Issues of Kyrgyz syntactic annotation
within the Universal Dependencies framework

TurkLang 2023, 2023-10-20

Aida Kasieva¹

Gulnura Dzhumalieva¹

Anna Thompson²

Murat Jumashev²

Bermet Chontaeva³

Jonathan Washington⁴
  1. Kırgız-Türk Manas Üniversitesi, Kyrgyzstan
  2. Independent Scholar
  3. Universität Tübingen, Deutschland
  4. Swarthmore College, USA

Overview

Our talk in a nutshell

We've been working on UD annotation of Kyrgyz material, mostly proverbs and literature.

The next ~13 minutes:

  • Background
    • Turkic Universal Dependencies
    • Kyrgyz syntax
  • Background on corpus
  • Selected annotation issues
    • Copula tokenisation
    • "Small" words (да, эле, бар/жок, керек)
    • Null-headed clauses (RCs, locative & genitive expressions)
    • Inflection vs. derivation

Background on corpus

Timeline

  • 2020: Anna puts together 86-sentence proverbs corpus, with supervision from Jonathan
  • 2021: Aida and Gulnura start working with students on proverbs, literature, and novel sentences
  • 2023: current state:
    • sentences: ≥270
    • tokens: ≥1910
    • Ongoing efforts to combine, check, unify annotations, add metadata, and expand
  • Future: submit to UD project as one large corpus

Tools leveraged in creation

  • Apertium morphological transducer (apertium-kir): (Washington et al. 2012)
    to pre-process sentences (tokenisation, morphological analysis)
  • UD Annotatrix (Tyers et al. 2017): dependency annotation interface
  • Manas-UdS Kyrgyz Corpus (Kasieva et al. 2020): source for sentences

Challenges: copula tokenisation

Defective copula verb э-

  • Irregular attested forms:
    • non-past tense: endings on predicate (үйдөмүн, not үйдө эймин)
    • direct past tense: эле (not эди) + past-tense person endings
    • irregular-looking past verbal noun: экен (not эген)
  • other words derived from э-: эмес not, экен evidently
  • Missing other forms:
    • tenses, non-finite forms: эчү, эген, өө, эш, ээр, эп, эй, etc.
    • negative forms: эбеле

Other UD corpora

  • Kazakh (Tyers & Washington 2015)
    • Tokenise copula separately (like we do)
    • Except not for 3rd person (when null)
  • Kyrgyz (Benli 2023)
    • no separate tokenisation
    • predicates with copula affixes sometimes analysed as verbs
    • эле analysed as VERB and compound:svc dependent of non-verbal predicate
  • The problem: tokenise present and past in parallel way
  • A solution: supertokens
Мен үйдөмүн. I am at home.

# labels = cg-checked-jnw dep-tocheck
# text[gloss] = I at.home am .
1	Мен	мен	PRON	_	_	2	nsubj	_	_
2-3	үйдөмүн	_	_	_	_	_	_	_	_
2	үйдө	үй	NOUN	_	_	0	root	_	_
3	мүн	э	AUX	_	_	2	cop	_	_
4	.	.	PUNCT	_	_	2	punct	_	_
							
Мен үйдө элем. I was at home.

# labels = cg-checked-jnw dep-tocheck
# text[gloss] = I at.home was .
1	Мен	мен	PRON	_	_	2	nsubj	_	_
2	үйдө	үй	NOUN	_	_	0	root	_	_
3	элем	э	AUX	_	_	2	cop	_	_
4	.	.	PUNCT	_	_	2	punct	_	_
							

Challenges: э- vs. other copula-like verbs

Мен үйдөмүн. I am at home.

# labels = cg-checked-jnw dep-tocheck
# text[gloss] = I at.home am .
1	Мен	мен	PRON	_	_	2	nsubj	_	_
2-3	үйдөмүн	_	_	_	_	_	_	_	_
2	үйдө	үй	NOUN	_	_	0	root	_	_
3	мүн	э	AUX	_	_	2	cop	_	_
4	.	.	PUNCT	_	_	2	punct	_	_
						
Мен үйдө болом. I am / will be at home.

# labels = cg-checked-jnw dep-tocheck
# text[gloss] = I at.home am .
1	Мен	мен	PRON	_	_	3	nsubj	_	_
2	үйдө	үй	NOUN	_	_	3	xcomp	_	_
3	болом	бол	VERB	_	_	0	root	_	_
4	.	.	PUNCT	_	_	3	punct	_	_
						
Мен үйдө элем. I was at home.

# labels = cg-checked-jnw dep-tocheck
# text[gloss] = I at.home was .
1	Мен	мен	PRON	_	_	2	nsubj	_	_
2	үйдө	үй	NOUN	_	_	0	root	_	_
3	элем	э	AUX	_	_	2	cop	_	_
4	.	.	PUNCT	_	_	2	punct	_	_
						
Мен үйдө болдум. I was at home.

# labels = cg-checked-jnw dep-tocheck
# text[gloss] = I at.home was .
1	Мен	мен	PRON	_	_	3	nsubj	_	_
2	үйдө	үй	NOUN	_	_	3	xcomp	_	_
3	болдум	бол	VERB	_	_	0	root	_	_
4	.	.	PUNCT	_	_	3	punct	_	_
						

Other UD corpora

  • Kazakh (Tyers & Washington 2015)
    • Treated same as copula constructions
  • Kyrgyz (Benli 2023)
    • Complements of бол- analysed as amod dependents

Challenges: small words: эле

occurs after

  • nouns: бала эле just a child
  • adjectives: кичинекей эле not that big
  • numbers: эки эле just two
  • adverbs: кечээ эле just yesterday
  • adverbial clauses: үч күн өткөндөн кийин эле only after three days passing

ADV, advmod:emph


# text[gloss] = three day pass-VN-ABL after just book-ACC I.turned.in .
1	үч	_	NUM	_	_	2	nummod	_	_
2	күн	_	NOUN	_	_	3	nsubj	_	_
3	өткөндөн	_	VERB	_	_	7	advcl	_	_
4	кийин	_	ADP	_	_	3	case	_	_
5	эле	_	ADV	_	_	3	advmod:emph	_	_
6	китепти	_	NOUN	_	_	7	obj	_	_
7	тапшырдым	_	VERB	_	_	0	root	_	_
8	.	_	PUNCT	_	_	7	punct	_	_
					

Other analyses

  • Kyrgyz (Benli 2023)
    • ADV, advmod (except e.g. in чын эле, fixed)
  • Kazakh (Tyers & Washington 2015)
    • (ғана/қана) ADV, advmod

Challenges: small words: да

Uses of да

  1. Post-predicate "modal particle"
    Натыйжалар жарыяланыптыр да, ээ?
    ‘The results have been announced, haven't they?’
  2. Conditional intensifier
    Оозу кыйшык болсо да, байдын уулу сүйлөсүн.
    Even if his mouth is crooked, let the rich person's son speak.’
  3. General contrastive intensifier
    Тамашада да чындыктын үлүшү бар.
    Even in a joke is some element of truth.’
  4. General conjoining adverb
    Атам да каршы болду.
    ‘My father was also against it.’ (or: ‘Even my father was against it.’)
  5. Correlative conjunction
    Атам да, апам да каршы болду.
    Both my father and my mother were against it.’

Uses 2&3, 3&4, 4&5 seem very similar: we treat the same

Challenges: small words: да

Use 1: PART, discourse


# sent_id = 15
# text = Натыйжалар жарыяланыптыр да, ээ?
# text[en] = The results have been announced, haven't they?
1	Натыйжалар	натыйжа	NOUN	n.pl.nom	Case=Nom|Number=Plur	2	nsubj	_	_
2	жарыяланыптыр	жарыяла	VERB	v.tv.pass.ifi.evid.p3.pl	Evident=Nfh|Mood=Ind|Number=Plur|Person=3|Tense=Past|VerbForm=Fin|Voice=Pass	0	root	_	_
3	да	да	PART	mod	_	2	discourse	_	SpaceAfter=No
4	,	,	PUNCT	cm	_	2	punct	_	_
5	ээ	ээ	PART	mod	_	2	discourse	_	SpaceAfter=No
6	?	?	PUNCT	sent	_	2	punct	_	_
						

Uses 2-5: ADV, advmod:emph


# sent_id = proverbs:42
# author = Сыргабековна&Сабыровна
# source = Сыргабековна&Сабыровна:413
# text = Тамашада да чындыктын үлүшү бар.
# text[eng] = Even in a joke is some element of truth.
# labels = cg-checked-jnw dep-initialchecked-jnw
1	Тамашада	тамаша	NOUN	n.loc	Case=Loc	5	obl	_	_
2	да	да	ADV	postadv	_	1	advmod:emph	_	_
3	чындыктын	чындык	NOUN	n.gen	Case=Gen	4	nmod:poss	_	_
4	үлүшү	үлүш	NOUN	n.px3sp.nom	Case=Nom|Number[psor]=Plur,Sing|Person[psor]=3	5	nsubj	_	_
5-6	бар	_	_	_	_	_	_	_	_
5	бар	бар	ADJ	adj	_	0	root	_	SpaceAfter=No
6	_	э	AUX	cop.aor.p3.sg	Mood=Ind|Number=Sing|Person=3|Tense=Aor|VerbForm=Fin	5	cop	_	SpaceAfter=No
7	.	.	PUNCT	sent	_	5	punct	_	_
						

Other analyses

  • Kyrgyz (Benli 2023)
    • CCONJ, mark
  • Kazakh (Tyers & Washington 2015)
    • ADV, advmod, dependent on root (not preceding word)

Challenges: small words: бар/жок

ADJ, acl


# sent_id = proverbs:47
# author = Сыргабековна&Сабыровна
# source = Сыргабековна&Сабыровна:282
# text = Колунда токочу бар бала сүйгүнчүктүү.
# text[eng] = The child with a loaf (of bread) in their hand is darling.
# text[gloss] = in.their.hand their.loaf present child darling is .
# labels = cg-checked-jnw dep-checked-jnw
1	Колунда	кол	NOUN	_	Case=Loc|Number[psor]=Plur,Sing|Person[psor]=3	3	obl	_	_
2	токочу	токоч	NOUN	_	Case=Nom|Number[psor]=Plur,Sing|Person[psor]=3	3	nsubj	_	_
3	бар	бар	ADJ	_	_	4	acl	_	_
4	бала	бала	NOUN	_	Case=Nom	5	nsubj	_	_
5-6	сүйгүнчүктүү	_	_	_	_	_	_	_	_
5	сүйгүнчүктүү	сүйгүнчүктүү	ADJ	_	_	0	root	_	_
6	_	э	AUX	_	Mood=Ind|Number=Sing|Person=3|Tense=Aor|VerbForm=Fin	5	cop	_	_
7	.	.	PUNCT	_	_	5	punct	_	_
					

ADJ (SUBST, LOC), advcl


# sent_id = proverbs:2
# author = Сыргабековна&Сабыровна
# source = Сыргабековна&Сабыровна:363
# text = Өрдөк жокто, чулдук бий.
# text[eng] = Upon the absence of the/a duck, (the) sandpiper (is) ~king.
# text[gloss] = duck absent.LOC , sandpiper ~king is .
# labels = cg-checked-jnw dep-revised1x-jnw
1	Өрдөк	өрдөк	NOUN	_	Case=Nom	2	nsubj	_	_
2	жокто	жок,	ADJ	_	Case=Loc	5	advcl	_	_
3	,	,	PUNCT	_	_	5	punct	_	_
4	чулдук	чулдук	NOUN	_	Case=Nom	5	nsubj	_	_
5-6	бий	_	_	_	_	_	_	_	_
5	бий	бий	NOUN	_	Case=Nom	0	root	_	_
6	_	э	AUX	_	Mood=Ind|Number=Sing|Person=3|Tense=Aor|VerbForm=Fin	5	cop	_	_
7	.	.	PUNCT	_	_	5	punct	_	_
					

Challenges: small words: керек

ADJ


# text[gloss] = I to.my.family needed am .
1	Мен	_	PRON	_	_	3	nsubj	_	_
2	үй-бүлөмө	_	NOUN	_	_	3	obl	_	_
3-4	керекмин	_	_	_	_	_	_	_	_
3	керек	керек	ADJ	_	_	0	root	_	_
4	мин	э	AUX	_	_	3	cop	_	_
5	.	_	PUNCT	_	_	3	punct	_	_
					

# text[gloss] = I book-ACC my.turning.in needed is .
1	Мен	_	PRON	_	_	3	nsubj	_	_
2	китепти	_	NOUN	_	_	3	obj	_	_
3	тапшырышым	_	VERB	_	_	4	csubj	_	_
4	керек	_	ADJ	_	_	_	_	_	_
5	_	э	AUX	_	_	4	cop	_	_
6	.	_	PUNCT	_	_	4	punct	_	_
					

Challenges: substantivised verbal adjectives

  • Verbal adjectives with "empty head"
  • Can be read in English as "the person/thing/one"
  • Two options:
    • Treat simply as "clause": ccomp, csubj
    • Treat as noun (not present): obj, nsubj ()

# sent_id = proverbs:21
# author = Сыргабековна&Сабыровна
# source = Сыргабековна&Сабыровна:283
# text = Колуң менен кылганды, мойнуң менен тартасың.
# text[eng] = You will pull with your neck what you make with your hands.
# text[gloss] = your.hand with made.⌷-ACC, your.neck with you.will.pull .
# labels = cg-checked-jnw dep-revised2x-jnw CHECK
1	Колуң	кол	NOUN	_	Case=Nom|Number[psor]=Sing|Person[psor]=2	3	obl	_	_
2	менен	менен	ADP	_	_	1	case	_	_
3	кылганды	кыл	VERB	_	Case=Acc|Tense=Past|VerbForm=Part	7	obj(/ccomp)	_	_
4	,	,	PUNCT	_	_	7	punct	_	_
5	мойнуң	моюн	NOUN	_	Case=Nom|Number[psor]=Sing|Person[psor]=2	7	obl	_	_
6	менен	менен	ADP	_	_	5	case	_	_
7	тартасың	тарт	VERB	_	Mood=Ind|Number=Sing|Person=2|Tense=Aor|VerbForm=Fin	0	root	_	_
8	.	.	PUNCT	_	_	7	punct	_	_
					

# sent_id = proverbs:16
# author = Сыргабековна&Сабыровна
# source = Сыргабековна&Сабыровна:84
# text = Балалуу болбогон кубанганды билбеген.
# text[gloss] = child with not.being.⌷ being.happy-ACC has.not.known .
# text[eng] = One who has not had children has not known being happy.
# labels = cg-checked-jnw dep-checked-jnw
1-2	Балалуу	_	_	_	_	_	_	_	_
1	Бала	бала	NOUN	_	_	3	xcomp	_	_
2	луу	луу	ADP	_	_	1	case	_	_
3	болбогон	бол	VERB	_	Case=Nom|Polarity=Neg|Tense=Past|VerbForm=Part	5	nsubj(/csubj)	_	_
4	кубанганды	кубан	VERB	_	Case=Acc|Tense=Past|VerbForm=Ger	5	obj	_	_
5	билбеген	бил	VERB	_	Mood=Ind|Number=Sing|Person=3|Polarity=Neg|Tense=Past|VerbForm=Fin	0	root	_	_
6	.	.	PUNCT	_	_	5	punct	_	_
					

# sent_id = proverbs:22
# author = Сыргабековна&Сабыровна
# source = Сыргабековна&Сабыровна:465
# text = Шашкан шайтандын иши.
# text[gloss] = rushing shaytan's work is .
# text[eng] = Hurrying is what a shaytan does. / Hurrying is the devil's work.
# labels = cg-checked-jnw dep-checked-jnw
1	Шашкан	шаш	VERB	_	Case=Nom|Tense=Past|VerbForm=Ger	3	csubj	_	_
2	шайтандын	шайтан	NOUN	_	Case=Gen	3	nmod:poss	_	_
3-4	иши	_	_	_	_	_	_	_	_
3	иши	иш	NOUN	_	Case=Nom|Number[psor]=Plur,Sing|Person[psor]=3	0	root	_	_
4	_	э	AUX	_	Mood=Ind|Number=Sing|Person=3|Tense=Aor|VerbForm=Fin	3	cop	_	_
5	.	.	PUNCT	_	_	3	punct	_	
					

Challenges: -DAGI

  • Attributed form of locative -DA, normally nmod:loc
  • Can be substantivised
    • results in two participants encoded in one form
    • can have mismatches in number, case, etc.
    • if just analysed as one token, analysis is misleading
      (текче isn't the subject of кайда)
    • poses major challenge for analysis without retokenisation

# sent_id = 7
# text = Макул, анда китеп текчесиндегилер кайда эле?
# text[en] = Okay, then where are/were the ones on the bookshelf?
1	Макул	Макул	INTJ	ij	_	7	discourse	_	SpaceAfter=No
2	,	,	PUNCT	cm	_	7	punct	_	_
3	анда	анда	ADV	adv	_	7	advmod	_	_
4	китеп	китеп	NOUN	n.nom	Case=Nom	5	nmod:poss	_	_
5-6	текчесиндегилер	_	_	_	_	_	_	_	_
5	текчесиндеги	текче	NOUN	n.px3sp.loc.attr	Case=Loc|Case=Nom|Number[psor]=Plur,Sing|Person[psor]=3	6	nmod:loc	_	SpaceAfter=No
6	лер	_	NOUN	n.nom.pl	Case=Nom|Number=Plur	7	nsubj	_	_
7	кайда	кайда	ADV	adv	_	0	root	_	_
8	эле	эле	AUX	cop.ifi.p3.pl	Mood=Ind|Number=Plur|Person=3|Tense=Past|VerbForm=Fin	7	cop	_	SpaceAfter=No
9	?	?	PUNCT	sent	_	7	punct	_	_
					

Challenges: -NIKI

  • substantivised form of genitive
  • also must treat as having a null head
  • retokenisation is needed

# sent_id = 1B
# text = Жубайым дачадагы балдар бөлмөсүнүн терезесин шаша-буша тазалап, ашкананыкын тазалабай коюптур.
# text[en] = My spouse cleaned the windows of the children's room at the summer house hastily, but didn't clean the kitchen's.
# text[tr] = Eşim yazlıktaki çocuk odasının penceresini apar topar temizlemiş, mutfağınkini temizlememişti.
1	Жубайым	жубай	NOUN	n.px1sg.nom	Case=Nom|Number[psor]=Sing|Person[psor]=1	13	advcl	_	_
2	дачадагы	дача	NOUN	n.loc.attr	Case=Loc|Case=Nom	3	nmod:loc	_	_
3	балдар	бала	NOUN	n.pl.nom	Case=Nom|Number=Plur	4	nmod:poss	_	_
4	бөлмөсүнүн	бөлмө	NOUN	n.px3sp.gen	Case=Gen|Number[psor]=Plur,Sing|Person[psor]=3	5	nmod:poss	_	_
5	терезесин	терезе	NOUN	n.px3sp.acc	Case=Acc|Number[psor]=Plur,Sing|Person[psor]=3	9	obj	_	_
6	шаша	шаш	VERB	v.iv.gna_impf	Aspect=Imp|VerbForm=Conv	9	advcl	_	SpaceAfter=No
7	-	-	PUNCT	guio	_	6	punct	_	SpaceAfter=No
8	буша	буш	VERB	v.iv.gna_impf	Aspect=Imp|VerbForm=Conv	6	fixed	_	_
9	тазалап	тазала	VERB	v.tv.gna_perf	Aspect=Perf|VerbForm=Conv	13	advcl	_	SpaceAfter=No
10	,	,	PUNCT	cm	_	12	punct	_	_
11-12	ашкананыкын	_	_	_	_	_	_	_	_
11	ашкананы	ашкана	NOUN	n.gen	Case=Gen	12	nmod:poss	_	SpaceAfter=No
12	кын	_	NOUN	n.acc	Case=Acc	13	obj	_	SpaceAfter=No
13	тазалабай	тазала	VERB	v.tv.neg.prc_perf	Aspect=Perf|Polarity=Neg|VerbForm=Cov	0	root	_	_
14	коюптур	кой	AUX	vaux.ifi.evid.p3.sg	Evident=Nfh|Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin	13	aux	_	SpaceAfter=No
15	.	.	PUNCT	sent	_	13	punct	_	_
					

Other analyses

  • Kyrgyz (Benli 2023)
    • single token: ADJ, amod
  • Kazakh (Tyers & Washington 2015)
    • multiple tokens, but second one is case dependent on first
    • doesn't capture the analysis of two separate participants in the event

Challenges: level of analysis

Inflected


# text[gloss] = person error-less is.not , marsh frog-less is.not .
1	Адам	адам	NOUN	n	_	4	nsubj	_	nom
2-3	катасыз	_	_	_	_	_	_	_	_
2	ката	ката	NOUN	n	_	4	xcomp	_	_
3	сыз	сыз	ADP	post	_	2	case	_	_
4	болбос	бол	VERB	v	_	0	root	_	iv|neg|fut|p3|sg
5	,	,	_	cm	PUNCT	4	punct	_	_
6	саз	саз	NOUN	n	_	9	nsubj	_	acc-ind
7-8	бакасыз	_	_	_	_	_	_	_	_
7	бака	бака	NOUN	n	_	9	xcomp	_	_
8	сыз	сыз	ADP	post	_	7	case	_	_
9	болбос	бол	VERB	v	_	4	conj/parataxis	_	iv|neg|fut|p3|sg
10	.	.	PUNCT	sent	_	9	punct	_	_
						

Derivational


# text[gloss] = person errorless is.not , marsh frogless is.not .
1	Адам	адам	NOUN	n	_	3	nsubj	_	nom
2	катасыз	катасыз	ADJ	n	_	3	xcomp	_	_
3	болбос	бол	VERB	v	_	0	root	_	iv|neg|fut|p3|sg
4	,	,	_	cm	PUNCT	3	punct	_	_
5	саз	саз	NOUN	n	_	7	nsubj	_	acc-ind
6	бакасыз	бакасыз	ADJ	n	_	7	xcomp	_	_
7	болбос	бол	VERB	v	_	3	conj/parataxis	_	iv|neg|fut|p3|sg
8	.	.	PUNCT	sent	_	7	punct	_	_
						

Conclusion

Чоң ракмат!

Katta rahmat!