There has been little research on natural-language processing of Buryat, in part due to the absence of language resources. In this article we present the first syntactic treebank for Buryat. It is an evolving project based on the Universal Dependencies (UD) annotation guidelines. We report on the procedure of constructing the treebank and explain some features of the labelling scheme with respect to linguistic phenomena in Buryat. The annotation of 919 sentences was done manually from scratch without using any automatic labelling tools. We evaluate the performance of various language models with UDPipe and describe the plans for future work.
In this article we present the first publicly available treebank for Buryat, the second
official and national language of the Republic of Buryatia, located within Russia
in Southern Siberia. According to UNESCO report, Buryat is considered to be
an endangered language and at risk to disappearing
Although there are no computational tools, Buryat possesses an extensive range
of written-language data, it is the only language in Siberia which has its own
historical records and there are regularly published Buryat newspapers, journals, books,
films, television and radio programmes. The development of language technologies
for Buryat is an essential step towards its revitalisation and documentation
The remainder of the paper is organised as follows. Section 2 gives some background typological information on Buryat. Then section 3 gives an overview of the corpus used for the treebank and how the annotation was done. Section 4 discusses some features of Buryat and how they were dealt with in the annotation guidelines. Furthermore, in section 5 we report preliminary parsing experiments for Buryat using the state-of-the-art NLP pipeline UDPipe. We test different combinations of features. In section 6 we present a summary and suggest avenues for future research and development. 2
Buryat (in Buryat Буряад хэлэн) is a Mongolic language spoken by the Buryats
Buryat people all Baikal lake love.
‘All Buryat people love lake Baikal .’
There are five main dialect groups of Buryat, but the literary standard is based
on the Khori dialect group, spoken to the east of Lake Baikal. This is the dialect
group with the greatest number of speakers
Buryat has seven cases, and two numbers. Nominative case and singular number are unmarked. There are a large number of non-finite verb forms (both participles and verbal adverbs) which are used for clause-level subordination and forming relative clauses. 3
Currently the Buryat treebank consists of 919 annotated dependency trees and 10,146
tokens. The annotation was made using Brat, an online tool for text annotation
udtwenty wikipedia proverbs buryad-unen translation
Grammar Encyclopaedic Proverbs Newspaper Grammar
Statistics about the corpus can be found in Table 1. The text source udtwenty
consists of 20 sentences illustrating different grammatical features available through
the UD project. The wikipedia text source consists of sentences extracted from the
Buryat Wikipedia. Given their public domain status, we also included three Buryat
proverbs from the proverbs source. Finally, after we were given permission to
redistribute them under a free licence, we included articles from the newspaper
Buryaad Ünen and some translations of example sentences from the Technical
Report Syntax Annotation Guidelines for the Turku Dependency Treebank
Having translated text as such a large portion of the treebank has three motivations. The first is that we aim to expand the treebank with more sentences from different domains of Buryat, so this percentage will decrease with time. The second is that we wanted to cover a wide range of syntactic structures, and the third is that we wanted to include translated text as a domain in itself. 3.1
Before annotating in the brat tool, the corpus was preprocessed and manually tokenised. Tokenisation was on space after splitting punctuation characters. The automatic tokenisation for each sentence was fixed manually before the sentence was annotated.
In order to add lemmas to the corpus, we created a lookup table of tuples (surface form, part-of-speech) to lemma. For example: (харандаашууд, noun) (pencils, noun) ! ! харандааш pencil
We then applied this lookup table to the corpus deterministically, for each (surface form, part-of-speech) pair inserting the lemma found in the lookup table. We discovered no ambiguity in lemmatisation. The lookup table is available online under a free/open-source licence.1 4
The annotation guidelines for Buryat were based on Universal Dependencies
In the following subsections we describe some examples from the Buryat
treebank, making emphasis on the most relevant and more typologically interesting
aspects. Thus, for reasons of space, we have omitted a discussion of finite-verbs in
simple clauses as they function more or less as one may expect. Further information
on the annotation guidelines can be found in
As previously mentioned, Buryat has seven cases: Nominative, accusative, genitive, dative,footnoteThe dative case in Buryat has double use as dative and locative. It is sometimes referred to as dative/locative, but in our work we refer to it simply as dative. ablative, instrumental and possessive. All of these apart from nominative are marked with suffix morphology.
The first two can be considered core cases. They have the following syntactic functions: The nominative case is used for the subject and the non-specific direct object of a simple clause. The overt accusative suffix is used for a specific direct object. Regardless of the case used the subject depends on the verb with the label nsubj and the direct object depends on the verb with the label dobj (see Figure 1).
The oblique cases are the dative, ablative, instrumental. The dative and ablative are locational cases used for both spatial and temporal meanings. In addition the dative is used in some non-verbal predicate constructions (see section 4.3) and as the case of the actor in passive constructions, and the ablative is used as the referent of comparison. In all cases these are marked with the nmod label, and when the locative is used to indicate possessor it is marked with nmod:own.
1https://svn.code.sf.net/p/apertium/svn/incubator/apertium-bua/dev/lexicon. bxr.tsv Би pron Case=Nom
I nsubj морин noun Case=Nom a horse dobj
root харанаб verb see Би pron Case=Nom
I
root nsubj
dobj мориие
noun Case=Acc the horse харанаб verb see (a) Non-specified direct object (b) Specified direct object
When formed from animate nouns the instrumental indicates the active participation of a second participant in the action, while when formed from inanimate nouns it can indicate use of a tool, means of transport, time period, etc. In both of these it receives the label nmod.
The genitive case expresses various kinds of adnominal attribution. Genitive modifiers precede their heads. Nouns in the genitive case depend on the nominal they modify and receive the label nmod. The possessive case may be used both adverbially, where it denotes accompaniment, and adnominally where it denotes simple possession. In both cases it receives the label nmod 4.2
Numerals have a regular case paradigm and can be used attributively, adverbially and predicatively. When used attributively to indicate quantity then they receive the label nummod, when used adverbially to indicate a number of times they receive the label advmod and when used predicatively in an equative construction they are the root of the sentence.
root root nsubj nummod dobj nsubj nmod advmod Би pron Case=Nom
I дүрбэн моринүүдые харанаб num noun verb
Case=Acc four the horses see Би Улаан-Үдэ дурбэ ошобо pron propn num verb Case=Nom Case=Dat
I Ulan-Ude four times visited (a) Numeral expressing quantity (b) Adverbial numeral use
There are a number of other uses of numerals, and these are annotated according to their function. 4.3
In Buryat, constructions involving non-verbal predicates (nominals, adverbials) require a copula. This copula (a form of бай- ‘to be’) can be omitted in the present tense (see Figure 3). Constructions with a copula can be split into a number of categories: Equation, attribution, location, possession and existential.
Equation describes sentences such as Борбилоо – шубуун ‘The sparrow is a bird’ (lit. Sparrow – bird) where the predicate is a noun. Attributive sentences such as Figures 3a and 3b are where the predicate is an adjective. Locational sentences have a locative nominal or adverbial as the head (see Figure 3c). In all cases the non-verbal predicate is the head of the clause, and the copula is a dependent on the predicate with the label cop.
(a) Copula omission (b) Overt copula Тэнгери
noun Case=Nom
Sky nsubj
punct – punct
root сэлмэг adj clear
root nsubj
case Ном noun Case=Nom
Book стол noun Case=Dat table nsubj
cop root сэлмэг adj clear байгаа
aux Tense=Past
was Тэнгери
noun Case=Nom
Sky cop дээрэ adp on байгаа
aux Tense=Past
was (c) Locative with copula
Existential sentences have a similar structure to locational sentences, but with reversed word order. In addition, in existential sentences the existential nominals буй ‘existing’, бии ‘existing’ (emphatic) or үгэй ‘not existing’ may be employed. These then become the head of the clause, as in (4b).
Possessive sentences do not have their own structure, and instead there are three patterns for possessives, two of which follow the attribution pattern (4a and 4c) and one which follows the locational pattern (4b). Энэ det This det бууш noun Case=Nom gun nsubj минии pron my det
root нүхэрэйл
noun Case=Gen of friend nmod:own саарһан
noun Case=Nom paper nsubj root үгэй adj not existing (a) Predicative use of genitive (b) Locative possessive Намда pron Case=Dat
To me root
punct discourse discourse һэн part nsubj Та pron Case=Nom
You мүнгэтэй
noun Case=Poss with money (c) Predicative use of possessive гү part qst ?
In the attribution pattern and in bare locative possession (i.e. without бии, буй or үгэй), the item being owned is the head of the clause and the owner is the subject, while in the locative/existential construction, the item being owned is the subject and the owner in the dative case is annotated with the relation nmod:own. 4.4
Buryat has a large number of morphemes for participles and verbal adverbs2 which are marked for tense, aspect and mood.
Participle clauses modify a head nominal, effectively allowing a whole verb phrase to act as an adnominal modifier. This is the way in which Buryat forms relative clauses, so we annotate them with the dependency relation acl, per UD documentation. Examples are provided in Figures 6 and 5. There are also headless relative clauses where the inflection that would normally attached to the head nominal appear on the participle.
2In much of the Mongolic and typological literature verbal adverbs are referred to as converbs. nmod acl
nsubj Намда гэртээ ошохо гэһэн pron noun verb part At me home go уялга бии noun adj obligation existing
Verbal adverbs create adverbial subordinate clauses which may have an independent subject. These receive the label advcl. When they have an independent subject, this is marked with possessive suffix on the verb.
dobj punct amod nsubj discourse Кино буулгахада noun verb Film shooting
, punct горитойхон мүнгэн гаргашалагдадаг гээшэ
adj noun verb part significant money is spent In contrast to the universal dependency rule about headedness in conjunction and multiword constructions, which prescribes taking the first token as the head, in Buryat we take the last token as the head. This is because in Buryat only the last word in the phrase is inflected. In line with treebanks for the Turkic languages, such as Kazakh (Tyers et al., 2015), we decided to annotate these as right-headed. det nsubj appos
nmod name root cop Абань минии Father my noun det
таабайн Гомбо Убушеевичай grandfather Gombo Ubušeevič noun propn propn
дүүниинь little brother noun болоно
is verb
nmod
advcl
root
root
Buryat possesses a great deal of uninflected words, traditionally called частицанууд
‘particles’ which may encode the following categories: negation, mood, tense
I pron Case=Nom nsubj
nmod шамда to you pron Case=Dat root хэлээ say verb
discourse discourse
punct hэн
– part шууб
– part .
. punct
Following the Universal Dependencies guidelines we attach these to the head of the most relevant nearby clause in a flat structure.3 5
In order to test the treebank in a real setting, we evaluated a parser trained with
UDPipe
The options to UDPipe were that we trained a single model for parsing and two models for tagging. The reason for this is that according to the documentation, the system works better if a different models are used for predicting part-of-speech tags and lemmas.
We first held out 19 sentences at random in order to have a neatly divisible number of sentences. Then we performed 10-fold cross-validation by randomising the order of sentences in the corpus and splitting them into 10 equally-sized parts. In each iteration we held out one part for testing (90 sentences) and used the rest for training (810 sentences). We calculated the labelled-attachment score (LAS) and unlabelled-attachment score (UAS) for each of the models. In addition, for those 3We agree with one reviewer who suggested that annotating these with the discourse relation makes it difficult to separate interjections from grammatical function words. In version 2.0 of the treebank these will be annotate these with the relation aux in keeping with the new guidelines. combinations of features that do not include the lemmas and part-of-speech tags from the gold tags we also calculate the part-of-speech tagging and lemmatisation accuracy.
To test how well models trained from the treebank perform, we made a number of configurations of the testing data. The first configuration was just to use the surface forms in the input, with the model providing lemmas and part-of-speech tags. This gives the current state-of-the art results in end-to-end processing of Buryat for both lemmatisation, part-of-speech tagging and dependency parsing.
The second configuration was to use the surface form and part-of-speech in the input. The objective of this configuration was to have a baseline that we could use to compare with the configuration where we include lemmas. The final configuration was to use both lemmas and part-of-speech from the treebank. This can be considered the upper-bound on parsing performance using the treebank.
Surface Surface+POS Surface+Lemma+POS
[78.5, 85.9] [76.0, 78.9] [63.6, 69.5] – – [69.2, 75.4] – – [70.7, 76.3]
LAS The results of the experiments are presented in Table 2. While the results for end-to-end parsing are far from the state of the art, they provide a useful baseline for future work on natural language processing of Buryat.
We are unfortunately not able to compare performance on tools trained on our
treebank with the Mongolian treebank of Loglo et al., (2014) as their treebank is
not freely available, although the same authors in
In our paper we have presented the first dependency treebank for Buryat, and the first publically available treebank for a Mongolic language. It is the fundamental step towards the further development of natural-language processing and language technology for Buryat. During the process a number of discussions about some linguistic phenomena in Buryat were initiated which will be further investigated.
There are a number of avenues for future work. The first is to include
morphological features. Buryat is a highly inflecting language and it has been shown for
highly-inflecting languages that including morphological features makes it possible
to learn more accurate parsing models. To this end we intend to build a finite-state
transducer compatible with the treebank to do morphological analysis. Secondly,
following the release of the version 2.0 guidelines for Universal Dependencies, we
plan to convert the treebank to the new guidelines. Many aspects will be able to be
automatically converted, but we expect to manually revise all cases of ellipsis.
Finally, we intend to synthesise the contents of this paper and MA thesis
The authors would like to express their gratitude to the editors of the newspaper Buryat Unen who gave us permission to include text from their articles for our treebank. We would like to thank Zhargal Badagarov and Dizhyd Markhadaeva who walked us through the most complicated issues of Buryat linguistics. We thank Jonathan Washington for his recommendation of the grammar sources on Mongolic languages and Martin Popel for his help to adjust UDPipe for our language model. We are also thankful to Koldo Gojenola and Gosse Bouma for sharing with us their valuable views on our project. In addition we are grateful to the anonymous reviewers whose detailed comments and suggestions allowed us to improve this article. Straka, M., Hajič, J., and Straková, J. (2016). UDPipe: Trainable Pipeline for Processing CoNLL-U Files Performing Tokenization, Morphological Analysis, POS Tagging and Parsing. In: Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16). Portorož, Slovenia: European Language Resources Association (ELRA). ISBN: 978-2-9517408-9-1. Tyers, F. and Washington, J. (2015). Towards a Free/Open-Source Universal-Dependency Treebank for Kazakh. In: Proceedings of the International Conference ”Turkic Languages Processing” (Turklang 2015). Росстат (2010). Russian Census.