From undergrad, working with UVA's Bioinformatics Laboratory. The lab had built MUST-CNN, a multilayer shift-and-stitch convolutional network for per-position sequence tagging, developed for biological sequences. The open question was whether it would generalize to natural language.
To test that, I wrote a small pipeline with NLTK to convert the RST-WSJ treebank — the Rhetorical Structure Theory annotations over Wall Street Journal text — into the per-position tagging format the model expected: building a tag dictionary, splitting text into elementary discourse units, and aligning words to their EDU tags. Four scripts cover the pipeline end to end, run over the full RST-WSJ-main-1.0 corpus.
Old work, kept here as the earliest project I have a repo for.