Lexicographic POS tagging experiment for isiXhosa.
Data is from here: "Linguistically enriched corpora for conjunctively written South African languages" by Martin Puttkammer and Tanja Gaustad.
Best model (ours) is bolded:
| Model | Accuracy (including noun classes) | Accuracy (not including noun classes) |
|---|---|---|
| du Toit & Puttkammer | 74.3590% | 75.6410% |
| Ours | 82.0513% | 87.1795% |
| NLAPOST21 shared task winner, crf, comp sum, bigram | (does not attempt to tag noun classes) | 64.1026% |
| NLAPOST21 shared task winner, lstm, comp sum, bigram | (does not attempt to tag noun classes) | 60.2564% |
NB: It is very important to note that this is on words as they would appear in a dictionary, e.g. on IsiXhosa.click. This means it is an 'unfair' comparison for these taggers, which are trained on a sentence level. These results do not mean that our tagger is better for any usecase other than dictionary entries formatted like ours are on IsiXhosa.click.
I had to hack this together a bit as a proof-of-concept so the code is in a not-so-great state. If you want to actually use this, please contact me and I'll gladly clean it up for you. This is mostly here for posterity + transparency and is released as-is.
All code is licensed under Apache 2.0. Data retains its original licensing, and we merely redistribute it for convenience.
- du Toit, J. S., & Puttkammer, M. J. (2021). Developing Core Technologies for Resource-Scarce Nguni Languages. Information, 12(12), 520. https://doi.org/10.3390/info12120520
- Pannach, F., Meyer, F., Jembere, E., Dlamini, S. Z. (2022). NLAPOST2021 1st Shared Task on Part-of-Speech Tagging for Nguni Languages. Journal of the Digital Humanities Association of Southern Africa, 3(01). https://doi.org/10.55492/dhasa.v3i01.3865