perspective, you may not even realize to what extent a voice assistant is biased.

Is gender bias even relevant in industrial natural language processing?

As we know, language is a product of culture. If we collect huge volumes of data from, say, online forums or social media, and use it for training a natural language processing model without filtering it first, any stereotypes contained in the data will be transferred to the models trained on that data. That means natural language processing might end up exhibiting, say, gender bias.

That said, industrial use cases tend to be more about objects, processes or customer roles. Our challenge is to develop high-quality voice assistant technologies even for use cases where not much data is available. Gender bias emerges when you work with vast amounts of data collected in an automated and uncontrolled way – and typically when the data doesn’t relate to a specific domain. But when the use cases are more specific, it’s not that common as you have a more direct control over the data.

In our work, it’s less about gender bias and more about avoiding unwanted bias in general. Although we haven’t yet come across gender bias in the industrial use cases we deal with, we do look carefully at the data quality – and you can apply that principle to any form of bias. 

Is there awareness of this topic at Fraunhofer IIS?

Yes. One aspect of that is an effort in raising awareness of gender bias. There are communication guidelines on gender-neutral language use and bias awareness is part of the Fraunhofer culture. This awareness can help identify problem cases in advance, recognize them more easily in the future, and ultimately avoid them altogether. 

Generally speaking, are people who work in speech processing already thinking about this topic? 

In the natural language processing community there is more and more focus on the fact that the data used to train the AI components for our voice assistants needs to be gender-neutral. So in that respect, yes, the researchers are already very conscious of the issue and have strategies to deal with it. We’re also aware that there are studies advocating the use of a neutral voice in speech processing. Ultimately, however, it’s up to our customers to decide whether they want to have a female voice, a male voice, both or something neutral. We supply technology components based on customer requirements, not end products.

Personally, I am passionate about this topic. At the Fraunhofer IIS Girls’ Day, where we seek to widen diversity in STEM subjects, I gave a presentation on gender bias. It’s important to learn about these things early on. Bias reinforces prejudice, and I believe that everyone should be able to build their lives based on their own interests and abilities – and not be held back by clichés.