[4Xin1] Poster session 2

[4Xin1-71] Disentangling Style and Content in Text Generation from Data

〇Yumi Hamazono^1,2, Yui Uehara², Tatsuya Ishigaki², Yusuke Miyao^3,2, Hiroya Takamura², Ichiro Kobayashi^1,2 (1.Ochanomizu University, 2.National Institute of Advanced Industrial Science and Technology, 3.The University of Tokyo)

Keywords:Natural language processing, Text Generation, Data-to-text

In recent studies, the task that automatically generates text describing data, called data-to-text, has shown high accuracy by using end-to-end learning.
Some data-to-text studies especially using real-world data and text show that it may not be possible to predict the target output text only from the input data for a dataset constructed.
When using datasets containing such unpredictable attributes, it has been found that the unpredictable attributes are fed as input, thereby improving the accuracy of text generation and more correctly describing the content of the data.
In this study, we apply a method of disentangling the latent representations of style and content in language models to data-to-text and verify sentence generation using input data and stylistic representations obtained from the sentences.
Furthermore, by classifying the style representations obtained from the sentences, we verify a method for extracting the attributes of the sentences that are not predicted from the input data.