Blog 1
Your team name
Baby Jarvis
A list of the team’s members
Divye Pratap Jain, Vardhman Mehta, Kevin Zhao
The top three project ideas you’re excited about
- An end-to-end trainable neural task oriented dialogue system that can perform tasks in multiple domains: paper
- A context aware neural dialogue system that can have long conversations: paper
- Implement transformers in a machine translation task: paper
A minimal viable action plan with stretch goals for each project idea
Action Plan 1: Task Oriented Dialogue System
- Determine the task orientation of the dialogue system
- Start with Restaurant tasks as that is the one mentioned in the reference paper.
- Gather dialogue data surrounding the task
- Use existing datasets made public by the paper and the Stanford task oriented dataset.
- Understand the format of the data and parse it to feed the network.
- Project Scope
- Start by adapting the public version of the model to our own implementation.
- Adapt to more independent tasks. For example
- Have a more generic cross domain performance and see if we can leverage knowledge of one domain to do tasks in another
- Evaluate the cross domain model using blue evaluation method and human evaluation (Ask our peers)
- Stretch Goal: Add Active learning for other related domains, as the model should be able to leverage existing domains to learn new tasks quickly.
Action Plan 2: Conversation Oriented Dialogue System
- Implement the network and implement a baseline version as discussed in paper
- Project Scope
- Replace RNNs with transformers and evaluate performance changes
- Stretch goal: Leverage BERT in combination with this paper to have a more robust and better dialogue system
- Evaluate our dialogue system (Ask our peers)
Action Plan 3:
- Determine a suitable language for the a machine translation model (from English)
- Data is available here for a number of English language pairs
- Brainstorm the general architecture of the translation model: perhaps taking influence from the model described in the following paper
- Implement the machine translation model with various RNN’s and transformers (with various hyperparameter changes) to determine how each neural architecture performs
- Perhaps implement a user interface for people to test and interact with the best performing model (translation from English to another language)
- Stretch goal: Implement a generic architecture that works with multiple languages (inspired by GPT)
The github URL for your project
https://github.com/Divye02/baby-jarvis
Other relevant links