EAGER: Automatic Story Generation in Support of Early Vocabulary Learning

In child development, small early differences can compound into big long-term effects. One example of this is the relationship between early vocabulary size, literacy, and later academic achievement. With this relationship in mind, many vocabulary enrichment programs based on shared reading with a caregiver have been developed, with mixed success.

Evidence suggests that individualizing target-vocabulary selection can improve learning, but manually generating stories that include personalized target words for every child is infeasible. Automatic story generation using natural language processing techniques has the potential to solve this problem. Although there has been some progress in automatic story generation for adults, this is an unsolved and particularly challenging problem when stories are targeted for preschoolers, because both content and complexity need to be tailored to the age group.

Thus, the researchers explore multiple innovative machine learning methods to generate engaging, high-quality child-directed stories that contain specific words that will enrich a child’s vocabulary. Furthermore, preschoolers and their caregivers participate in story-sharing activities to investigate if the automatically generated stories are effective tools for teaching words to children.

This research is particularly critical for low-income families and dual language learners, who are more likely to exhibit vocabulary delays while, at the same time, being less likely to receive intervention support.

This EArly Grant for Exploratory Research makes novel and potentially transformative contributions to the area of automatic story generation by taking necessary exploratory steps towards flexible, adaptive technology that can automatically generate personalized, engaging, and effective stories for toddlers and their caregivers to share at home as a vehicle for early vocabulary enrichment. Specifically, the first part of this project consists of the following: 1) an investigation of multiple computational models with regards to their suitability for preschooler-directed story generation; 2) a study of strategies to avoid the generation of content that is not suitable for children by machine learning-based story generation models; and 3) an exploration of how to automatically incorporate a set of predefined target words into generated stories.

Furthermore, the team of researchers investigates the quality of story generation models and the stories' effectiveness for word learning via the following: 4) obtaining feedback from families in the local community as to whether the automatically generated stories are appropriate and engaging for preschoolers and 5) conducting a laboratory study in which stories will be shared by caregivers and their children in a setting that resembles a natural home environment and subsequently comparing the children’s knowledge of target words against control words.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.