Latvian Romani is a Northeastern Romani dialect with a limited number of publicly available sources. Two large archival collections of texts in Latvian Romani, compiled primarily in the 1930s in Latvia and Estonia, have been recently digitized as images and made available online for a wider public. In our study, we focus on one of these collections, the Latvian Romani folklore texts collected by Jānis Leimanis in interwar Latvia. In this paper, we describe how initial manual transcriptions, most of which have been created with the help of a special crowdsourcing platform, were integrated in the handwritten text recognition (HTR) workflow in Transkribus. We present two HTR models trained on the basis of Leimanis' collection and discuss various issues related to the work on these texts. Folklore Texts (http://garamantas.lv/en/collection/886320/Romani-folklore-collection-of-Janis-Leimanis), (https://fennougrica.kansalliskirjasto.fi/handle/10024/87064).
This study focuses on the development of a corpus for Latvian Romani (Lotfitka), one of the
understudied Romani dialects. It belongs to the group of Northeastern Romani dialects [
and
Riga the collection by Paul Ariste, a brilliant Estonian linguist, archived in the Estonian literary museum and available online at the
Digitization of manuscripts with field notes in an endangered, understudied and obviously
lowresource language variety using automatic handwritten recognition still represents a rather novel
direction in applying computational methods to such data. Among numerous public models available
in Transkribus, the Evenki-Russian bilingual model trained on Konstantin Rychkov’s manuscripts from
the 1910s seems to be the only example of similar research, see [
2022 Copyright for this paper by its authors.
century ago and not previously published in full. Another important recent initiative focusing on
digitizing old manuscripts with texts in minor languages is Manuscripta Castreniana
(https://www.sgr.fi/manuscripta/), an impressive collection of texts transcribed, translated and
commented by the linguist Matthias Alexander Castrén in the middle of the 19th century, see [
Jānis Leimanis (1886-1950) was a prominent Romani figure of interwar Latvia (see [
Leimanis’ archive in LFK comprises 75 copybooks (three of them currently unavailable), with about 500 folklore units of different genres (463 of them accessible) and 1254 manuscript pages in total. A diversity of genres is impressive: there are about 65 longer narrative texts (fairy tales and stories), a number of songs, short forms (often incorporated in longer texts, but still classified as separate folklore units), description of traditional practices (e.g., Latvian pirts ‘sauna’), proverbs and parables, as well as an appendix with a list of obsolete and disappearing words. This collection is mostly bilingual, as all Romani texts have Latvian translations provided by Leimanis himself: overall, 884 pages contain such bilingual texts. This gives us a unique opportunity to compile a parallel corpus for a Romani dialect.
In 2014, LFK launched a massive crowdsourcing campaign on deciphering the handwritten
manuscripts which were scanned and uploaded to a special platform (http://garamantas.lv/en/). This
initiative became a huge success in terms of the citizens’ voluntary engagement into the preservation
of cultural heritage (see more in [
The orthography used by Leimanis for Romani texts is based on Latvian; no extra symbols or diacritics (for instance, to mark stress) are used. Leimanis’ handwriting is very clear, and the text is almost always very accurately adjusted to the lines. No doubt, such handwritten texts make an ideal case for automatic recognition.
One of the well-known and freely available HTR tools is Transkribus
(https://transkribus.eu/Transkribus/). It is possible to use it as a desktop application with the access to
the server or work directly in the browser in Transkribus Lite. The platform can be used for transcribing
texts and training various text recognition models on them. Transcription can be manual or automatic,
conducted with the help of the existing models (OCR models for printed texts are also available for
users), see more details in [
In January 2021, we intensively worked on transferring the deciphered texts to the Transkribus platform as ground truth data for the Leimanis collection. Preliminary transcription experiments were conducted already in 2020. At the initial stage, the texts were taken from finished garamantas.lv transcriptions, with some minor editing in most obvious cases. The almost perfect quality of the transcriptions available at garamantas.lv accelerated the process of ground truth preparation. The major difficulty was related to the distribution of transcription blocks between the lines, as Transkribus requires a very detailed image-based reproduction of texts in transcriptions. In contrast, the guidelines for crowdsourcers ask for normalized form of erratic words to be put in square brackets. This information is not based on the scanned images, and such additions had to be removed from the transcriptions.
Even though this work was initially conducted just by one person, a trained linguist with only
theoretical knowledge of Latvian Romani, we hope that access to the resources on this and other
dialects, most importantly, the dictionary [
Due to the standard format of the copybooks used by Leimanis, the standard size of the files is about 20–34 lines (usually two-page layouts with the Latvian Romani text on the left and the Latvian translation on the right side). It takes about 6–10 minutes to copy and paste a previously deciphered text and distribute it correctly across recognized lines, as well as to correct minor things (e.g., add strikethrough annotation for some words which were not relevant for the transcribers, but which are more important for the HTR training in Transkribus). Figure 2 shows what the text from Figure 1 looks like after having been transcribed.
It is claimed that training a good Transkribus model requires about 15000 words, or 75 pages, of ground truth material. In our case, an average file with a two-page layout has about 160 words. This gives the following preliminary calculation: at least 94 files (15000/163) are needed to train the HTR model. As there were more deciphered files available at garamantas.lv, it was decided to add at least 200 deciphered pages as ground truth transcriptions.
The first model, Leimanis_test, was trained on 212 pages and validated on 23 pages. About 30 pages in the training data are copybook covers, which have only several shortened lines and describe metadata on the copybook content. The second model, Leimanis_updated, has the first model as its base model. Only several pages were additionally transcribed and used as ground truth data. The proportion of train and validation data was shifted to the improvement of the latter in terms of size. Still, using the base model and a higher number of epochs (100 in the second model vs. 50 in the first model) resulted in a considerable improvement of quality, as character error rate (CER) was already below the threshold of 5%. Both models were trained with the CITlab HTR+ method. The comparison of our models is given in Table 1 below.
A preliminary layout analysis had been previously performed with the help of the CITlab Advanced tool with default settings. For ground truth pages, it was always manually corrected, when necessary, but for many pages it is still left uncorrected: this is seen as a preparatory step before each automatic page recognition procedure, so it is conducted consecutively.
After obtaining the HTR model and transcribing the rest of the files automatically, one should have a look at the quality of the obtained transcriptions. There is some chance that the initial quality of the model would not be enough, then adding more manual transcription could probably improve the training. In any case, in order to continue work with the texts, one needs to correct all possible errors in automatic transcriptions. The project page gives information on character error rate about 5% for some of the available models, which, of course, requires some postprocessing of the automatically annotated files.
For evaluation of the quality of recognition of the two available models and their comparison, we launched HTR on a page outside the ground truth data (page 1245 in our collection, originally notebook 68, page 53). It is a fragment of the unit 426 (http://garamantas.lv/lv/unit/400700/LFK-1389-426), a fairytale ‘How a baron married a Roma girl, and how she preferred forest and her people’ (“Kā skaisto čigānieti apprecēja barons, un ka tai mežs un tautieši labāk patika...” / “Sir šukārune romane čha lija baronos, te sir lake, vešs te roma fidīr kamdžapes nasir filačin te...”).
Figure 3 shows the result of automatic layout analysis (CITlab Advanced with default settings). The text in notebooks is very consistently and clearly written line-to-line, and corrections are minor, in this case only as a superscript word in line 5 on the right page. The layout recognition works very well, and in this case all the errors are minor and expected and belong to the regularly occurring types. First, the very first line is split into two parts; this happens occasionally, but still not too frequently to noticeably disturb the process of post-correction. Second, superscripts are usually assigned separate lines in automatic layout analysis. This is not incorrect from a purely visual perspective, but for structural reasons we want to incorporate such superscript words in the lines where they truly belong by merging them with the next line. As Transkribus provides additional markup for superscript and subscript (also for bold, italic, underlined, and strikethrough, among others), the post-corrected result can be easily used for training next models and for the restoration of the original. 3Here by pages we mean double pages as scanned and uploaded to garamantas.lv; page numbers are also reflected in the file names (1389-68-05 at the website and 1389_68_05.jpg in our collection).
See the tutorial available at https://readcoop.eu/transkribus/howto/how-to-train-a-handwritten-text-recognition-model-intranskribus/ (accessed on 15/02/2021).
WER 2.70 2.36 4.73 5.41
Recognition errors are often related to diacritics either in the Latvian or the Latvian Romani text; sometimes diacritics can also be incorrectly added, as in our case with sūne (incorrect šūne). In addition, some problematic cases are really challenging, as they are related to less standard letter combinations occasionally made in handwritten texts (e.g., when one letter is distorted, or when borderlines between adjacent letters are not as clear as they normally are). Sometimes, capitalization errors are made, if Leimanis doesn’t make clear size difference between the first capital letter and the following one (see the error Viņa / viņa in Figure 4). Finally, sometimes different letters are indeed similar in their graphic forms, therefore automatic recognition models make mistakes like with human recognition, as we as transcribers also sometimes struggle, trying to decipher some letters; in our case, the capitalized F in the word Fidir got interpreted as either L or T.
It has been shown how the opportunities provided by Transkribus can be successfully exploited with respect to unique handwritten Latvian Romani data. Initial volunteer-based manual deciphering efforts have been successfully integrated in the development of automatic handwritten recognition for Latvian Romani; we are thankful for being granted access to such preliminary ground truth data.
The character of handwriting and accurate original copybook layout made it possible to train recognition models, the best of which has very satisfactory quality, so that the further deciphering process is based on post-correction of automatically recognised pages. Errors made by the model Leimanis_updated are not numerous and in many cases can be compared to similar difficulties faced by ordinary transcribers.
We hope that our initial effort at digitizing Leimanis’ archive will result in the development of a bigger Latvian Romani corpus. This would, of course, imply more serious normalization of texts, additional translations (at least in English), developing morphological annotation for Latvian Romani, etc. The bilingual character of Leimanis’ collection makes it possible to consider the compilation of a parallel corpus, with such options as, for instance, word alignment. It is also worth mentioning that for several texts later translations are available in [15] and in the materials from Pasakas.net5. Various Latvian Romani materials are already available in our repository (https://github.com/LatvianRomani/Lotfitka).
We are thankful to Sanita Reinsone for giving access to the collection and kind support and to Ieva Tihovska for her help with text transcriptions and sharing her knowledge about the history of Latvian Roma people and Jānis Leimanis in particular. We would also like to recognize the contribution by Anette Ross who helped us with proofreading the automatically recognized transcriptions and gave us access to various texts in Latvian Romani.
5As the original domain was taken by other people, a duplicated archival copy is currently stored at asakas.net. Heyer, L. Hirvonen, T. Hodel, M. Jokinen, Ph. Kahle, M. Kallio, Fr. Kaplan, Fl. Kleber, R. Labahn, E. M. Lang, S. Laube, G. Leifert, G. Louloudis, R. McNicholl, J.-L. Meunier, J. Michael, E. Mühlbauer, N. Philipp, I. Pratikakis, J. Puigcerver, Pérez, H. Putz, G. Retsinas, V. Romero, R. Sablatnig, J. Andreu Sánchez, Ph. Schofield, G. Sfikas, Chr. Sieber, N. Stamatopoulos, T. Strauß, T. Terbul, A. Héctor Toselli, B. Ulreich, M. Villegas, E. Vidal, J. Walcher, M. Weidemann, H. Wurster, K. Zagoris. Transforming scholarship in the archives through handwritten text recognition: Transkribus as a case study, Journal of Documentation 75, 5 (2019) 954-976. doi: 10.1108/JD-07-2018-0114 [15] S. Brice (Ed.), Laimes puķe: čigānu tautas pasakas, Sprīdītis, Rīga, 1992.