Suggested Approaches to Mediate Learning Pathways

Approach with Navigator

Route: In this work, we propose a model to automatically generate learning pathways from available open learning resources, such that the generated pathways are semantically coherent and pedagogically progressive. The proposed model has two components – a Greedy Generator and a Validator based on Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) models respectively. The Greedy Generator chooses the next resource in the learning pathway based on local considerations and the Validator validates the learning pathway as a whole. They work in tandem with each other connected by a feedback loop.

Reinforcement Learning Suggest (Reroute):
Based on student interactions and performance on resources, the Learning Navigator makes real-time suggestions for students. The “suggest” algorithm is a multivariate optimization technique based on “search ranking.” Vectors are used to personalize the recommendations using principles of learning. Event, Condition, Learning Principle, Action (ECpA) rules will be created to aid in operationalizing each learning principle. This process will be used to create both detailed user journey maps and inform the personalized learning pathways. Navigator currently utilizes five different learning principles to guide delivery and sequencing of resources: (1) engage learners to build from prior knowledge to construct new knowledge, (2) maximize a student’s zone of proximal development, (3) continually revisit a concept to foster more complex and abstract knowledge development, (4) provide frequent embedded assessments to better estimate the learner’s skills and comprehension, and (5) offer choice to the learner to foster engagement. The goal is to continually operationalize more learning principles to implement in the study logic on Navigator.
This algorithm offers a weighted list of resources based on the set of weights obtained from the ECPA rules and the set of weights for each rule, maintained for each learner

RLUpdate:
This algorithm will be called after a learner responds to the suggestions made. A learner can do one of three things with a suggestion: click on the suggestion, ignore the suggestion, cancel the suggestion which results in reinforcement signals to update the weights for the suggestions

Conversational Reroute: In this work we propose a model of conversational learning navigator which works with the learner to make learning interactive, interesting and personally meaningful to the learner. The conversational navigator is like a companion to the learner; it replies to the questions of the learner regarding the subject; it understands the learner’s knowledge of concepts and re-routes the learner when needed. The conversational agent uses Natural Language Understanding and Generation to create overviews of the pathway, answer learner queries, re-route the learner and to automatically generate questions.

Trailer Generation: Trailers usually provide the viewer with an eagle’s eye view of any Narrative. With the enormous growth rate in user-generated academic resources such as videos, audios, texts, courses, efficient navigation of these resources has become very important and it will be useful to have academic trailers of such resources to get a quick sense of the narratives. Leveraging capacity of machine learning and automation we can have intelligent systems to generate academic trailers with minimum human interference. We propose generating trailers of academic resources by using text processing, NLP and machine learning techniques.

Search Ranking:
Use multivariate optimization techniques to personalize search by ranking content not only on content metadata but also based on the principles of learning & the understanding of the learner.

AI Techniques and Tools

• Route
  • Word embeddings -Word2Vec
  • Document embeddings – Paragraph2Vec
  • LDA (Latent Dirichlet Allocation)
  • SVM (Support Vector Machine)
  • LSTM (Long Short-Term Memory)
    
• Reroute
  • Rule based algorithms / Semantic AI techniques
    
• Reinforcement
  • Markov Decision Process
  • Reward based Reinforcement learning
    
• Conversational Reroute
  • Open AI’s GPT-2
  • Word embeddings – Fasttext and word2vec
  • Rasa framework for developing AI powered chatbots
    
• Trailer Generation
  • GPT-2 – Text Generation Model
  • IBM Watson – Text-to-Speech Engine
  • Word embeddings – Word2Vec (semantic representation of word)
  • GRU – Gated Recurrent Unit (for modeling sequential data)
    
• Search
  • ELS based Ecosystem
  • Navigator Feature Set for Ranking (taxonomy mapping, content quality, usage pattern, user profile)