Different Results from Identical Models: A Deep Dive into Pre-trained Word Embeddings and Keras Architectures