CapsNet for Natural Language Processing
A capsule is a group of neurons whose activity vector represents the instantiation parameters of a specific type of entity such as an object or an object part.
This repository shows how to use a CapsNet architecture for Natural Language Prcoessing tasks like sentiment analysis.
Capsules are introduced by Geoffrey Hinton. We use a CapsNet implementation from 苏剑林 as git submodule. The implementation can be found here.
Here are some papers where capsules and the CapsNet architecture are introduced:
|Dynamic Routing Between Capsules||Sara Sabour, Nicholas Frosst, Geoffrey E Hinton||here|
|Matrix capsules with EM routing||Geoffrey E Hinton et al.||here|
|Transforming Auto-encoders||Geoffrey E. HintonAlex Krizhevsky, Sida D. Wang||here|
The CapsNet implementation is included via git submodule. So the first step after cloning this repository is to initialize the git submodules. This can be done via:
git submodule update --init --recursive
We use the IMDB dataset for sentiment analysis with CapsNet. We use a bidirectional GRU before the capsnet layer.
The training can be started with:
It takes several minutes per epoch. It is highly recommended to use a GPU for training. All experiments are done with a GTX 1060 (6GB).
The following experiments are done on IMDB dataset:
- Model a): we use a bidirectional GRU with a hidden size of 256. Number of capsule is set to 10. Number of routings is set to 3.
A recent version of Keras, TensorFlow and h5py is needed. Only Python 3.x is currently supported.
Contact (Bugs, Feedback, Contribution and more)
For questions about the capsnet-nlp repository, please open an issue here. If you want to contribute to the project please refer to the Contributing guide!
To respect the Free Software Movement and the enormous work of Dr. Richard Stallman
the software in this repository is released under the GNU Affero General Public License
in version 3. More information can be found here