A Case Study on Context-Based Classification with Small Data Using SetFit

Witty Works was developing a writing assistant that detects non-inclusive language and faced the problem of context-dependent word classification. The existing spaCy-based approach extracted and filtered words based on their linguistic features, but it had a limitation in that it could not distinguish words whose meaning changes depending on context (e.g., 'fossil fuels' vs 'old fossil').

Through guidance from Hugging Face experts, the following technical transition was achieved:

• Adoption of Sentence Transformers: Instead of word embeddings, sentence embeddings that reflect the meaning of the entire sentence were used to improve context comprehension.
• Utilization of SetFit: To reduce the cost of large-scale data annotation, the SetFit library, optimized for few-shot fine-tuning, was adopted.

Through this, Witty Works was able to efficiently build a classification model that accurately grasps context with only a small amount of training data.