This paper responds to the great interest to humanities researchers who are concerned with the study of the Romanian language in its diachronic evolution: developing a set of tools allowing (semi-)automatic interpretative transcription of scanned Romanian documents written in Cyrillic, in print as well as manuscript forms. The corpus contains old data, belonging to the 19th20th centuries, in order to develop an automatic recognition and interpretative transcription of Romanian historical newspapers from Cyrillic (Cy) into Latin (La), in both manuscript and printed forms. We think that the present study will have an important impact the humanities research, including that of paleography, history, archaeology and that field of linguistics interested in the study of the language in diachrony, but it will also help the researchers in the field of computational linguistics that develops models for old language, in order to elaborate a diachronic POS tagger so necessary to recover old lemmata.
It is well known that the operation of interpretative transcription of texts written in Cyrillic is extremely laborious, but it will solve a problem of great interest to humanities researchers who are concerned with the study of the Romanian language (including also Bessarabia texts) in its diachronic evolution.
From the perspective of Digital Humanities (history, paleo-linguistics, to name only a few disciplines), this study is innovative because it will open a huge field of research, making feasible automatic indexing and online content-based search in collections of old Romanian documents. The transcripts produced by the machine will be complemented by linguistic annotations, such that the researcher will be able to seize simultaneously: the original CyrillicRomanian script, its Latin alphabet transcription, annotated elements as in modern language dictionaries (tokens and lemmas), and even elements of grammar, such as syntactic structures, etc.
The novelty of this study includes two major components: developing a diachronic corpus,
called RODICA (ROmanian DIachonic Corpus with Annotations)1, still in its infancy (here,
approximatively 4.5 million words) and defining a method to implement a set of tools allowing
(semi-)automatic interpretative transcription of scanned Romanian documents written in
Cyrillic [
Research will focus on the automatic transliteration of Cyrillic Romanian texts belonging to the 19th-20th centuries, from journalistic genre (written in the Cyrillic alphabet and transliterated in the Latin alphabet). In order to define a methodology to investigate Romanian old language, we consider that the corpus RODICA responds well. It is sufficient to allow an analytical demarche that aims to identify the deviations from the norm that occur in a language, in epochs that are themselves automatically identified statistically.
This study is focused on semi-automatic interpretative transcription according to the way Romanian language written in Cyrillic could be conserved. In this paper, the main objective is the creation of an electronic corpus of old Romanian texts written with the Cyrillic alphabet, from the 19th to the 20th centuries, belonging to journalistic genre in both manuscript and printed types in order to develop an automatic recognition and interpretative transcription of Romanian language tool from Cyrillic into Latin. As training data in the recognition process, we will use 60% of our corpus in Cyrillic alphabet.
The rest of the paper will be organized as follows; in section 2, we mention a few works related to resources and tools related to the analysis of the old language. In section 3, we will state the problem, present our methodology for the developing an automatic recognition and interpretative transcription of Romanian historical heritage writings from Cyrillic into Latin, using the corpus called RODICA. Finally, the paper contains some conclusive statements and suggestions for future work. 2
The future of a language depends on early exposure and on a large number of people who has access to it.
A technology as the one proposed in this project has never been created before for
Romanian old texts. Optical character recognition (OCR) has made remarkable progress in the last
decade, current systems almost reaching the performance of human readers who are ignorant on
the target language. In particular, the character set of the Latin alphabet is recognized at high
rates, which decrease in the case of other alphabets (among them – the Cyrillic one [
Recognition becomes much more problematic in the case of handwritten characters, with their quasi-infinite diversity of forms, being the next phase in our research. Especially problem
atic is the Cyrillic handwriting recognition. Promising results have been obtained recently in
recognizing isolated characters and cursives [
Very few results are known about OCR of Romanian printed with Cyrillic. In [
Our method opens a new perspective for the study of our historical heritage, as conveyed by
Cyrillic Romanian, in both manuscript and printed form, by using full text search technology. It
will enable adaptation of modules that perform linguistic processing: segmentation of the text at
the word level (tokenization), morphosyntactic tagging, syntactic parsing, recognition and
classification of proper names, disambiguation of word senses, and others. The modern deep
learning (DL) technologies, based on neural networks, which allowed us [
An adequate metaphor for this research is a bridge that covers the long way from pixels to
content. Indeed, unstructured grouping of pixels in images representing pages of old
RomanianCyrillic documents will be interpreted and their inner messages deciphered. In order to
acquire a collection of digitised Romanian-Cyrillic resources with their corresponding metadata
and interpretative transcriptions; organise training sets, we will establish the set of norms
(representation formats of intermediary steps in the process of transforming an image of a page
(viewed as a sequence of pixels) into a structured display of textual content. Among these
representation formats, we mention: the metadata describing a source document and XML
representations of the original Cyrillic characters and the final Latin transcribed content (as much as
possible, conforming to TEI [
After we acquire an important collection of digitized resources (printed, semi-uncial and handwritten) containing Romanian language in Cyrillic writings, covering all historical periods, of various conditions of quality (noise level, uneven characters, etc.), with and without supralinear writing, we will add metadata to this corpus, which will provide details about: main language (which must be Romanian), second language(s) (if the document includes words or passages of text in other languages), year of publication, document script (printed, semi-uncial writing, or cursive manuscript), document source (typography), author, level and types of noise (degraded pages, ink stains, creases, dirt, etc.), inclusion or not of supra-linear writing, if there is any critical edition of the text (with indication of source), etc.
Also, in order to build the parallel corpus of original page images in Cyrillic and their interpretative transcriptions in Latin Romanian (UTF-8), we will annotate this corpus with respect to Elements of Content (EoC) in the layout: characters, words, glued words (scriptio continue), lines of writing, paragraphs, supra- and infra-linear writing (words/characters placed above and under lines), border notations, etc. Extract out of this parallel corpus sample images of characters in context, together with equivalent codes, to be used both in training and in evaluation.
To improve the quality of original images, to segment images down to elementary EoCs, to recognize the language different words or sequences of words are written in, to index and search documents based on their EoCs, and to evaluate the recognition processes, we will develop or adapt state-of-the-art visual segmentation software to distinguish EoCs in context. Experiments will be carried out with commercial and open source packages (ABBYY FineReader, for instance). The segmentation software should allow alignment of the EoCs identified in the original digital format of the document and their deciphered textual equivalents. These pointers have a triple role: to link the textual index back into the source document, to support annotations of the expert users in the original document, and to support their corrections related to the interpretative transcription.
We will also train a language recognition system to distinguish among (sequences of) words those belonging to different languages (often used in old Romanian texts: Romanian, Slavonic, Hungarian, Greek, Latin, etc.). This will enable to distinguish foreign words in human translations.
RODICA is a lexical resource developed based on an important newspapers collection [
Importantly, the corpus RODICA represents a first iteration towards building a Romanian Gold corpus, centred on diachronic meta-annotation, and contains over 4.5 million lexical tokens in Latin. The punctuation, the words with less than two characters and the number from the ―Total words‖ have been rNemotoevtehda.t part of this corpus has been transliterated from Cyrillic to Latin, see Table 2.
1817-2015 1829-2015 1829-2015 1837-2015 Total words 643084 959010 1372610 1609230 4583934 Total unique old words 53029 56790 67050 210180
1817-2015 1829-2015 1829-2015 1837-2015
Chirilic words 51084 18010 32610 89230 190934 % (Total
Chirilic words /Total words) 7,94 1,88 2,38 5,54 For illustration, the Figure 1 contains some examples of unconventional writing. For instance, in Figure 1, it can be observed that period does not always mark the end of a sentence, also, it can be noticed that Arabic numbers are used instead of Romans, that a Cyrillic character must be transcribed in two Latin letters, depending on the letters preceding that character. In addition, the capital letters do not always mark the beginning of a phrase or their own name but are often used without a grammatical explanation. There are also missing letters due to mistakes made by the scribe. Ш ш Ш ш Ъ ъ Ь ь Ѣ ѣ Ю ю Ѩ ѩ Ѥ ѥ Ѧ ѧ Ѫ ѫ Ѯ ѯ Ѱ ѱ Ѳ ѳ
We consider that this research responds well both for applicative goals (for enabling effective language chronology analysis using different lexical resources) and for scientific objectives (for exploring the evolution of journalistic language).
Automatic transliteration to the current Latin script and added annotation referring to modern Romanian language are two highly challenging objectives, from both the technological and linguistic points of views, and will open unprecedented research avenues for Romanian scientists and not only.
The success or failure of the study will be estimated according to a combination of the temporal criteria, genre (journalistic) and printed script criteria, as follows: for each historical period of 50 years, a random-per-script sample of 30 pages will be considered.
In the future, we will expand this study for texts belonging to the 16th - 19th centuries, in order to testing the automatic recognition and interpretative transcription of Romanian historical heritage writings from Cyrillic into Latin, in printed as well as manuscript forms.
This survey was published with the support of the PN-II-PT-PCCA-2013-4-1878 Partnership PCCA 2013 grant, having as partners „Alexandurzua‖ IoUanniveCrsity of Iași, SIVECO Romania, and „Ștefan Cel Mare‖ University of Suceava and of the granta- of the tional Authority for Scientific Research and Innovation, CNCS/CCCDI – UEFISCDI, project number PN-III-P2-2.1-BG-2016-0390, within PNCDI III.
Romanian
N Procee