MemN2N-tensorflow

Why we included this project

This is a TensorFlow port of the End-To-End Memory Network from the Sukhbaatar et al. paper, focused on the language modeling setup described in Section 5. The code follows the paper's multi-hop attention design closely, and most architectural choices, from memory size to the number of hops, are exposed as command-line flags, so it reads well as a study aid. You can train it on the bundled Penn Tree Bank corpus and check your perplexity against the published results. It is not a production NLP library, so treat it as a learning tool rather than something to wire into a live system. The whole thing is compact enough to read through in one sitting, which helps if you want to see how memory networks differ from ordinary recurrent language models.