Project

Project Ideas

The following are possible ideas for project (but not limited to this).

Develop a working MT/ASR/TTS/ST system for some new (no high-quality available MT) and low-resource languages (e.g. Spanish-to-Tamil), explore and solve challenges along the way
Improving methods to better utilize monolingual data
Extending and improving Vocabulary and Tokenization for NMT
Improving evaluation quality and efficiency, certain human-assisting tools for evaluation, conduct study.
Computer-assisted and interactive translation methods
MT for multimodal data, e.g. video translation, speech translation
Integrating domain knowledge into MT/ASR/TTS system
Novel hardware-based MT/ASR/TTS/ST system, e.g. Compress MT model to very small size and build a system (with inference but not training) on mobile phones, or extending existing CUDA library (e.g. LightSeq) to support more complex models.
Extending a massive NMT (e.g. LegoMT) to a few more languages

Project Proposal

Please submit your project proposal (one submission per team and be sure to include all your teammates with the submission) to Canvas. The proposal document should be about 1 page (maximum of 2 pages). Be sure to describing the following elements:

What research problem are you planning to address?
What multilingual challenge(s) are you planning to address as part of this project?
What are the existing state-of-art methods on this problem? Is the source code/model available?
Possible directions for proposed method.
What datasets are you planning to use? Please provide the link to the dataset accompanied by a small description. If you have yet to download the dataset, please explain any issues you may be having.
What is the evaluation metric?
Who is your team and how are you planning to split the workload between team members? Can you provide a rough timeline/milestones you plan to follow?
What CPU, GPU and storage infrastructure do need for this project? Are you interested in using Amazon Web Services (AWS) or Google Cloud Platform (GCP) as part of your project? Please estimate the amount of computation time required.

Project Final Report

The final report is intended to be structured as a workshop/conference paper, as mentioned in the first lecture slides. Y'all may have gotten an idea of the typical sections involved in a paper in HW3. Nonetheless, to help with the broad sections typically involved in a research paper, please see below: Kindly note that this is meant as a guide, and you can be flexible in case you have experience writing research papers. There is an upper-limit cap of 8 pages and Appendix length is unrestricted.

Introduction/Motivation This essentially lays out the motivation for the problem, talks about why we need to work on it, the key contributions presented in the work.
Related Work/Background This talks about key papers/works that provide context to your current work. Instead of listing down multiple past works, talk about the ones that minimally differ from your work, and how.
Methodology This section talks about your method, raises research questions and how you are going to address them. Experimental Section/Results This section can describe your experiments and the results you obtain.
Experiments can also be merged with the methodology section if that is more appropriate.
Analysis/Ablations Typically, you would have multiple factors involved in your experimental setting. Analysis sections help you probe deeper into the results and help piece out contributions from individual modeling decisions made.
Conclusion/Discussion This would list the main takeaways from your work, discuss some future ideas (if any) and engage in discussion.
Limitations This section lays out some known limitations of your work.
[Project Only] Team Member Contributions List out each individual's contributions in this section.

Kindly make note of the grading scheme that will be followed for the project. Please note that each project is different so some things can be subjective (especially in the A to A+ range). However, we will try our best to follow this scheme:

A+: A respectable research contribution that is novel and effective, and could be submitted largely as-is as a paper to an academic conference. All elements above are high quality.
A: A respectable contribution that is largely complete and promising, but the description, sophistication of methodology, experiments, or analysis are not as fleshed out and complete as A+ assignments.
A-: All required elements are present, but one or more are lacking to some extent.
B+ or B: The project is complete, but one or more of the required elements is seriously lacking.
B- or below: The project or description thereof is seriously incomplete.

If you are unsure about what project to pick / need more information / want to discuss, Instructor/TAs are here to help you! Please come to office hour (as a group) to discuss about your proposal.