Sentence Detailing and Its Applications

Abstract

The purpose of this study is to observe whether there is a performance improvement when we train the multilingual Text-to-Text Transfer Transformer (mT5), which is a transformer model in Natural Language Processing (NLP), with a dataset having sentences constructed using a set of given words as a pre-process before we train the model with a text-from-title dataset directly. Given words were considered as concept-set and the model was expected to learn the concept like commonsense knowledge from news to generate appropriate sentences after being trained with title-to-text dataset as well. We named this method “Sentence Detailing” due to its feature of generating sentences by adding details to a set of words. In addition to the text generation from title, we also examined this method under the topic of data augmentation.