dual_encoder_udc

An implementation of dual encoder using Keras (more details are on my blog post https://basmaboussaha.wordpress.com/2017/10/18/implementation-of-dual-encoder-using-keras/).

In this repository I implemented the dual encoder used in this paper https://arxiv.org/pdf/1506.08909.pdf using Keras. The motivation of this work is that the available implementations of this work are in Theano https://github.com/npow/ubottu and Tensorflow https://github.com/dennybritz/chatbot-retrieval/

I think that both implementations are hard to understand and to re-use (my point of view) so I decided to implement a simple code using Keras with Theano in backend.

Requires:

• Python 3.5
• Theano 0.9.0
• Keras 2.0.8

Before running the model:

• Creat three folders in the local directory (dual_encoder_udc): dataset, model and embeddings.
• Download Glove Embeddings into embeddings directory

  cd embeddings
  wget http://nlp.stanford.edu/data/glove.840B.300d.zip
  unzip glove.840B.300d.zip
  

• Download the dataset (train, test and dev):
  • You can download it directly in format .pkl into dataset folder: These pkl files were generated using the utilities/prepare_dataset_from_csv.py script which separate the context and response from each of the train.csv, test.csv and dev.csv

  https://drive.google.com/file/d/1VjVzY0MqKj0b-q_KXnaHC49qCw3iDqEY/view?usp=sharing
  

  • Or you can download raw files in csv format from here https://github.com/rkadlec/ubuntu-ranking-dataset-creator and run the utilities/prepare_dataset_from_csv.py script by yourself.

  cd utilities
  python prepare_dataset_from_csv.py
  

Running the model:

python3.5 dual_encoder.py

Results:

Running The above python script should give you the results below. Note that the best results are obtained after epoch 2, using the default hyperparameters: --emb_dim=300, --hidden_size=300, --batch_size=256. Note that embeddings are intilized with glove and are fine tuned during training.

Perform on test set after Epoch: 2...!

group_size: 2
recall@1 0.898414376321353
group_size: 5
recall@1 0.7420190274841437
recall@2 0.898784355179704
group_size: 10
recall@1 0.6213002114164905
recall@2 0.7822410147991543
recall@5 0.948784355179704