goal-oriented-chatbot
inzva AI Projects #3 - Training A goal oriented chatbot agent using reinforcement learning
Team Members
Semih Yağcıoğlu
Mustafa Sercan Amaç
Cemil Güney
Efehan Danışman
Introduction
Chatbots became a popular field of research due to it’s room for improvement and practical implications. Hence, under the framework of Inzva AI Projects, we would like to make our humble contribution to the field by training a goal-oriented chatbot. Below you can find what we have done. Special thanks to Inzva team to create a space for everyone who would like to make a contribution!
Literature Review
There were several works that inspired us to contribute the efforts on goal oriented chatbots. First was the paper on End to End Task-Completion Neural Dialogue Systems by researchers from Microsoft Research and National Taiwan University which is also our basis to make improvements. The dialogue system at the paper has different components such as dialogue state tracker to keep track of the status of the dialogue from start to end, natural language understanding (NLU) to understand the user input, natural language generation (NLG) to respond the user and user simulator to train the model.
Nonetheless, model is lacking NLU and NLG components which will be added at our work by using state of the art methods. For language understanding, Liu and Lane’s work on Attention-Based Recurrent Neural Network Models for Joint Intent Detection and Slot Filling using encoder-decoder structures with Recurrent Neural Networks. On the Natural Language Generation component; Semantically Conditioned LSTM-based Natural Language Generationfor Spoken Dialogue Systems from Cambridge University looks promising.
Replication
We replicated results of the study with Korean restaurant reservation dataset as in this repository.
Improvements
Would you like to try?
You can try our chatbot via Telegram. COMING SOON!