Salesforce proposes new ways to reduce AI gender bias

Working with researchers at the University of Virginia, Salesforce has come up with new ways to help reduce aI gender bias. Typically, researchers need to feed a number of ai models for training, but inevitably docused with some implicit, or explicit, gender bias. These AIs can then become susceptible to bad habits while performing language translation or other predictive tasks.

Salesforce proposes new ways to reduce AI gender bias

Double hard de-bias gesture

With this in mind, the team tried to correct certain regularities, such as the frequency of words in the big data set, so that AI “purified” embedded content before reasoning, discarding sexist words.

The scheme, which captures the semantics, syntax, and relationships between words, has been used by many natural language processing (NLP) schemes but has been criticised for the inevitable gender bias.

Previous remedies were introduced in the reprocessing process to remove gender discrimination-related elements, but their effectiveness was greatly limited, such as being restored after de-biasing operations.

Salesforce proposes new ways to reduce AI gender bias

Double hard de-biaser benchmark test results

To that end, Salesforce has proposed a new solution called Double-Hard Debias to transform embedded spaces into seemingly gender-free subspaces.

Then, before performing another anti-bias operation, it “projects” the gender component along this dimension to get the modified embedded content. To assess the results, the researchers tested the WinoBias dataset.

The data set consists of sentences that favour gender stereotypes and oppose gender stereotypes, and the performance gap reflects how the algorithmic system executes on both sentence groups and leads to “gender bias” scores.

Salesforce proposes new ways to reduce AI gender bias

tSNE Embedding Projection

The results show that while retaining semantic information, the two-hard de-bias scheme uses the GloVe algorithm to obtain an embedded deviation score, cut from 15 to 7.7 points. It is also more evenly mixed than other scenarios on visualizations (tSNE projections) that model the embedding.

Even so, some experts argue that it is impossible to completely eliminate bias from word embedding. A recent study by the Technical University of Munich, for example, showed that there is “no natural neutral text”. Because the semantic content of words is always associated with the socio-political environment.