Possesives in Turkish

[ilker...@gmail.com]

Hi,

In English I see separate tokens for possesive case as in:

"Lisa's room is always clean while Ross's room is always messy."

<Token index=1;words=[<Word index=1;text=Lisa;lemma=Lisa;upos=PROPN;xpos=NNP;feats=Number=Sing;governor=3;dependency_relation=nmod:poss>]>

<Token index=2;words=[<Word index=2;text='s;lemma='s;upos=PART;xpos=POS;feats=_;governor=1;dependency_relation=case>]>

But in Turkish there is no seperate token for possesive additon for "'nın" (while getting lemma right):
"Gül'ün odası her zaman dağınıkken Lisa'nın odası hep temiz."
<Token index=6;words=[<Word index=6;text=Lisa'nın;lemma=Lisa;upos=PROPN;xpos=Prop;feats=Case=Gen|Number=Sing|Person=3>]>
<Token index=7;words=[<Word index=7;text=odası;lemma=oda;upos=NOUN;xpos=Noun;feats=Case=Nom|Number=Sing|Number[psor]=Sing|Person=3|Person[psor]=3>]>

Another case for capitalization:
"Bir Hırvat seçim gözlem STK'sı olan GONG seçimlerde 700 gözlemci görevlendirdi."
<Token index=5;words=[<Word index=5;text=STK'sı;lemma=Stk;upos=PROPN;xpos=Prop;feats=Case=Acc|Number=Sing|Person=3;governor=4;dependency_relation=compound>]>
STK is an abbreviation and its lemma should be STK as well but it is converted to "Stk"

Is it caused by library or model? 
If it is model, could you guide me how to fix a situation like this?

[Peng Qi]

I believe this is largely a result of the Universal Dependencies annotations we used for Turkish, which is baked into the model. If you want the possessives to be segmented off the main word, the best course of action is probably either a) re-annotating your own dataset and training models of the stanfordnlp pipeline on that data, or b) post-editing the annotations from stanfordnlp following rules.

That being said, stanfordnlp does give you clues about what is going on here. For instance, in the case of "Lisa'nın", the morphological features contains Case=Gen, which means this word is a genitive, and from which one could infer the relation between Lisa and the room (although the model probably should've identified it as a possessive in this case). See also: https://universaldependencies.org/u/feat/Case.html#Gen

Hope this helps!

[ilker...@gmail.com]

I guess for the long term it will be better to fix "Universal Dependencies annotations".

Thanks for your reply.

[ilker...@gmail.com]

Hi Peng,

Is there any example to post-editiong the annotation?

For example adding another token for tokens carrying feature Case=Gen case.

Or changing POS of token?

Thanks in advance.

[Peng Qi]

Unfortunately not. But the page linked above (https://universaldependencies.org/u/feat/Case.html) and this one (https://universaldependencies.org/u/feat/Poss.html) should contain information about identifying possessives. (Case=Gen and Poss=Yes are usually strong indicators)

One caveat is that the model's prediction is not always perfect, so you might not want to rely on just one of these features.

Hope this helps!