Scalable Conversational AI for Literacy Development
The overall objective of this project is to evaluate the feasibility of using automatic question-answer (QA) generation to facilitate the learning and assessment of narrative comprehension skills, a critical component of reading proficiency. This innovation can enable large-scale, cost-effective development and production of learning and assessment resources. There are three components:
- Creating a high-quality QA pair dataset for model training and evaluation (link to the dataset here)
- Developing AI models that expand the state-of-the-art deep neural network techniques (e.g., BERT) for machine comprehension tasks over the created dataset, and customizing the training to meet the unique requirements of an educational context;
- Building an interactive reading system with QA functionalities (i.e., the agent asking students questions and assessing their answers) to enhance and evaluate students’ comprehension.
Interactive Reading System Design
The system consists of five components: user book uploading, question-answer pair generation, dialogue generation, chatbot deployment, and student profiling.
- In Step 1, users will upload books, either individually or by batch, to a web-based interface or a mobile app.
- In Step 2, the question-answer generation module will take users’ book input and automatically generate a list of narrative comprehension questions and their associated answers, utilizing the dataset and the QA models we are currently working on. In addition, we will allow users to customize the questions based on their specific needs.
- In Step 3, the dialogue module will take the QA pairs and generate dialogue flows. Specifically, each dialogue moment will involve a conversational agent 1) asking children a question, 2) providing tailored feedback and explanation to children’s response, and/or 3) rephrasing the original question (usually open-ended) to a multiple-choice question as a way of scaffolding if the children do not answer the original question or answer it incorrectly.
- In Step 4, the chatbot building module will take the dialogue flow and create a chatbot that students will interact with. This chatbot can be deployed across multiple platforms, including smart speakers (e.g., Alexa, Google), mobile apps, or web-based systems. The agent will employ state-of-art learning analytic models that automatically assess children’s responses through multiple metrics, such as response length, vocabulary use, relevance, and accuracy.
- In Step 5, the system will record children’s performance during each book reading activity so as to track children’s development of comprehension skills over time. There will be a dashboard for parents or teachers to access this information.