maryszmary/conllu-perceptron-tagger

A simple perceptron tagger for CoNLL-U files

Data

Serbian

Results

Before (default features; best of 5)

Metrics	Precision	Recall	F1 Score	AligndAcc
Tokens	100.00	100.00	100.00
Sentences	100.00	100.00	100.00
Words	100.00	100.00	100.00
UPOS	95.73	95.73	95.73	95.73
XPOS	100.00	100.00	100.00	100.00
Feats	100.00	100.00	100.00	100.00
AllTags	95.73	95.73	95.73	95.73
Lemmas	100.00	100.00	100.00	100.00
UAS	100.00	100.00	100.00	100.00
LAS	100.00	100.00	100.00	100.00

After (see feachers below; best of 5)

Metrics	Precision	Recall	F1 Score	AligndAcc
Tokens	100.00	100.00	100.00
Sentences	100.00	100.00	100.00
Words	100.00	100.00	100.00
UPOS	96.19	96.19	96.19	96.19
XPOS	100.00	100.00	100.00	100.00
Feats	100.00	100.00	100.00	100.00
AllTags	96.19	96.19	96.19	96.19
Lemmas	100.00	100.00	100.00	100.00
UAS	100.00	100.00	100.00	100.00
LAS	100.00	100.00	100.00	100.00

Changes

What I tried

• changing the size of the context analysed, from 1 to 3
• changing the size of prefix, from 0 to 0:3
• changing the size of suffix, from last 4 to last 2
• adding a feature: lengh of token (didn't help)
• adding a new feature: the last letter in the token
• adding a new feature: containsNumber (helped a lot
• adding a new feature: word[0].isupper()

What I used

• changing the context: -3 +1 words as opposed to -2 +2
• changing the size of prefix: 0 → 0:3
• adding a new feature: the last letter in the token
• adding a new feature: word[0].isupper()

Overall improvement

From 95.73% to 96.19%.

conllu-perceptron-tagger

A (very) simple perceptron tagger for CoNLL-U files, intended for use as a teaching aid.

This is wholely based on the following code:

• https://explosion.ai/blog/part-of-speech-pos-tagger-in-python
• https://github.com/sloria/textblob-aptagger

I've basically taken the code and wrapped it for parsing CoNLL-U format files.

Usage

Like UDpipe:

Train:

cat kk-ud-train.conllu | python3 tagger.py -t model.dat

Predict:

cat kk-ud-test.conllu | python3 tagger.py model.dat > output