Long Short-Term Memory Neural Networks for Artificial Dialogue Generation

This paper investigates both of user and system modeling to extend an existing corpus of human-machine dialogue data with simulated/artificial dialogues. To simulate and generate such artificial dialogues, a long-short term memory (LSTM) neural network system is proposed. The LSTM neural network is an Encoder-Decoder built on a bidirectional multilayer architecture where the input sequence to the encoder is a list of user dialogue acts and the decoder output sequence is a list of system dialogue acts. All dialogue acts are defined at the intent level and are extracted from the TownInfo corpus for tourist information provided by the FP7 CLASSiC Project funded by European Union. The proposed LSTM configuration is compared to a fully connected Hidden Markov Model (HMM) based architecture where the states are the user dialogues acts and the observations are the system dialogue acts. After carrying out different experiments, the results obtained on the TownInfo corpus showed that the LSTM-based system outperforms the HMM-based system.