Simulating Learning from Language and Examples

Abstract

Simulations of human learning can be used as computational models for evaluating theories of learning. They can also be taught interactively to author intelligent tutoring systems. Prior simulated learner systems have learned inductively from worked examples and correctness feedback. This work introduces a mechanism where simulated learners can also learn from natural language. Using a neural grammar parser with additional symbolic processing steps, we simulate the production of loose interpretations of verbal instructions. These interpretations can be combined with worked examples to resolve the ambiguities of either form of instruction alone. We find that our system has practical benefits over an alternative method using github Copilot and slightly better accuracy.