Witty Works, a startup focused on promoting inclusive language in the workplace, turned to Hugging Face's Expert Acceleration Program to improve its AI writing assistant. The tool, which started as a web app for inclusive job ads and evolved into a browser extension, faced a major challenge: context-dependent non-inclusive words that a rule-based approach couldn't handle.
Initially, Witty Works used a combination of linguistic analysis and a knowledge base of 2,300 non-inclusive words and idioms. While this worked for 85% of cases, it failed for words that change meaning based on context—like "fossil" in "old fossil" versus "fossil fuels." To tackle this, they needed a context-aware classifier.
Hugging Face experts guided them away from vanilla transformers (which required large labeled datasets) toward Sentence Transformers and the SetFit library. Sentence Transformers generate embeddings for entire sentences, capturing semantic meaning, while SetFit enables few-shot fine-tuning with as few as 15–20 labeled sentences per word. This approach saved time and money by avoiding the need for massive annotated datasets.
"We generate contextualized embedding vectors for every word depending on its sentence, then keep only the embedding for the 'problem' word's token and calculate cosine similarity," explained Elena Nazarenko, Lead Data Scientist at Witty Works. The resulting sentence embeddings feed into a classic classifier (KNN or logistic regression) for final classification.
The solution not only improved accuracy but also aligned with Witty Works' commitment to actively managing bias—keeping manual review manageable with small training sets. Today, the assistant offers real-time suggestions across English, French, and German, helping users write more inclusively in emails, job ads, and social media posts.