We performed initial runs using the NNDial repository to run the codebase for the end-to-end trainable task-oriented dialogue system. The models trained successfully and the interactive mode enabled us to perform live tests on the trained models.
Trying to interpret the code written in Theano was extremely difficult and major parts of the code were complex and convoluted. Converting this code to pytorch would be a difficult task without enough knowledge about Theano. We also contacted the author and he suggested that it would be quite time consuming to convert the codebase to pytorch. We finally concluded to move ahead with our backup idea still in the dialog system domain. Since we did not want to risk being without a project at the end to the quarter. This new idea is described below.
Instead of focusing on all the models and components that go into making a dialogue system, we are now focusing on the "belief tracker." This model is the main component of the dialogue system as it keeps track of what the user wants.
Our overall goal for this class is to build a dialog system using a belief tracker.
https://arxiv.org/pdf/1606.03777.pdf
This paper introduces a novel way of belief tracking using a hand crafted rule based mechanism that achieves state of the art performance on the WOZ 2.0 dataset.
https://arxiv.org/pdf/1805.11350.pdf
This paper builds up on the earlier version and implements a differentiable training mechanism along with a NBT-CNN to track beliefs.
The most important part of a dialog system is the belief tracker that is able to accurately track the user's beliefs given the context and the history. There exists relevant research such as Fully Statistical Neural Belief Tracking to track the belief of the user's utterance. The paper tracks the belief using fully connected multi layer networks that take in word embeddings of user utterances, slot - value word embeddings and word vectors of preceding system acks (historical beliefs). This paper implements a state of the art network to track beliefs on the WOZ data. The dataset and codebase are available here.
This paper build on top this idea of a Neural Belief tracker (NBT) that was initially mentioned in this paper.
Our goal is to improve the accuracy for the slot tracking that the current state of the art NBT model achieves for slot tracking.