Description
This repository hosts the Hierarchical Encoder Decoder RNN model (HRED) and the Latent Variable Hierarchical Recurrent Encoder-Decoder RNN model (VHRED) for generative dialog modeling as described by Serban et al. (2016a) and Serban et al. (2016c).
Truncated BPTT
Both models are implemented using Truncated Backpropagation Through Time (Truncated BPTT). The truncated computation is carried out by splitting each document (dialogue) into shorter sequences (e.g. 80 tokens) and computing gradients for each sequence separately, such that the hidden state of the RNNs on each subsequence are initialized from the preceding sequences (i.e. the hidden states have been forward propagated through the previous states).
Creating Datasets
The script convert-text2dict.py can be used to generate model datasets based on text files with dialogues. It only requires that the document contains end-of-utterance tokens </s> which are used to construct the model graph, since the utterance encoder is only connected to the dialogue encoder at the end of each utterance.
Prepare your dataset as a text file for with one document per line (e.g. one dialogue per line). The documents are assumed to be tokenized. If you have validation and test sets, they must satisfy the same requirements.
Once you're ready, you can create the model dataset files by running:
python convert-text2dict.py <training_file> --cutoff <vocabulary_size> Training python convert-text2dict.py <validation_file> --dict=Training.dict.pkl Validation python convert-text2dict.py <test_file> --dict=Training.dict.pkl <vocabulary_size> Test
where <training_file>, <validation_file> and <test_file> are the training, validation and test files, and <vocabulary_size> is the number of tokens that you want to train on (all other tokens, but the most frequent <vocabulary_size> tokens, will be converted to <unk> symbols).
NOTE: The script automatically adds the following special tokens specific to movie scripts:
- end-of-utterance: </s>
- end-of-dialogue: </d>
- first speaker: <first_speaker>
- second speaker: <second_speaker>
- third speaker: <third_speaker>
- minor speaker: <minor_speaker>
- voice over: <voice_over>
- off screen: <off_screen>
- pause: <pause>
If these do not exist in your dataset, you can safely ignore these. The model will learn to assign approximately zero probability mass to them.
Model Training
If you have Theano with GPU installed (bleeding edge version), you can train the model as follows:
-
Clone the Github repository
-
Create a new "Output" and "Data" directories inside it.
-
Unpack your dataset files into "Data" directory.
-
Create a new prototype inside state.py (look at prototype_ubuntu_HRED for an example)
-
From the terminal, cd into the code directory and run:
THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python train.py --prototype <prototype_name> > Model_Output.txt
where <prototype_name> is a state (model architecture) defined inside state.py. Training a model to convergence on a modern GPU on the Ubuntu Dialogue Corpus with 46 million tokens takes about 1-2 weeks. If your GPU runs out of memory, you can adjust the bs (batch size) parameter in the model state, but training will be slower. You can also play around with the other parameters inside state.py.
(CURRENTLY NOT SUPPORTED) To test a model w.r.t. word perplexity run:
THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python evaluate.py <model_name> Model_Evaluation.txt
where <model_name> is the model name automatically generated during training.
Model Sampling & Testing
To generate model responses using beam search run:
THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=gpu python sample.py <model_name> <contexts> <model_outputs> --beam_search --n-samples=<beams> --ignore-unk --verbose
where <model_name> is the name automatically generated during training, <contexts> is a file containing the dialogue contexts with one dialogue per line, and <beams> is the size of the beam search. The results are saved in the file <model_outputs>.
To compute the embedding-based metrics on the generated responses run:
python Evaluation/embedding_metrics.py <ground_truth_responses> <model_outputs> <word_emb>
where <ground_truth_responses> is a file containing the ground truth responses, <model_outputs> is the file generated above and <word_emb> is the path to the binarized word embeddings. For the word embeddings, we recommend to use Word2Vec trained on the GoogleNews Corpus: https://drive.google.com/file/d/0B7XkCwpI5KDYNlNUTTlSS21pQmM.
Citation
If you build on this work, we'd really appreciate it if you could cite our papers:
A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues. Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron Courville, Yoshua Bengio. 2016. http://arxiv.org/abs/1605.06069
Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models. Iulian V. Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, Joelle Pineau. 2016. AAAI. http://arxiv.org/abs/1507.04808.
Datasets
The pre-processed Ubuntu Dialogue Corpus and model responses used by Serban et al. (2016a) are available at: http://www.iulianserban.com/Files/UbuntuDialogueCorpus.zip. These can be used with the model states "prototype_ubuntu_LSTM", "prototype_ubuntu_HRED", and "prototype_ubuntu_VHRED" (see state.py) to reproduce the results of Serban et al. (2016a) on the Ubuntu Dialogue Corpus.
The original Ubuntu Dialogue Corpus as released by Lowe et al. (2015) can be found here: http://cs.mcgill.ca/~jpineau/datasets/ubuntu-corpus-1.0/
Unfortunately due to Twitter's terms of service we are not allowed to distribute Twitter content. Therefore we can only make available the tweet IDs, which can then be used with the Twitter API to build a similar dataset. The tweet IDs and model test responses can be found here: http://www.iulianserban.com/Files/TwitterDialogueCorpus.zip.
The MovieTriples script is also available for research purposes only by contacting Iulian Vlad Serban by email, although we strongly recommend researchers to benchmark their models on Ubuntu and Twitter, since these datasets are substantially larger and represent more well-defined tasks.
References
A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues. Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron Courville, Yoshua Bengio. 2016a. http://arxiv.org/abs/1605.06069
Multiresolution Recurrent Neural Networks: An Application to Dialogue Response Generation. Iulian Vlad Serban, Tim Klinger, Gerald Tesauro, Kartik Talamadupula, Bowen Zhou, Yoshua Bengio, Aaron Courville. 2016b. http://arxiv.org/abs/1606.00776.
Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models. Iulian V. Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, Joelle Pineau. 2016c. AAAI. http://arxiv.org/abs/1507.04808.
The Ubuntu Dialogue Corpus: A Large Dataset for Research in Unstructured Multi-Turn Dialogue Systems. Ryan Lowe, Nissan Pow, Iulian Serban, Joelle Pineau. 2015. SIGDIAL. http://arxiv.org/abs/1506.08909.