Synthetic data helps models only if it lands near real examples in a specific way.

Computer models often struggle to understand human language when certain meanings are rare. Researchers studied how words act in conversation by analyzing four different uses of the word look. Because training examples were scarce, they used an AI system to create fake, extra examples. Normally, people just focus on making lots of different fake examples. Instead, these researchers tested where these new examples landed mathematically compared to the real data. They kept the total number of fake examples the same and tested six setups. Adding synthetic examples helped the models sort the meanings better than using only real data. Examples placed very close to real data boosted a balance score called macro F the most. Meanwhile, a balanced mix of close and far examples gave the highest overall accuracy. However, fake data only moved the dividing line between categories rather than truly improving probability guesses.

Why it matters

This changes how researchers handle data scarcity in language tasks. Instead of just generating more fake data, they need to place it strategically to improve model performance, which could help with many low-resource classification problems.

Questions answered