Kenneth Steimel @ksteimel@scholar.social

#introductions Hey everyone. I am a PhD Student in Computational Linguistics. My main research interest is computational linguistics for under resourced languages. I am particularly interested in computational linguistics for Bantu languages because of the extremely complex morphology that they bring to the table.

I am also doing research in sentiment analysis on stance detection and irony/sarcasm detection. Most of my research is on machine learning approaches.

tag attribute pairs.
This means that unlikely combinations of tags and even completely unseen combinations of tags are no problem.
Looking at the results, they are much more accurate. I'll probably make a blog post with actual numbers but it definitely represents an improvement.
I needed better morph tags so that I can leverage them in my cg3 rules and trust that they're probably not leading me astray and now I think I can.

(Both the forward and backward context in addition to characters are used).
This means the model does not work with individual features represented in the feature string. If the model is confronted with a word it hasn't seen, it may get all the tags wrong.

With the multilabel model, the individual pieces of the morph tag are broken apart. The first impact of this is that the tag distribution isn't so skewed. Where there were 3,000+ complex tag combos, there are 82 2/3

Successfully doing multilabel morphological tagging for UD-style tags in Swahili!

In essence, a word like 'vyekundu', could have the features BantuNounClass=8|Number=Sing where the features are a series of attributes and values separated by a pipe symbol.
Previously, I used a naive model where the whole feature string was used as a single label and the model was just predicting the most likely label given the input sequence of words. 1/3

I tracked the error down and it is due to not having models in stanza for Swahili. Then I remembered that stanza uses UD corpora. I'm annotating the first UD corpus of Swahili for the thesis I was procrastinating on. Thus, in the end, my procrastination was thwarted by itself.

So I trained a neural translation model for English to Swahili while procrastinating on thesis stuff a week ago using the scripts provided by argos translate and the global voices corpus i'm expanding on for my thesis work.
It works with argos nicely and the translations are kinda okayish. Might be better not using a transformer model. Anyways, I was wanting to test going the other direction (swh -> eng) but I was getting errors due to a missing tokenizer for swh.