guruprasad68 / few-shot-pos_tagging

Few-Shot POS-Tagging

This project was done as a part of UW-Madison's course CS839: Learn more with less supervision by Prof. Fred Sala.
Co-Author: Julia Kroll

We study Few-Shot POS-Tagging using prototype learning and self-training methods following huang et al.,'s work on NER, a similar sequence labeling task. The code was adapted from their codebase. We implement the two methods on three datasets: (i) CoNLL2000 (ii) UDPOS-English Web Treebank (iii) TED.

To run the code:

pip install -r requirements.txt
cd src
bash train_lc.sh
bash train_proto.sh

The results obtained were as follows:

Dataset	Model	Accuracy	shot
CoNLL	Self-Training	0.9898	20-shot
CoNLL	Prototype Learning	0.9848	5-shot
UDPOS	Self-Training	0.8166	20-shot
UDPOS	Prototype Learning	0.8790	5-shot
TED	Self-Training	0.9692	20-shot
TED	Prototype Learning	0.9610	5-shot

More details can be found in the report.