Story Generation using Hierarchical Convolutional Networks
Open Access
Author:
Naphade, Saniya
Graduate Program:
Computer Science and Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
October 15, 2021
Committee Members:
Chitaranjan Das, Program Head/Chair C Lee Giles, Thesis Advisor/Co-Advisor Rui Zhang, Committee Member Ting-Hao Huang, Thesis Advisor/Co-Advisor
Keywords:
story generation convolutional networks hierarchical text generation fusion mechanism
Abstract:
Many writers suffer from writer’s block when composing long stories with no fixed direction. A creative artificial intelligence system that can generate well reasoned and eloquent texts with limited inputs can provide valuable insights. This type of system can help writers to explore various directions in which the story could progress. However, the automated story generation task is usually experimented with toy datasets comprising of disconnected topics. These toy datasets usually contain artificial instances and topics (usually short sentences) making it easier for the model to learn and generate coherent passages. This work explores whether hierarchical story generation model can construct fluent follow-up passages depending on the instance from a real human story as input. All the instances are generated using the real-stories from the BookCorpus dataset. The instances are constructed by using extractive summarization on portions of the real story before being given as input Human readers have ranked the generated passages higher than the ground truth passages.
