Characterizing Variation in Crowd-Sourced Data for Training Neural Language Generators to Produce Stylistically Varied Outputs

Juraj Juraska, Marilyn Walker


Abstract
One of the biggest challenges of end-to-end language generation from meaning representations in dialogue systems is making the outputs more natural and varied. Here we take a large corpus of 50K crowd-sourced utterances in the restaurant domain and develop text analysis methods that systematically characterize types of sentences in the training data. We then automatically label the training data to allow us to conduct two kinds of experiments with a neural generator. First, we test the effect of training the system with different stylistic partitions and quantify the effect of smaller, but more stylistically controlled training data. Second, we try a method of labeling the style variants during training, and show that we can modify the style of the output to some extent using our stylistic labels. We contrast and compare these methods that can be used with any existing large corpus, showing how they vary in terms of semantic quality and stylistic control.
Anthology ID:
W18-6554
Volume:
Proceedings of the 11th International Conference on Natural Language Generation
Month:
November
Year:
2018
Address:
Tilburg University, The Netherlands
Venues:
INLG | WS
SIG:
SIGGEN
Publisher:
Association for Computational Linguistics
Note:
Pages:
441–450
URL:
https://www.aclweb.org/anthology/W18-6554
DOI:
10.18653/v1/W18-6554
Bib Export formats:
BibTeX MODS XML EndNote
PDF:
https://www.aclweb.org/anthology/W18-6554.pdf