Lab 4, SDS 2016

Content

• Fork the project sds-tracker from Github

Homework - RNN DST

• Report bugs
• Come up with a RNN model encoded as Tensorflow computation graph
  • Implement it in tracker/GRUmodel.py
  • Use RNN for encode inputs from each turn sys_utt + DELIM + user_utt
  • For each turn predict the slots based on the last state from the RNN
    • This model does in fact more SLU (spoken language understanding) because it does care about history
    • If you have this model working start refactoring the code for using encoding dialogue history with another RNN