Nowadays, the development of high-quality parallel aligned text corpora is one of the most relevant and advanced directions of modern linguistics. Special emphasis is placed in creating parallel multilingual corpora for low resourced languages, such as the Kazakh language. In the study, we explored texts from four Kazakh bilingual news websites and created the parallel Kazakh-Russian corpus of texts that focus on the criminal subject at their base. In order to align the corpus, we used lexical compliances set and the values of POS-tagging of both languages. 60% of our corpus sentences are automatically aligned correctly. Finally, we analyzed the factors affecting the percentage of errors.
As of today, linguistic resources are not only a part of any linguistics study but an important base for designing any NLP applications. Such resources typically include dictionaries, thesauri, linguistics ontologies, monolingual and multilingual corpora. In order to create these linguistics resources lexicographic researches, analysis of the lexical structure of languages, exploring the text characteristics and similar studies are being conducted.
Design and creation, development and use of high-quality text corpora are one of
the most relevant and advanced directions of modern linguistics [
There are a lot of types of corpora. There are specialized corpora (genre, time, place), general corpora, multilingual corpora, learner corpora, historical or diachronic corpora, monitor corpora and multilingual corpora. Multilingual corpora, in turn, are divided into comparative (comparable corpus) and parallel or the corpus of the translations (translation corpus).
In our opinion, parallel text corpora are particulary important in studying language and features of the translation, various parsing, tasks of speech recognition, etc. For instance, in the tasks of foreign language training, such corpora allow finding possible equivalents of the analyzed lexicon, tracking its values and functions in some contexts.
Furthermore, the concept of the parallel corpus is an integral part of the broader and more difficult concept, such as – machine translation. It is known that machine translation is still the unresolved task of computational linguistics, despite the rapid growth of the various program and empirical resources. In some times the quality of machine translation also depends on the amount of parallel sentences used in training.
For the last decade in the world there was created the set of bilingual and
multilingual corpora, among which, in our view, the most exciting are:
EUROPARL20.000.000 word usage, the open corpus of European Parliament in 11 languages,)
(https://www.isi.edu/~koehn/publications/europarl/); CHEMNITZ
GERMANENGLISH TRANSLATION CORPUS – 1.000.000 word usage
(http://www.tuchemnitz.de/phil/InternetGrammar); KACENKA (Korpus anglicko-cesky; Czech)
3.000.000 word usage (http://www.phil.muni.cz/angl/kacenka/kachna.html); OPUS
(5 languages) (https://aclanthology.info/papers/L04-1174/l04-1174); English-French
Canadian Hansard [
The most prominent and the greatest Kazakh language corpora are: Almaty Corpus
of Kazakh (http://web-corpora.net/KazakhCorpus/search/), containing more than 40
million word usage, 86% of word usage have grammatical analysis; Kazakh text
corpora on Sketch Engine [
At the same time, despite the existence of a large number of parallel multilingual corpora, for low resourced languages, such as the Kazakh language, the task of parallel corpora creating is vital. The task becomes more complex when we say about the development of parallel corpora for not similar languages, the languages from different families. For instance, one language belongs to the Turkic language family and the other belongs to the Indo-European language family, as Kazakh and Russian.
In our study, we explored texts in two languages (Russian and Kazakh) from Kazakh bilingual news websites and created the parallel Kazakh-Russian corpus at the base of these sites' texts. Moreover, texts of our parallel corpus do not belong to fiction or another broad theme; they focus on the criminal subject that makes them limited-field. Therefore, we were able to apply the dictionary method to align the corpus. In addition, to improve the quality of sentence alignment we have made POS-tagging of the texts in both languages and then exploit the labelling. Finally, we calculate the percentage of correct aligned sentences and analyzed the factors affecting the percentage of error.
Parallel corpora contain the text of the original and its translation into some other language. Additionally, these two texts are not just opposed each other, they have to be aligned: particular fragments of the original text have to coincide with the corresponding fragments of the translation. We can say that a parallel corpus is only useful when it is aligned.
In most studies, two levels of alignment are explicitly or implicitly distinguish: sentence alignment and lexical alignment. Generally, the task of automatic comparison of sentences or words in one text to their equivalents in translation is very labor intensive as this consistency between words or sentences is sometimes not “one to one”. For instance, a few paragraphs in source language can correspond to one paragraph in the target language, in translation some words can be deleted or replaced with very distant synonyms or fixed phrases which can be absolutely various in different languages, etc.
We can classify sentence alignment methods into 3 three categories. Methods of
the first category are based on the use of lengths of sentences and paragraphs [
The second group of methods uses lexical information, which can be received from
the corpus [
The third group of texts alignment algorithms in parallel corpora is based on the
POS-tags which are contained in an annotated corpus or use spelling similarity [
However, the use of any method of these groups has a number of inaccuracies and
weaknesses. Accordingly, nowadays interest in the development of systems which
apply the combination of all approaches constantly grows. For instance, Varga et al.
[
Rico Sennrich and Martin Volk [
Another approach to sentence alignment is described in the article [
Vondricka [
Authors of the article [
The creation of aligned parallel corpora faces the additional challenge of the search
of textual resources for a parallel corpus. Nowadays there are a lot of researches
relating to obtaining parallel sentences from non-parallel or comparable data. For
example, such linguistic data can be obtained from Wikipedia. This is an extremely
valuable resource for extracting parallel sentences, as the document alignment is already
provided and Wikipedia articles on the same topic can be in different languages. In
addition, they are connected via “interwiki” links, which are annotated by users [
Consequently, we can make a conclusion that the problem of parallel corpora
alignment has still not been solved up to the end and the universal method has not
been found. Furthermore, to date, most studies consider that the choice of the
alignment method depends on the researched language pair, thematic focus of texts and
types of documents represented in the corpus [
The parallel corpus creation task includes several parts. The first and one of the most important tasks is to collect text material for such corpora. Despite the fact that the Internet contains a large number of websites which are bilingual and multilingual, the choice of needed bilingual resources constitutes an important part of the parallel corpus elaboration. This task becomes more complicated by the fact that we process such different languages as Kazakh and Russian. Additionally, it is necessary to use specialized tools and various techniques not only for language processing but also for collecting necessary material for corpora.
We suggest that parsing of the websites is the best way to automate the process of collecting and saving information. For our study, we have developed our own special software for automatic parsing of the websites, which allow parsing websites that can be similar in design, content and structure.
Four bilingual websites zakon.kz caravan.kz, lenta.kz, nur.kz were chosen for the developed parser. The selected websites represent well-known and reliable news websites of Kazakhstan that are the main news sources on the criminal subject. They contain a large number of articles according to the criminal information, for example, different offences such as robberies, car thefts, murders, car incidents and others, which is one of the goals of our study. Additionally, the websites can switch text information between two languages: Russian and Kazakh.
As a result of a program runtime, we have received the general set of 3000 texts in two languages: Russian and Kazakh. From them, we have selected manually the test set for the creation of the aligned parallel corpus of the Russian-Kazakh texts on criminal information. The corpus size is more than 50410 words, about 24800 of them in Russian and about 25600 words in Kazakh.
On the next step, we determined the structure of the corpus organization. Nowadays such structure can be very diverse, depending on the pragmatical purposes of its creators or users: in the form of the traditional text with reference to the translations, in a tabular "mirror" form that is more convenient for perception and comparison, in the form of the database.
For our study, we chose the database structure as it is the most convenient way for storage of a large amount of data with a possibility of its further increasing.
All news articles are stored in the table of the database which includes their ID, title, the address of the website and the text of an article.
At the following stage, we carried out POS-tagging of the corpus. For a Russian corpus labelling, we chose the pymorphy2 (https://nlpub.ru/Pymorphy) Python packet which is specially developed for morphological analysis of Russian and Ukrainian texts. The libraries of the packet use the OpenCorpora (https://www.pydoc.io/pypi/gensim-3.2.0/autoapi/corpora/dictionary/index.html) dictionary and make hypothetical conclusions for non-recognized words.
In turn, the complexity, structural and typological characteristic of Kazakh marking is connected with the fact that it belongs to agglutinating languages. Structure of this language is rather difficult and unusual, since your native language is inflectional. The agglutinative formation is opposite inflectional where every formant has several inseparable meanings at once (for example, a case, gender, number, etc.). In this reason, we make POS-tagging of Kazakh texts via the regular expression tagger based on RegexpTagger class of nltk Python (https://www.nltk.org/) package. For example, we can identify some types of nouns in Kazakh texts via the following list of regular expressions: patterns=[(r’.*бен$','NN'), (‘r.* пенен $','NN'), (‘r.* басшылық $','NN'), (r’.* іпқону $','NN'), (r’.* тармен $','NN'), (r’.* герлермен $','NN'), (r’.* здар $','NN')] Additionally, to increase recall and precision of our POS-tagging of Kazakh texts we combine regular expressions with the system that includes seven rules. For instance, "If a word followed by words from the special list — the word is marked as Verb". 4
At the first step of the automatic alignment of our corpus, we were guided by punctuation symbols, capital letters and paragraphs. At the next step, it is possible to select two basic approaches to sentence alignment. The first approach that provides significantly higher productivity is based on sentence length. In the second, more resourceintensive approach, the lexical compliances set in by a word alignment method. In our research, the first approach will not yield exact and objective results as the Kazakh language is agglutinating. It means that the form of a word is formed by addition affixes as well as auxiliary additional words carrying semantic and morphological information. In this reason, the use of alignment on the length of sentences or paragraphs of inflectional and the agglutinating languages is not an effective method.
Upon detailed studying of this area, we revealed that for the languages belonging to different language groups and further for specialized, thematic texts it is the best of all to apply the dictionary method of alignment. On the basis of this conclusion, we exploit the lexical compliances set and the values of POS-tagging obtained in previous stages of preparation. The main reason why we were not able to use the first easier approach to sentences alignment is a huge difference between syntax and semantics of Kazakh and Russian languages.
As a lexical compliances set we use our own Kazakh-Russian dictionary, which is based on the English-Kazakh-Russian dictionary that contains about 50 000 entries. Figures 2 shows the fragment of the English-Kazakh-Russian dictionary, which we use as a background one. The dictionary contains about 50 000 entries. Figures 3 shows the fragment of Kazakh-Russian dictionary, which we apply to align parallel Kazakh - Russian corpus.
To improve the sentence alignment, apart from the dictionary method, we use knowledge about the POS-tagging of the words in sentences. Such an approach will allow improving results of the dictionary method as the Kazakh words have several variants of the translations. Thanks to the correct marking of the words in the texts it is possible to improve the best translation equivalent. 5
Our parallel Kazakh-Russian corpus contains texts from certain Kazakh news sites for the period 2018 – 2019. The corpus includes more than 50410 words, about 24800 of them in Russian and about 25600 words in Kazakh.
In order to assess the correctness of aligned sentences, we leverage experts’ opinion, which are native speakers of the Kazakh language as well as the Russian language.
A well-designed special application allows the experts to choose the text in any (Russian or Kazakh) language and automatically load the parallel file of text. When working with a corpus, the expert may mark texts, save them with marking and align them manually. Figure 4 shows the user interface of our special universal application for working with aligned parallel corpora.
As a result of estimating at least of three experts, it was determined that about 60% of sentences in our parallel Kazakh-Russian corpus are automatically correctly aligned. The rest of the sentences need to be aligned manually.
In our opinion, such a percentage is connected to the following factors. 1. Not full coincidence by the number of sentences in corpora. In connection with the complexity of syntactic structures of the Kazakh language, some sentences do not correspond on the structure to Russian equivalents and divide to a few sentences. 2. The complexity of dictionary base creation. The basis of the dictionary method lies in the qualitative-designed dictionary. As the Russian and Kazakh languages are in the distant language groups, during the creation of such a dictionary, there are difficulties with accurate translation. 3. The complexity and limitation of using comparative grammar for the Kazakh and Russian languages in our study. The analysis and further development of this approach will allow improving the result of the alignment. 4. The dictionary method does not consider proper nouns. Texts on criminal subject contain a large number of such words, especially in headings.
All these mismatches can significantly affect the results of alignment. 6
Parallel corpora are used for the meeting of different challenges, such as development and setting the machine translation systems, comparative studying of languages, language training. Development of such corpora is particularly important for lowresource languages and for pairs of languages related to different groups, for example, for Russian and Kazakh.
The developed parallel Kazakh-Russian corpus is created on the basis of four multilingual news websites of Kazakhstan from which the specialized criminal information was selected. The corpus contains 50410 words from which 25600 relates to Kazakh, and 24800 to Russian.
The corpus is aligned with the use of the specially configurated dictionary and knowledge of POS-tagging of both texts. Additionally, the corpus is provided with the special software application allowing adding specialized information to the corpus. The expert assessment of the automatic alignment is 60% of correctly aligned texts. In the next phase of the study, we plan to classify and analyze the mistakes connected with the alignment of the corpus. For that, we will involve a group of philologists of the Kazakh and Russian languages to the professional analysis and assessment of the results.
The developed aligned Kazakh-Russian parallel corpus can be used as training data for machine translation, identification and extraction of the texts connected with crime and for various NLP tasks.
The following step of our study is greater involvement in a stage of the information alignment using POS-tagging which is limited by the complexity of such full marking for the Kazakh texts now. 7
This research is supported by the Committee of Science of the Ministry of Education and Science of the Republic of Kazakhstan (project No. AP05131073 – Methods, models of retrieval and analyses of criminal contained information in semi-structured and unstructured textual arrays).