Text Embellishment using Attention Based Encoder-Decoder Model

Text embellishment is a natural language generation problem that aims to enhance the lexical and syntactic complexity of a text. i.e., for a given sentence, the goal is to generate a sentence that is lexically and syntactically complex while retaining the same semantic information and meaning. In contrast to text simplification (Wang et al., 2016), text embellishment is considered to be a more complex problem as it requires linguistic expertise, and therefore are difficult to be shared across different platforms and domain. In this paper, we have explored this problem through the light of neural machine translation and text simplification. Instead of using a standard sequential encoder-decoder network, we propose to improve text embellishment with the Transformer model. The proposed model yields superior performance in terms of lexical and syntactic embellishment and demonstrates broad applicability and effectiveness. We also introduce a language and domain agnostic evaluation set up specifically for the task of embellishment that can be used to test different embellishment algorithms.