On-the-Fly Rectification for Robust Large-Vocabulary Topic Inference