Identifying Latent Traits of Questions for Controllable Machine Generation
PROCEEDING
Alexander Maas, Taku Kawada, Kazunori Yamada, Toru Nagahama, Tatsuya Horita, Tohoku University, Graduate School of Information Sciences, Japan
EdMedia + Innovate Learning, in Online ISBN 978-1-939797-65-0 Publisher: Association for the Advancement of Computing in Education (AACE), Waynesville, NC
Abstract
In the Natural Language Processing domain, original and controllable text generation techniques are producing creative, original, and realistic results (Keskar et al. 2017). However, there has been limited focus on applying these controlled generation techniques to English language education. This is partly due to the availability of good quality, labelled data, since manually labelling data is costly, time consuming and error prone. To address this issue, this paper proposes to cluster and label questions using latent variables found via a transformer-based neural network for use in future, controlled generation tasks. By labelling questions this way, the time and cost of preparing the data is reduced and the human error component is eliminated.
Citation
Maas, A., Kawada, T., Yamada, K., Nagahama, T. & Horita, T. (2022). Identifying Latent Traits of Questions for Controllable Machine Generation. In T. Bastiaens (Ed.), Proceedings of EdMedia + Innovate Learning (pp. 42-47). Online: Association for the Advancement of Computing in Education (AACE). Retrieved December 6, 2023 from https://www.learntechlib.org/primary/p/221267/.
© 2022 Association for the Advancement of Computing in Education (AACE)
