Automatic Text Evaluation through the Lens of Wasserstein Barycenters
This work addresses the problem of automatic text evaluation for researchers and practitioners in NLP, offering an incremental improvement over existing metrics like MoverScore and BertScore.
The paper introduces BaryScore, a new metric for evaluating text generation using deep contextualized embeddings, which outperforms other BERT-based metrics and shows more consistent behavior, particularly in text summarization tasks.
A new metric \texttt{BaryScore} to evaluate text generation based on deep contextualized embeddings e.g., BERT, Roberta, ELMo) is introduced. This metric is motivated by a new framework relying on optimal transport tools, i.e., Wasserstein distance and barycenter. By modelling the layer output of deep contextualized embeddings as a probability distribution rather than by a vector embedding; this framework provides a natural way to aggregate the different outputs through the Wasserstein space topology. In addition, it provides theoretical grounds to our metric and offers an alternative to available solutions e.g., MoverScore and BertScore). Numerical evaluation is performed on four different tasks: machine translation, summarization, data2text generation and image captioning. Our results show that \texttt{BaryScore} outperforms other BERT based metrics and exhibits more consistent behaviour in particular for text summarization.