Mark-Evaluate: Assessing Language Generation using Population Estimation Methods
This work addresses the challenge of assessing quality and diversity in language generation for NLP researchers, offering a novel approach that improves evaluation accuracy.
The authors tackled the problem of evaluating language generation by proposing a family of metrics derived from population estimation methods in ecology, such as mark-recapture and maximum-likelihood techniques, which showed higher correlation to human evaluation than existing metrics on tasks like unconditional language generation, machine translation, and text summarization.
We propose a family of metrics to assess language generation derived from population estimation methods widely used in ecology. More specifically, we use mark-recapture and maximum-likelihood methods that have been applied over the past several decades to estimate the size of closed populations in the wild. We propose three novel metrics: ME$_\text{Petersen}$ and ME$_\text{CAPTURE}$, which retrieve a single-valued assessment, and ME$_\text{Schnabel}$ which returns a double-valued metric to assess the evaluation set in terms of quality and diversity, separately. In synthetic experiments, our family of methods is sensitive to drops in quality and diversity. Moreover, our methods show a higher correlation to human evaluation than existing metrics on several challenging tasks, namely unconditional language generation, machine translation, and text summarization.