This code is an implementation of the Dual-Encoder LSTM introduced in The Ubuntu Dialogue Corpus: A Large Dataset for Research in Unstructured Multi-Turn Dialogue Systems.
The data can be found here and must be placed in the data/ directory.
This project has some (WIP, and therefore messy) improvements on the model introduced in the paper. I've experimented with an attention mechanism and CNN-based encoders. I intend to clean all of this up soon and add a variety of other improvements, in the very near future.
Simply run python3 train.py. Edit the hyperparemters at the the top of the file.
Pre-requisites are PyTorch, CUDA, numpy and NLTK.
The predict.py file contains two methods of interest for inference, using the pre-trained model provided in the repo. To run inference with an alternate model, replace the code on line 7 of the file.
predict_val, given a context (sequence of messages delimited by particular tokens) and a reply (single message) determines the likelihood of the reply following the context.
predict, given a string (be it a message or a sequence) provides the output of the encoder for the string. This output can be utilizes as a mesasge embedding for various purposes.
Feel free to contact me at [email protected] if you have any questions regarding this implementation.
If you use this repository for any research, please cite:
@inproceedings{mehri2017chat,
title={Chat disentanglement: identifying semantic reply relationships with random forests and recurrent neural networks},
author={Mehri, Shikib and Carenini, Giuseppe},
booktitle={Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
volume={1},
pages={615--623},
year={2017}
}