УДК 81’32
AN ASSESSMENT OF UNIVERSAL DEPENDENCY ANNOTATION GUIDELINES FOR TURKIC LANGUAGES
Francis M. Tyers1, Jonathan Washington2, Çağrı Çöltekin3, Aibek Makazhanov4
1School of Linguistics, Higher School of Economics, Moscow;
2 Linguistics Department, Swarthmore College, Swarthmore;
3 Seminar für Sprachwissenschaft, Universität Tübingen;
4 National Laboratory Astana, Nazarbayev University, Astana [email protected]
Annotated corpora of three Turkic languages – Turkish, Kazakh, and Uyghur – were released as part of version 2 of the Free/Open-Source Universal Dependencies (UD) syntactic and morphological annotation guidelines. The objective of these guidelines is to provide consistent dependency annotation to facilitate cross-linguistic comparison.
This paper presents the current state of each of the three UD-annotated Turkic corpora, along with an evaluation of the performance of parsers trained on these corpora.
Overall, the UD annotation guidelines for Turkish, Kazakh, and Uyghur are fairly compatible – a testament to the careful design of the guidelines. However, the specifi c annotation guidelines for each of these languages were developed mostly independently; because of this, differences between the three standards exist. Moving forward with Turkic annotation standards in UD, attempts will be made to reconcile the differences. These differences are overviewed in this paper.
Furthermore, a number of issues in annotation have arisen and have yet to be resolved. Some of these issues require further investigation of the phenomena, and some require consultation within the UD community to determine whether solutions may be determined based on similar phenomena in other languages.
A number of these open issues are discussed, including tokenisation (how to deal with words that include an orthographic space, or multiple words that do not include an orthographic space), the difference between core and oblique arguments of verbs, complex predicates (including structures where there is a combination of a non-fi nite form which governs argument structure and contributes to TAM and a fi nite-form which contributes to TAM and takes person agreement), multiple derivation (multiple causative or causative–passive combinations), and use of copulas instead of auxiliaries in what appear to be auxiliary constructions.
Keywords: Turkish; Kazakh; Uyghur; treebank; dependency grammar;
Universal Dependencies.
ОЦЕНКА КРИТЕРИЕВ МОРФО-СИНТАКСИЧЕСКОЙ РАЗМЕТКИ ДЛЯ ТЮРКСКИХ ЯЗЫКОВ В ПРОЕКТЕ
«UNIVERSAL DEPENDENCIES»
Francis M. Tyers1, Jonathan Washington2, Çağrı Çöltekin3, Aibek Makazhanov4
1Школа лингвистики, Высшая школа экономики, Москва;
2Департамент лингвистики, Суортмор-колледж, Суортмор;
3Школа лингвистики, Тюбингенский университет;
4Национальная Лаборатория Астана, Назарбаев Университет, Астана
Аннотированные корпусы трех тюркских языков – турецкого, казах- ского и уйгурского – были выпущены в составе второй версии проекта
«Universal Dependencies», предоставляющего свободно распространяемые рекомендаций к универсальной морфо-синтаксической разметке. Целью этих рекомендаций является предоставление единой схемы разметки для упрощения межязыкового анализа.
В настоящей работе описано текущее состояние каждого из трех тюрк- ских корпусов размеченных по принципам UD, а также оценка эффектив- ности синтаксических парсеров, обученных на этих корпусах.
Схемы разметки UD для турецкого, казахского и уйгурского языков во многом совместимы, что свидетельствует о тщательной проработке уни- версальности принципов UD. Однако конкретные рекомендации по аннота- ции для каждого из этих языков разрабатывались в основном независимо;
из-за этого существуют различия между тремя стандартами. При дальней- шей разработке схем разметки для тюркских языков будут предприняты попытки сгладить данные различия, которые также рассмотрены в данной работе.
Кроме того, возник ряд вопросов по разметке определенных конструк- ций. Ответы на некоторые из этих вопросов требуют дальнейшего изучения природы соответствующих явлений в языке. В других случаях ответы мо- гут быть получены на основе анализа схожих явлений в других языках; для этого потребуются консультаций с членами сообщества UD. В данной ра- боте обсуждаются ключевые вопросы разметки, в частности: токенизация (считать ли слова, включающие в себя орфографические пробелы отдель- ными единицами разметки, и наоборот, разбивать ли несколько синтакси- ческих слов, являющихся частью одного орфографического, на отдельные единицы разметки); разница между актантами и сирконстантами; сложные предикаты (включая структуры, где существует комбинация нефинитной формы, которая управляет аргументной структурой и несет временные и
аспектно-модальные функции, и финитной формы, которая также несет временные и аспектно-модальные функции и принимает личное оконча- ние); множественная деривация (комбинации из нескольких каузативов и каузативов-пассивов); использование копулы вместо вспомогательного глагола в конструкциях, напоминающих сочетание главного и вспомога- тельного глаголов.
Ключевые слова: Турецкий язык; Казахский язык; Уйгурский язык;
грамматика зависимостей; проект «Universal Dependencies».
1. Introduction
Universal Dependencies (UD, Nivre et al. 2016) is a Free/Open- Source set of guidelines for syntactic and morphological annotation of corpora, which aims to provide consistent dependency annotation to facilitate cross-linguistic comparison. In addition to the guidelines, an- notated corpora are made available under a Free/Open-Source license.
This paper overviews recent work that has gone into making UD annotation guidelines for Turkic languages based on the UD standard.
The current status of UD-annotated corpora of Turkic languages is overviewed in section 2. Three separate efforts have resulted in fairly compatible guidelines for Turkish (southwestern Turkic, §2.1), Ka- zakh (northwestern Turkic §2.2), and Uyghur (southeastern Turkic,
§2.3), which is a testament to the careful design of the UD guidelines.
However, because the specifi c annotation guidelines for each of these languages were developed mostly independently, differences between them exist, as described in section 2.5. Moving forward with Turkic annotation standards in UD, attempts will be made to reconcile the existing differences. In addition to efforts with these three languages and annotation standards, annotated corpora of some other Turkic lan- guages have been begun as well (§2.4).
Furthermore, a number of issues in annotation have arisen and have yet to be resolved. Some of these issues require further investigation of the phenomena, and some require consultation within the UD com- munity to determine whether solutions may be determined based on similar phenomena in other languages. A number of these open issues are discussed in section 4, including tokenisation (§4.1), the difference between core and oblique arguments of verbs (§4.2), complex predi- cates (§4.3), multiple derivation (§4.4), and use of copulas in auxiliary constructions (§4.5). Section 5 wraps up.
2. Current status
In this section we briefl y describe treebanks of Turkic languages that have been or are about to be released in UD.
Table 1. Turkic UD treebanks at a glance
Treebank Language Sentences Words Annotation Genre Kazakh-UD Kazakh 1 047 10 032 manual
annotation Wikipedia, fi ction IMST-UD Turkish 4 660 48 093 semi-auto. conversion news,
social media Turkish-
PUD Turkish 1 000 16 886 auto./manual annotation translated news
Turkish-GK Turkish 2 803 17 800 manual
annotation grammar examples Uyghur-UD Uyghur 100 1 662 semi-auto. conversion fi ction
In addition to the released Turkish (§2.1), Kazakh (§2.2), and Uy- ghur (§2.3) corpora, there has been some work on UD annotation of other Turkic languages (§2.4). Section 2.5 outlines the main differ- ences between the annotation standards of the released corpora.
2.1 Turkish
Turkish is relatively well represented in the UD with two tree- banks. The IMST-UD treebank (Sulubacak et al. 2016a) is the result of a semi-automatic conversion of the IMST treebank (Sulubacak et al.
2016b) which, in turn, was based on METU-Sabancı treebank (Ofl azer et al. 2003). The second Turkish treebank, Turkish-PUD, in the of- fi cial UD repository is part of the parallel treebanks released during CoNLL 2017 UD parsing shared task (Zeman et al. 2017). Besides these two treebanks, the treebank reported in Çöltekin 2015 (Turkish- GK) is annotated for the purpose setting UD annotation guidelines for Turkish. To cover a wide range of morphosyntactic phenomena, the Turkish-GK treebank annotates the example sentences from compre- hensive grammar book. This treebank follows UD version 1.3 annota- tion scheme, and currently not converted to version 2.0.
2.2 Kazakh
Kazakh is represented in UD by a single treebank (Makazhanov et al. 2015; Tyers and Washington 2015), which was fi rst released in UD v1.3, and at the moment of writing contains 1109 trees (sentences) and a total of 10894 tokens. The annotation scheme of the treebank defi nes 16 UD POS tags, 45 “category=value” feature pairs, and 34 depen- dency relations of which four are language-specifi c. Tokenisation and morphological processing strategies in the Kazakh UD treebank follow the principles of Turkic lexica as defi ned by the Apertiun project1. One reason for this is to keep the UD corpus compatible with the morpho- logical analysers developed by the Apertium Turkic working group.
Currently the treebank is partially compatible with UD v2.0 stan- dard, with the choice of head direction in some constructions being one of the major discrepancies. The standard requires coordination and some compounds (e.g. names) to be left-headed, while the treebank developers believe that in Kazakh (and other Turkic languages) such constructions should be right-headed due to the placement of morpho- logical locus, which is exclusive to the last (rightmost) element of such constructions. So far this issue has been resolved by an intermediate conversion step, where initially the annotation is performed in a right- headed fashion, and at the time of release a special script fl ips the heads of the constructions in question.
2.3 Uyghur
In Aili et al. (2016b), a treebank for Uyghur with 20,000 tokens is described. Tokens fi t into one of 12 part-of speech categories and there are 137 morphological tags. There are 23 total dependency relations, with adjuncts classifi ed by morphological case. In co-ordination, the conjunction is attached to the following conjunct and the preceding conjunct is attached to the following one (so-called ‘head-fi nal’ con- junction).
Aili et al. (2016a) present a conversion of the Uyghur dependency treebank Aili et al. (2016b) to Universal Dependencies. They used some default mapping rules to convert the parts of speech and de-
1 http://wiki.apertium.org/wiki/Turkic_lexicon
pendency relations, and then some limited rules based on the part of speech of the head to distinguish between ambiguous relations (for example mapping att → {amod, det, nummod}. The treebank contains surface forms, parts of speech and dependency relations, but no lemmas or morphological features.
2.4 Other
Ageeva and Tyers (2016) present two small treebanks for Tuvan and for Crimean Tatar of approximately 1,000 tokens each for use in testing a method of cross-lingual dependency parsing. They show that it is possible to take advantage of a morphological analyser and a tree- bank for another language in order to learn an improved delexicalised parser.
2.5 Main differences
At present, there are a number of differences in the dependency annotation standards for Kazakh, Turkish, and Uyghur. Quite a few of these differences are in the morphological annotation (part-of-speech tags and morphological features), but there are a handful of differences in tokenisation (how to approach words that include an orthographic space, or multiple words that do not include an orthographic space) and dependency annotation as well. In general, the Kazakh and Turk- ish annotation standards are more compatible with one another than either is with Uyghur.
One example of a difference in part-of-speech tagging is how loca- tional pronoun-derived adverbials are represented. In Turkish, words like nerede ‘where’ and nereden ‘from where’ are labelled as PRON (with the appropriate case indicated in the morphology features), and hence usually have the dependency relation of nominal adverbials, obl. In Kazakh, the corresponding words қайда ‘where’ and қайда
‘from where’ are labelled as fl at adverbs, or ADV, and hence have de- pendency relations of advmod. Which analysis is more appropriate is not clear: the fact that they are pronouns with case suffi xes in both languages argues for their annotation as pronouns, while the fact that these pronouns are defective (they can’t take all case suffi xes) in each language argues for their analysis as grammaticalised adverbs.
Figure 1. Two alternative analyses for speech with the adverbial “quotative word” use of the verb de- ‘say’. Both sentences mean “I asked, ‘what kind of woman is she?’” Analysis 1a shows the de- verb form as a case marker to the clause it governs, which is in turn an object of the main verb. In the analysis in 1b, the speech verb is treated as a verbal adverb adjunct to the main verb, with
its own clausal complement
The current Uyghur corpus does not have annotation for mor- phological features, and there are some notable differences between Kazakh and Turkish standards. One of these is the annotation of Person=3 in Turkish for any nominal–since as they trigger third-per- son agreement morphology as subjects and possessors. Kazakh does not have this feature. Similarly, Turkish annotates for Polarity val- ues (e.g., on verb forms) of both Pos ‘positive’ and Neg ‘negative’, whereas Kazakh only indicates polarity if the value is negative.
Table 2. Language specifi c relations: the tick mark () means that the re- lation is found in the treebank; the − mark means that the relation could
apply, but is not applied at present; and the asterisk (*) means that we consider this relation to be an erroneous classifi cation
Relation Comments Kazakh Turkish Uyghur
acl:poss Adnominal modifi cation with
possessive − −
acl:relcl Adnominal modifi cation with
verbal adjective − −
advmod:emph Adverbial emphasiser (mostly
-dA) −
aux:q Question word, -mI − −
compound:lvc Light verb
compound:redup Reduplication compound −
fl at:name Proper name − −
iobj:caus Causee − −
nmod:abl Oblique in the ablative * *
nmod:cau Causee * *
nmod:clas Noun-noun compound * *
nmod:comp Nominal modifi er [mostly
ablative] − −
nmod:poss Genitive possessive modifi er
nmod:tmod Time modifi er − −
obl:own Owner in -DA − −
In terms of tokenisation, the Turkish tokenisation standard always considers denominal adjectives formed with -/lI/ to be a noun followed by an adposition (i.e., two tokens), while in Uyghur these words are treated as lexicalised adjectives (i.e., as one token). The Kazakh tree- bank varies between these two approaches, currently in a somewhat unprincipled way. Another difference in terms of tokenisation is treat- ment of the so-called -/ki/ affi x, two uses of which are the formation of the attributive locative (-/DAki/ in Turkish, -/DAGI/ in Kazakh) from the locative case form (-/DA/ in both) and the formation of the sub- stative genitive (- /(n)Inki/ in Turkish and -/Nікі/ in Kazakh) from the genitive suffi x (-/(n)In/ in Turkish and -/NIң/ in Kazakh). In all three languages, forms with these “compound” affi xes are annotated with the appropriate relation (amod for attributive locative), but in Kazakh and Uyghur, these forms are not analysed as containing a separate -/ki/
suffi x. That is, in Kazakh and Uyghur, there is a single token, while in Turkish, a separate -/ki/ token has a case dependency to the noun.
One difference in annotation of dependency relations is the treat- ment of the adverbial “quotative word” (Turkish diye, Kazakh деп, Uyghur ﺩەپ). In all three languages, this form, morphologically speak- ing, is a verbal adverb (participle) form of the verb de- ‘say’, though in modern Turkish, the -/(y)A/ participle used in diye is not productively used as a verbal adverb. In Kazakh and Uyghur treebanks, it is analy- sed this way–thatis, it receives an advcl analysis (dependent on the main clause it comes in) and is labelled a VERB, with its relation to the head of its clausal complement labelled as a ccomp. This is shown in fi gure 1b. In the Turkish treebank, however, the speech verb is treated
as an ADP and has a case dependency relationship to the head of the
“quoted” phrase, as shown in fi gure 1a.
There are also a number of differences with how language-specifi c relations are used. Table 2 presents a summary of the language-spe- cifi c relations used in each treebank, and the applicability to the other treebanks.
3. Parsing performance
All three treebanks were included in the 2017 CoNLL shared task on Universal Dependency parsing from raw text data (Zeman et al.
2017). The results for the Turkic languages are presented in Table 3.
LAS and UAS stand for labelled-attachment score and unlabelled- attachment score respectively, while CLAS stands for content-labelled attachment score (Nivre and Fang 2017).
Table 3. Parsing performance in the CoNLL shared task. The column Train indicates the number of tokens in the training data, and the column
Dev indicates the number of tokens in the development data. Note that there were no separate training sets for Turkish and Turkish-PUD; the lat-
ter is only used as a test set for parsers trained on Turkish training data Language Train Dev Winning team (LAS) UAS LAS CLAS Kazakh 0 529 Dumitrescu et al. (2017) 45.72 29.22 25.14 Turkish 38 082 10 011 Dozat et al. (2017) 69.62 62.79 60.01 Turkish-PUD 38 082 10 011 Björkelund et al. (2017) 59.35 38.22 32.32 Uyghur 0 1662 Björkelund et al. (2017) 60.57 43.51 34.07 It is interesting to note the difference between the parsing perfor- mance on the Turkish section of the parallel treebank (Turkish-PUD) and on the IMST treebank (the main UD treebank for Turkish). Both of these treebanks have been converted from other treebank formal- isms: the METU-Sabancı formalism in the case of the IMST treebank and Google Universal Dependencies in the case of the Turkish section of the parallel treebank.
It is also curious as to why the Kazakh and Uyghur numbers are so different despite the data size being similar (that is, while the Uyghur dev set had three times as many tokens, it didn’t have lemmas or any morphology annotation). One explanation could be that there was a lot
more data sparsity when tagging Kazakh as opposed to Uyghur. It’s also striking that Björkelund et al. (2017) got slightly better results for Uyghur than Turkish-PUD, despite a much smaller development corpus size. It should be mentioned that in addition to the training and development data, the shared task included at least 10 000 tokens of both Kazakh and Uyghur testing data.
4. Open questions
In this section we discuss certain phenomena in Turkic languages that present challenges during annotation. Some of those phenomena, e.g. multiple derivation, can be fairly well understood, but are diffi cult to handle adequately given the present UD annotation guidelines. The nature of others may require further investigation within the depen- dency grammar formalism.
4.1 Tokenisation
One of the guiding principles of Universal Dependencies is about its lexicalism (Nivre et al. 2016). That is, words are the basic units of annotation. The guidelines explicitly state that word here refers to syn- tactic words. It is allowed for orthographic words to be split when it is necessary for the syntactic analysis.
In earlier Turkish dependency parsing/annotation work, on the oth- er hand, the words are split at all derivation boundaries, introducing a syntactic word (often called an infl ectional group in Turkish NLP literature) for each derivational morpheme. For example, the Turkish word sınırlandırılabilecek ‘that can be limited’ can be represented as six syntactic words, delimited by ˆDB ‘derivation boundary’, as shown in (1).
(1) sınır+Noun+A3sg+Pnon+Nom ˆDB+Verb+Acquire
ˆDB+Verb+Caus ˆDB+Verb+Pass ˆDB+Verb+Able+Pos ˆDB+Adj+AFuttPart
Although derivation in Turkic languages can be quite productive, as exemplifi ed by (1) above, arguing for necessity of this level of word segmentation is not always practical. Present Turkic UD treebanks seg-
ment the words when parts of the word may have confl icting morpho- logical features and/or parts of the word can participate in different/
confl icting syntactic relations (Çöltekin 2016). In (1) above, none of the syntactic words are necessary since UD morphological features can mark the effect of each morpheme and none of the parts can participate in different syntactic relations–i.e., the parts cannot be modifi ed by other words or ambiguously head other words in a sentence. However, there are also examples where the split is necessary. For example, fail- ing to split the copular suffi x in Figure 2 results in two nsubj relations headed by the same word1. Furthermore, the same word would be as- signed both Number=Plur and Number=Sing features.
Figure 2. Dependency analysis of the sentence Örnek bizim yazdıklarımızdandı
‘The example was from the ones we wrote’. Note how on the verbal form yazdıklarımızdandı there would be two values for Person, Number and Tense
were the copula not split from the non-fi nite form
How to treat very productive derivational suffi xes, which attach to phrases rather than single words, is also a challenge. These include the suffi xes -/LI/ ‘with’, -/sIz/ ‘without’, and -/LIK/ ‘-ness, -ed’, which appear in most Turkic languages. Very many forms that include these suffi xes are lexicalised, for example evsiz ‘homeless’ (lit. ‘without house’), evli ‘married’ (lit. ‘with house’), gözlük ‘spectacles’ (lit. ‘eye-
1 Note that an analysis of this sentence more in line with the annotation standards for Kazakh would have yazdıklarımızdan as the root, with dı as a cop dependent of it, but this again results in the problem of having two nsubj relations headed by the same word. Issues like this are recognised by v2 of the UD standard (cf., http://universaldependencies.org/v2/copula.html) and affect non-Turkic languages as well.
ness’)1, but in some cases, for example бір палаталы ‘one chambered, unicameral’, it would be advantageous to consider the suffi x separately since бір ‘one’ modifi es the stem, not the whole form. There are poten- tially ambiguous examples as well, such as iki gözlük, which can tech- nically mean either ‘two glasses’ or ‘for two eyes’ (e.g., the value of something–though this usage would be rare)2, depending on what level the lük suffi x is interpreted at. One possible solution would be to only split if the word to which the suffi x is attached has its own modifi ers of a certain type, although this sort of structural difference is diffi cult to segment in an NLP pipeline.
Another associated issue is related to syntactic words which contain multiple surface words. An example is the Turkish question marker which, when attached to predicates, may also carry some of the mor- phological features of the predicate. UD currently does not explicitly support syntactic words spanning multiple tokens, though the Kazakh treebank implements some things this way. Some rudimentary solu- tions exist in the present UD scheme–e.g., the goeswith relation or considering a space-separated token a single token–but ad hoc use of these without a standard could cause inconsistencies between Turkic language treebanks. Some issues related to syntactic words containing multiple surface words are discussed further in section 4.3.
Although some general guidelines exist for segmentation of words, there is a need for widely accepted, more concrete rules to ensure con- sistency among Turkic languages, and even among treebanks of the same language.
4.2 Core and oblique
In the UD v2 standard, the dependency relations obj, iobj, and obl are differentiated in the following way: obj is the most core ele- ment of a verb that is not its subject (i.e., a direct object), iobj is the
1 It should be mentioned that it’s not clear that these examples aren’t understood by native speakers of Turkish as compositional and productively formed. Instead, perhaps this interpretation relies on the translation of these words to other languages (a poor criterion!) – it is not necessarily “metaphorical”
(at least historically) that evli should mean ‘married’.
2 A reading that might be found in a wider range of real sources might be
‘two-division’ or ‘two-room’, based on another meaning of göz. In any case, any interpretation of such forms will depend on the context and whether an established lexicalised meaning exists.
next most core element that isn’t a subject or direct object, and oblique is a non-core object. This relies on the notion of a difference between core and non-core elements, or complements versus adjuncts, respec- tively.
In Turkic languages, there does not seem to be a simple and clear way to delineate complements and adjuncts. No element of a verb phrase is absolutely required to be included in a grammatical utterance, not even the subject. While agreement marking will show the existence of a semantically present subject, even if not included in the sentence, Turkic languages do not mark object or indirect object agreement on the verb. Furthermore, since most of the cases have a very wide range of uses, many phrases can be used in any verb phrase, although with a different interpretation depending on the verb.
It seems clear, at least, that typical “accusative direct objects” (and morphologically unmarked indefi nite direct objects) should be anno- tated with the obj relation. However, there is currently only one test that we can use to justify this and other relations: if the element par- ticipates in case promotion or demotion when the verb is made passive or causative, we consider it a core argument, to be labelled with obj if it seems “more core” and iobj if there is another element labelled obj. If the case marking does not change when the verb is made pas- sive or causative, then the element is considered oblique, and receives the obl dependency relation. A more apt solution may exist, but has yet to be identifi ed.
4.3 Complex predicates
In this subsection we discuss verbal (i.e., non-copular) complex predicates. Such predicates consist of two or more orthographic words that together convey single meaning, which is different from meanings (if any) of those words taken separately. Sometimes it is not at all clear how to classify the relationship between the constituents of complex predicates. For instance, a common Kazakh expression пайда бол, meaning appear or be established consists of what appears to be a noun пайда (‘benefi t’, when used on its own as a noun) and a verb бол (‘be’
or ‘fi nish’), which in this particular case loses its habitual copular and auxiliary functions. Thus, a verb that normally takes no arguments1 in
1 In UD copulas are subordinated to nominal predicates for the sake of cross- linguistic consistency (http://universaldependencies.org/u/overview/syntax.html).
this case governs what appears to be a noun; the question is with what syntactic relation.
Depending on the nature of their constituents, complex predicates in question can be roughly classifi ed into three categories: (i) non- verbal + verbal; (ii) verbal + non-verbal; (iii) verbal + verbal. Assum- ing that predicates are fi nite (i.e., non-clausal), in all of these cases the rightmost constituent carries a personal agreement marker (sometimes covert), and in the latter two categories the fi rst verbal constituent is usually non-fi nite1 and contributes to TAM. Also, in all of the cases
‘particles’ and conjunctions may be inserted between the constituents at will – e.g., compare пайда болды ‘it appeared’ and пайда да болды
‘and it also appeared’.
The fi rst category of complex predicates (non-verbal + verbal) in certain UD treebanks (including Kazakh) is sometimes handled as a special sub-type of compounds, namely a light verb construction.
This solution, while possible, relies on meaning, which is undesir- able. There are two alternatives: (i) treat such constructions as a single space-separated token (which in some cases is done in the Apertium Kazakh lexicon); (ii) sacrifi ce the meaning and treat such construc- tions just as normal verb-argument or nominal-copula relations. Both alternatives have pros and cons. While the fi rst one could be accom- modated at the level of morphological analysis and tagging, it is not clear how to handle embedded ‘particles’ and conjunctions. As for the second alternative it just seems wrong to impose literal (usually ab- surd) meaning on otherwise meaningful constructions2.
The second category of complex predicates (verbal + non-verbal) corresponds in Kazakh to negation of fi nite verbs which appears to consist of two tokens. In this construction, the fi rst element (before the space) is morphologically a verbal noun or adjective ending in -/GAн/
and the last element is either жоқ ‘non-existent’ or емес ‘not’ – e.g., айтқан жоқпын ‘I did not say’. Currently in the Kazakh treebank this case is handled at the morphological analysis step, and the construc-
1 The only exception that we are aware of is the non-morphological negation, where (at least in Kazakh) the initial verbal constituent may agree with the subject, e.g. мен олай айтқаным жоқ vs мен олай айтқан жоқпын.
2 Especially if a non-verbal constituent has no lexical meaning on its own and exists only in this sort of expression – e.g., міз бақпа ‘pay no attention’ or місе тұт ‘be satisfi ed’, where міз and місе do not exist as lexical items outside of these constructions.
tion is treated as a single token, as shown in fi gure 3a. It is currently unclear what to do when function words are embedded between the elements of constructions like this. Alternatively this sort of construc- tion could be treated as analytic negation with жоқ or емес being considered negation words and subordinated to the leading verb with the relation advmod:neg, as shown in fi gure 3b. This approach how- ever leads to non-verbal entities carrying personal agreement markers, which is undesirable. Although, this happens in the current conversion of the Turkish treebank with the question word -/mI/ which can carry person/number agreement and tense, as demonstrated in fi gure 3c.
Figure 3. Some examples of Kazakh multi-token negation (‘I didn’t say’) and the current analysis of Turkish multi-token question word (‘Would I say?’)
The third category of complex predicates (verbal + verbal) includes constructions which appear similar to auxiliary verb constructions, but which are probably best thought of as verbal adverb adjunct of main verb. The trailing fi nite verb does not contribute to TAM (tense, as- pect, and mood – which typical auxiliaries in Kazakh convey), and the meanings of these combinations “feel” lexicalised (though it’s unclear whether this is just due to how they translate to other languages). Some examples include болып табылады ‘is found to be’, болып саналады
‘is considered’, атап өтті ‘mentioned’, алып келді ‘brought’, etc. In such constructions the preceding verbs assume a form of -/(I)п/ verbal adverb which can be followed by an auxiliary. The trailing verbs, how- ever, are not always in the closed class of auxiliaries – and the ones that can be auxiliaries do not convey the normally associated auxiliary meaning, such as contributing to TAM. Both verbal elements give a combined meaning to the entire construction, e.g. ата ‘name (V)’ + өт ‘pass (V)’ combine as атап өт ‘mention’. Currently some of these constructions are treated as a single token in the Kazakh treebank due to the fact that they are lexicalised in the morphological analyser used to preprocess text for the treebank, but is designed for use in machine
translation pairs where such lexicalisation is useful. Because of this single-token analysis, the previously mentioned problem of function word embedding exists for these forms. However, other occurrences in the treebank are treated as a main verb with an advcl dependent. One disadvantage of this is that it does not match intuitions about the lexi- cal semantics of these constructions, although, again, perhaps these in- tuitions are based on the translation of these forms to other languages.
Another disadvantage of treating the verbs separately is that in some cases it can result in crossing dependencies, as shown in fi gure 4.
Figure 4. A separate-word analysis of a V+V compound verb in Kazakh, showing overlapping dependencies. The sentence translates to “I brought the book to school.” Here, the verb бар ‘go’ has the oblique dependent мектепке
‘to the school’, while the verb ал has the direct object кітапты ‘the book’
(accusative). Note that a different order of the words, Мен кітапты алып мек- тепке бардым would have a slightly different meaning, ‘I took the book and went to school’ or ‘Taking the book, I went to school’–and may not entail that
the book ended up being brought to school as the depicted sentence does The following facts all point to the interpretation of these verbs operating as a single, compound unit: that both verbs can contribute to the argument structure of the entire phrase, that the semantics are not always entirely compositional (but each verb usually contributes something), that the phrase seems to represent one event and not two, and that the verbs together share a single TAM reading. One possibility for how to deal with this would be to annotate these con- structions with compound (or a subtype of the compound relation, e.g. compound:v), with the fi nite verb governing the non-fi nite one.Another case that could be considered to fall under this category of complex predicates has not actually occurred in the Kazakh treebank, but is fairly frequent in speech: when a Russian infi nitive is followed by the verb ет ‘do, make’– e.g. звонить ет ‘make a telephone call’, обжаловать ет ‘appeal to a higher court’, etc. Due to the introduc- tion of a foreign word, these constructions could potentially be handled
with a special dep relation that preserves structure but bears no partic- ular grammatical function. Thus, the Russian infi nitive could be tagged as X (uncategorised word) and be subordinated to the trailing verb with the relation dep. If a function word is embedded in between, it can be subordinated to the Russian infi nitive with the same relation.
4.4 Multiple derivation
Many linguistic phenomena that are commonly expressed by syn- tactic means – e.g., relations between words – are expressed by mor- phological features in Turkic languages. In particular, Turkic verbs can be infl ected for a range of affi xes expressing features like tense, aspect, modality, voice and subject agreement. The UD specifi cations allow representing most of these features, and the changes in version 2.0 improved this representation considerably. However, in some cases the UD morphology specifi cation is still sub-optimal for expressing some morphological phenomena.
The issue mainly arises when multiple Aspect, Mood and Voice features are present on the same verb. The Turkish examples in (2) include multiple Voice (2a) and Aspect (2b) features on a verb.
(2) a. bekle -t -il -iyor wait CAUS PASS PROG
‘being stalled (=caused to wait)’
b. oku -yuver –iyor read RAPID PROG
‘he/she is reading quickly’
In (2a), the verb has both passive voice and causative voice. Simi- larly in (2b), the affi xes indicate that the action is done quickly, and it is in progress, both of which are typically defi ned as aspect (Gök- sel and Kerslake 2005). While the UD version 2.0 specifi cations al- low marking each of these feature values individually, there is no clear way to mark multiple values for a single feature. In Turkish UD treebank, these words are marked using language-specifi c feature values. For example, the morphological annotation of (2a) includes Voice=CauPass, and the multiple aspect suffi xes in (2b) are indi- cated by Aspect=ProgRapid.
A related issue is repetition of some of these features. Notably, the Turkish causative marker can be attached to a verb multiple times
without a principled limit, indicating a chain of causation1. Similarly, Turkish possibility/ability mood marker can be repeated two times in certain contexts. It is worth noting that this may also arise in non- Turkic languages – see e.g., Ainu in Senuma and Aizawa (2017).
Note that although the above method encodes the relevant infor- mation, it makes it diffi cult for an automated system (e.g., a parser), since the symbol CauPass does not clearly indicate the features Cau or Pass unless special attention is paid for this non-standard notation.
Since some of these combinations are rare2, it is diffi cult for a machine learning method to automatically discover that CauPass is equivalent to having both features marked individually. Similarly, a researcher, for example, looking for causative verbs in the language using a tree- bank search tool will likely to be misled by this ad hoc representation.
4.4 Use of copulas with non-fi nite verb forms
One issue that has arisen recently is how to analyse the use of cop- ulas (as opposed to auxiliary verbs) with non-fi nite verb forms that occur together with auxiliary verbs. There appear to be cases of this in many Turkic languages. Normal non-fi nite form + auxiliary forms are straightforwardly dealt with in UD, as in fi gure 5.
Figure 5. Dependency analysis of an auxiliary construction in Kazakh In both Kazakh and Turkish, a number of fi nite verb forms are formed with what appears to be a verbal noun or verbal adjective form,
1 Although real-life use is often limited, and multiple causative markers often (but not always) indicate emphasis rather than multiple levels of causation. In the Turkish-UD treebank there are no examples of multiple causative, but Turkish- GK includes examples with two causative suffi xes, also in combination with the passive suffi x.
2 Of 9113 verbs in Turkish-UD treebank, Voice=CauPass is the most common multiple-feature marking with 115 occurrences. Others include 5 instances of Mood=CndPot, 4 instances of Mood=GenNec and Mood=DesPot, 2 instances of Aspet=DurPerf, and single instances of Aspect=ProgRapid and Mood=NecPot.
followed by normal copula agreement1. This includes forms like Turk- ish okumuşum ‘I read (past)’ (with perfect form okumuştum ‘I had read’) and okurum ‘I read (non-past)’ (with perfect form okurdum ‘I would read’) and Kazakh оқығанмын ‘I read (past)’ (with perfect form оқыған едім ‘I had read’) and оқырмын ‘I may read’ (with perfect form оқыр едім ‘I would read’). These structures are entirely parallel, although the Kazakh forms have a space between the verb form and the past form of the copula. This construction lends itself to a number of different analyses, as shown in fi gure 62. There are also, in a smaller set of Turkic languages, “auxiliary” constructions that appear to be composed of a copula along with a form that cannot ever be verbal nouns or verbal adjectives and can only operate as a non-fi nite form together with an auxiliary. Because these look more like true auxiliary phrases, it’s clearer how they might be treated. One analysis is shown in fi gure 7.
Figure 6. Two different analyses of an apparent auxiliary construction in Kazakh that consists of a verbal noun or adjective and a copula, glossed as ‘I had read’.
These analyses differ in how they view the copula auxiliary: either as the auxilia- ry in an auxiliary verb construction or as the copula in a normal copula predicate
which happens to have a substantive or attributive verb form as the predicate
Figure 7. One analysis of an apparent auxiliary construction in Kazakh that consists of a non-fi nite form and a copula, glossed roughly as ‘I had read’
1 Note that this is distinct from the obvious verbal nouns in forms like Turkish okumaktayım or Kazakh оқудамын – both meaning “I am reading” (with a long- term or habitual sense), and can be analysed as a verbal noun followed by a locative suffi x, followed by a copula with person agreement.
2 Note that the part of speech of copulas is always AUX in UD.
The set of choices around copula-as-auxiliary constructions in- cludes several options for both dependency relations and morphologi- cal annotation, and Turkic languages have different orthographic strat- egies regarding tokenisation. This issue would be especially good to solve before further annotation.
5. Concluding remarks
We have presented an overview of the current status of the Tur- kic languages within the Universal Dependencies project and drawn attention to a number of inconsistencies that remain. We hope that this serves to inform and direct future research in the area of Turkic dependency parsing and Turkic language technology in general. Af- ter careful analysis we are convinced that the majority of substantial differences in the annotation schemes are a result of conversion from different grammatical traditions as opposed to real signifi cant gram- matical differences.
Acknowledgements
The authors would like to thank the UD community for thoughtful discus- sion and input on a range of issues discussed in this paper. We would also like to thank Deniz Uysal and Tolgonay Kubatova for help with native speaker judge- ments. The work of Aibek Makazhanov was supported by Nazarbayev Univer- sity under the research grant №129-2017/022-2017.
REFERENCES
1. Ageeva, Ekaterina and Francis M. Tyers (2016). “Combined morpho- logical and syntactic disambiguation for cross-lingual dependency parsing”.
In: Proceedings of TurkLang 2016.
2. Aili, Mairehaba, Weinila Mushajiang, Tuergen Yibulayin, A.
Kahaerjiang, and Yan Liu (2016a). “Universal dependencies for Uyghur”.
In: Proceedings of WLSI/OIAF4HLT. Osaka, Japan, pp. 44–50.
3. Aili, Mairehaba, Aziguli Xialifu, Maihefureti, and Saimaiti Maimaitimin (2016b). “Building Uyghur Dependency Treebank: Design Principles, Annotation Schema and Tools”. In: International Workshop on Worldwide Language Service Infrastructure, pp. 124–136.
4. Björkelund, Anders, Agnieszka Falenska, Xiang Yu, and Jonas Kuhn (2017). “IMS at the CoNLL 2017 UD Shared Task: CRFs and Perceptrons Meet Neural Networks”. In: Proceedings of the CoNLL 2017 Shared Task:
Multilingual Parsing from Raw Text to Universal Dependencies. Vancouver, Canada: Association for Computational Linguistics, pp. 40–51.
5. Çöltekin, Çağrı (2015). “A grammar-book treebank of Turkish”. In:
Proceedings of the 14th workshop on Treebanks and Linguistic Theories (TLT 14). Ed. by Markus Dickinson, Erhard Hinrichs, Agnieszka Patejuk, and Adam Przepiórkowski. Warsaw, Poland, pp. 35–49.
6. Çöltekin, Çağrı (2016). “(When) do we need infl ectional groups?” In:
Proceedings of The First International Conference on Turkic Computational Linguistics. Konya, Turkey.
7. Dozat, Timothy, Peng Qi, and Christopher D. Manning (2017).
“Stanford’s Graph-based Neural Dependency Parser at the CoNLL 2017 Shared Task”. In: Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies. Vancouver, Canada:
Association for Computational Linguistics, pp. 20–30.
8. Dumitrescu, Stefan Daniel, Tiberiu Boroş, and Dan Tufiş (2017). “RACAI’s Natural Language Processing pipeline for Universal Dependencies”. In: Proceedings of the CoNLL 2017 Shared Task:
Multilingual Parsing from Raw Text to Universal Dependencies. Vancouver, Canada: Association for Computational Linguistics, pp. 174–181.
9. Göksel, Aslı and Celia Kerslake (2005). Turkish: A Comprehensive Grammar. London: Routledge.
10. Makazhanov, Aibek, Aitolkyn Sultangazina, Olzhas Makhambetov, and Zhandos Yessenbayev (2015). “Syntactic Annotation of Kazakh:
Following the Universal Dependencies Guidelines. A report”. In:
Proceedings of the 3rd International Conference on Turkic Languages Processing, pp. 338–350.
11. Nivre, Joakim and Chiao-Ting Fang (2017). “Universal Dependency Evaluation”. In: Proceedings of the NoDaLiDa 2017 Workshop on Universal Dependencies (UDW 2017), pp. 86–95.
12. Nivre, Joakim, Marie-Catherine de Marneffe, Filip Ginter, Yoav Goldberg, Jan Hajič, Chris Manning, Ryan McDonald, Slav Petrov, Sampo Pyysalo, Natalia Silveira, Reut Tsarfaty, and Dan Zeman (2016). “Universal Dependencies v1: A Multilingual Treebank Collection”. In: Proceedings of Language Resources and Evaluation Conference (LREC’16).
13. Ofl azer, Kemal, Bilge Say, Dilek Zeynep Hakkani-Tür, and Gökhan Tür (2003). “Building a Turkish treebank”. In: Treebanks: Building and Using Parsed Corpora. Ed. by Anne Abeillé. Springer. Chap. 15, pp. 261–
277.
14. Senuma, Hajime and Akiko Aizawa (2017). “Toward Universal De- pendencies for Ainu”. In: Proceedings of the NoDaLiDa 2017 Workshop on Universal Dependencies (UDW 2017). Gothenburg, Sweden, pp. 133–139.
15. Sulubacak, Umut, Memduh Gökırmak, Francis Tyers, Çağrı Çöltekin, Joakim Nivre, and Gülşen Eryiğit (2016a). “Universal Dependencies for Turkish”. In: Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. Osaka, Japan, pp. 3444–3454.
16. Sulubacak, Umut, Tuğba Pamay, and Gülşen Eryiğit (2016b).
“IMST: A revisited Turkish dependency treebank”. In: Proceedings of the 1st International Conference on Turkic Computational Linguistics (TurCLing). Konya, Turkey.
17. Tyers, Francis Morton and Jonathan North Washington (2015).
“Towards a free/open-source dependency treebank for Kazakh”. In:
Proceedings of the 3rd International Conference on Turkic Languages Processing, pp. 276–289.
18. Zeman, Daniel, Martin Popel, Milan Straka, Jan Hajic, Joakim Nivre, Filip Ginter, Juhani Luotolahti, Sampo Pyysalo, Slav Petrov, Martin Potthast, Francis Tyers, Elena Badmaeva, Memduh Gökırmak, Anna Nedoluzhko, Silvie Cinkova, Jan Hajic jr., Jaroslava Hlavacova, Václava Kettnerová, Zdenka Uresova, Jenna Kanerva, Stina Ojala, Anna Missilä, Christopher D. Manning, Sebastian Schuster, Siva Reddy, Dima Taji, Nizar Habash, Herman Leung, Marie-Catherine de Marneffe, Manuela Sanguinetti, Maria Simi, Hiroshi Kanayama, Valeria dePaiva, Kira Droganova, Héctor Martínez Alonso, Çağrı Çöltekin, Umut Sulubacak, Hans Uszkoreit, Vivien Macketanz, Aljoscha Burchardt, Kim Harris, Katrin Marheinecke, Georg Rehm, Tolga Kayadelen, Mohammed Attia, Ali Elkahky, Zhuoran Yu, Emily Pitler, Saran Lertpradit, Michael Mandl, Jesse Kirchner, Hector Fernandez Alcalde, Jana Strnadová, Esha Banerjee, Ruli Manurung, Antonio Stella, Atsuko Shimada, Sookyoung Kwak, Gustavo Mendonca, Tatiana Lando, Rattima Nitisaroj, and Josie Li (2017). “CoNLL 2017 Shared Task:
Multilingual Parsing from Raw Text to Universal Dependencies”. In:
Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies. Vancouver, Canada: Association for Computational Linguistics, pp. 1–19.