vinhkhuc/MemN2N-babi-python

Implements the MemN2N architecture with multi-hop attention over memory slots to reason over supporting facts in narrative contexts, using soft attention mechanisms to retrieve and aggregate relevant information for answer prediction. Supports single-task and joint multi-task training across the 20 bAbI diagnostic tasks, with both command-line execution and Flask-based web demo for interactive question-answering. Built in pure NumPy following Facebook's original MATLAB implementation, enabling direct model inspection and research experimentation without framework dependencies.

576 stars. No commits in the last 6 months.