AI-Enabled Education Through Machine Learning and Crowdsourcing
Abstract
The ongoing artificial intelligence (AI) revolution is providing the opportunity to fundamentally change the way education and training are conducted. It is now possible to create so-called intelligent textbooks (ITBs) that provide courseware functionality (text, videos, simulations, etc.), learning analytics, personalized tutoring, and more. At the heart of an ITB is a knowledge graph (KG) that captures the relationships among the concepts in the domain being taught. Similar to an ontology, a KG is a complex and robust structure that enables AI algorithms to provide ITBs with functionality such as textsummarization, question generation, question answering, and more. However, the potential of ITBs to facilitate better learning has been extremely difficult to realize without major investments of time, money, and expertise. Indeed, it is the construction of the KG itself that presents the biggest barrier towards developing extensive libraries of ITBs. The reason is that KGs are currently constructed using human subject experts, and this process is extremelyexpensive and time consuming. This project will develop new approaches to scalably develop KGs so that the dream of ITBs can be realized. Our research will proceed in four interrelated research thrusts. Thrust 1: Automatic Knowledge Graph Construction using Machine Learning. We will develop new machine learning algorithms (ML) that enable automatic KG construction. In the language of graph theory, we will develop term extraction algorithms that identify the nodes of the graph and relationship extraction algorithms that identify the edges of the graph. Thrust 2: Human-Driven Knowledge Graph Construction using Crowdsourcing. To close this performance gap between automatic (ML) and human-expert-generated KGs, we investigate crowdsourcing KG construction using students and course instructors who interact directly with underlying course material. Thrust 3: Fusing ML and Crowdsourced Knowledge Graphs. The task of fusing the ML and crowdsourced KGs is complicated by the facts that 1) each participant KG will represent only a small portion of the full KG, and 2) the participant KGs will contain errors, including omitting required relationships, proposing relationships that do not exist, and mislabeling other relationships. Therefore, we will develop new graph denoising techniques that intelligently average over several crowdsourced KGs to create an output KG with fewer errors. Thrust 4: New Education Applications Enabled by Knowledge Graphs.Armed with an efficiently generated KG for a textbook/course, we will develop new algorithms that transform flat text content into an intelligent textbook that aids student learning. Particular features we will explore include automatic question generation (to support student practice), active dialog with the text (answering questions posed by students), and monitoring of student progress in a domain (learning analytics).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012534
Entities
People
- Richard G. Baraniuk
Organizations
- Office of Naval Research
- Rice University
- United States Navy