Paraphrase generation and evaluation - a view from the trenches

Wiktor Franus, Bartłomiej Twardowski, Paweł Zawistowski

In this paper we evaluate the current state of the art in natural language paraphrase generation using deep learning methods. The focus is put on the entire modeling pipeline from data gathering up to model evaluation. Specifically, we list the publicly available datasets suitable for this task, assess their quality and discuss procedures connected with data preparation and model training. Finally, we discuss problems related to the currently used evaluation approaches.

Keywords:
natural language processing, paraphrase generation, deep learning