add method to compute comet metric and a test script for it

[tanhaow]

Associated Issue(s): resolves #15

Changes in this PR

• Added compute_comet() function for computing COMET scores using HuggingFace's evaluate library
• Created test_scripts/test_compute_comet.py test script

Notes

• Automatically uses GPU acceleration (MPS on macOS, CUDA on Linux/Windows) when available
• The function takes three parameters: tr_text (translation), src_text (source), and ref_text (reference translation)
• Returns a float in the range [0, 1], where 0 indicates a poor translation and 1 indicates a perfect translation
• First execution will download the COMET model (~500MB) to ~/.cache/huggingface/hub/

Reviewer Checklist

• Review the code
• Run test script with sample data

[laurejt]

@tanhaow Is there specific sample data I should be using for the test script? Add it to the test dataset project drive if you haven't yet.

[tanhaow]

@tanhaow Is there specific sample data I should be using for the test script? Add it to the test dataset project drive if you haven't yet.

I didn’t create a specific one. You could just use a portion of the data we already have.

[laurejt]

It's good enough for now. If there's a way to reduce all of the generated warnings/logging from HuggingFace that'd be nice, it's a lot of text.

In the future, we should probably look into ways to avoid repeatedly loading the same HuggingFace models over and over again for COMET, but that can be a future task. It may be quite important for getting #16.

[laurejt]

While we're editing this file, let's get rid of the mention of BLEU since it's a notoriously bad evaluation metric for machine translation.

Suggested change

	including ChrF, COMET, and potentially BLEU and others in the future.
	including ChrF, COMET, and potentially others in the future.