Takumi Goto, Yusuke Sakai, Taro Watanabe
The paper introduces UOT-ERRANT, a new metric for evaluating grammatical error correction systems by optimally transporting edit vectors, showing improved performance and interpretability.
Evaluating systems that correct grammar errors in text is important to determine which ones work best. Traditional methods compare corrected sentences to reference sentences, but this doesn't always work well because many words don't change. This study introduces a new way to evaluate these systems by focusing on the changes made to the text, using a technique called unbalanced optimal transport. This method, called UOT-ERRANT, not only improves evaluation accuracy but also helps understand how the systems make corrections.
