=Paper=
{{Paper
|id=Vol-3160/paper2
|storemode=property
|title=New Information Extracting and Analysis Methodology for the Terminology Research Purposes: The Field of Biology
|pdfUrl=https://ceur-ws.org/Vol-3160/paper2.pdf
|volume=Vol-3160
|authors=Gints Jasmonts,Silga Sviķe,Karina Skirmante
|dblpUrl=https://dblp.org/rec/conf/ircdl/JasmontsSS22
}}
==New Information Extracting and Analysis Methodology for the Terminology Research Purposes: The Field of Biology==
https://ceur-ws.org/Vol-3160/paper2.pdf
New Information Extracting and Analysis Methodology for the
Terminology Research Purposes: The Field of Biology
Gints Jasmonts, Silga Sviķe and Karina Šķirmante
Ventspils University of Applied Sciences 1, Inženieru Street 101, Ventspils, LV-3601, Latvia
Abstract
This study is part of a larger research developing a database of organism names for preserving
linguistic diversity and for terminological studies. This article addresses the issue of retrieving
the names of organisms from digitized resources of specialized literature for subsequent
collection in the new database. The study used image processing algorithms, Optical Character
Recognition (OCR) and language recognition algorithms, altogether creating a methodology
that reflects the possibility of using data retrieval for the terms of the field of biology in Latvian
and other local languages – the names of organisms along with their scientific or Latin name.
Further on, these retrieved data are scrutinized and prepared for addition in the developed
database.
Keywords1
Information extracting, organism names, database, digitized resources, OCR methods
1. Introduction
Although currently there is a certain range of data available in various digital resources for
researching the diversity of the Latvian language, for example, Latvian text corpora (www.korpuss.lv),
they are often not useful for terminology research. The corpus included on this site – Balanced Corpus
of Contemporary Latvian Texts – contains various texts with metadata, but it comprises only 10 million
words and the ratio of scientific texts included in the corpus accounts for only 10% [1]. Of these 10%,
only 2.5% are devoted to the field of biological science, and the authors of the corpus point out that this
corpus is mostly developed for grammar research [2]. For example, when searching for the
terminological name of the plant genus Zinnia in the Latvian language cīnija (zinnia), only three sources
with the use of the name of this plant genus are found. In two sources from regional and specialized
periodicals – in the newspaper “Latgales Laiks” published on August 12, 2016 and in the gardening
magazine “Dārza Pasaule” published in 2012, as well as in a work of fiction – the book “Sniega laika
piezīmes” [“Snow Time Notes”] by Inga Ābele, published in 2004. Terminology research options in
this corpus are also limited by the inclusion of only modern sources in the corpus database – i. e. starting
from 1991, therefore it is not possible to compare changes in terms over a longer period of time. The
scientific or in other words Latin name of this plant is not found in this corpus.
When searching for the Latvian name of the same plant genus name – cīnija in the largest electronic
dictionary site available in Latvian – Tēzaurs (www.tezaurs.lv) it can be concluded that the entry for
cīnija [3] gives a definition and two species names of this genus, and three different sources are given
below, from which this information is compiled from – one is from the volume of the “Latviešu literārās
valodas vārdnīca” [“Dictionary of the Latvian Literary Language”], one is from the volume “Latvijas
enciklopēdija” [“Encyclopaedia of Latvia”] by the publishing house of Valērijs Belokoņs and one of
the volumes of the encyclopaedia “Latvijas daba” [“Nature in Latvia”] [3], however, the affiliation of
examples to the source used in this entry cannot be accurately determined and therefore all these sources
have to be considered separately by the researcher.
IRCDL 2022: 18th Italian Research Conference on Digital Libraries, February 24–25, 2022, Padova, Italy
gints.jasmonts@venta.lv (G. Jasmonts); silga.svike@gmail.com (S. Sviķe); karina.krinkele@venta.lv (K. Šķirmante)
0000-0001-8374-3237 (S. Sviķe); 0000-0002-2906-4660 (K. Šķirmante)
©️ 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
When searching for the Latvian plant name of the genus Zinnia, National Terminology Portal
(https://termini.gov.lv/) offers an entry from the Terminology Commission of the Latvian Academy of
Sciences, the 9th term collection “Agronomijas terminu vārdnīca” [“Agronomic Term Dictionary”] [4],
which indicates the Latvian name of the plant genus Zinnia – cīnija that is approved by the Terminology
Commission. The metadata to this resource are accurate, however, it is also the only resource included
in this portal with the name of that genus of plants. In practice, for example, the offer of the company
‘Kurzemes sēklas” shows that seeds of these plant species are sold under the Latvian name of cinnija
[5]. Also, in other specialized literature the Latvian name of the plant species Zinnia elegans is
mentioned as gleznā cinnija (common zinnia or elegant zinnia) [6, p. 50]. In this case, it would be
valuable to find out the usage frequency data of one or another variant of the Latvian name, as well as
their fixation in different types of resources. For example, in Pauls Galenieks’s work “Botāniskā
vārdnīca” [“Botanical dictionary”] for the plant genus Zinnia, the given Latvian name is cinijas [7,
p. 15]. These aggregated data already show that this Latvian name has changed over time. Such
terminological studies require a specific selection of data. For example, research on the use of botanical
or zoological terms, their changes or statistical research requires the use of diverse publications:
scientific articles, periodicals, popular science articles, teaching and specialized literature, reference
literature, laws and regulations, but these data are dispersed in different resources. Whereas, diachronic
and etymology research requires further use of other digital resources, such as the ancient text corpus
[8]. However, it only summarizes the texts of the 16th–18th century, therefore printed resources, as well
as other additional digital resources, should also be used, including the data set in articles from the 19th
and 20th centuries. In such studies, it would be important to use as diverse research material as possible
in order to obtain a more accurate result of the study.
Only one example of an ornamental plant’s name is described above. However, if one were to search
for the terminological name of a medicinal plant, where human life might depend on the result, problems
could occur. The principle of unambiguity in the terminology of Latvian plant names has been justified
by Inese Ēdelmane [9]. This part of the lexicon is very important and it is crucial to select, structure and
make it accessible, that would add significantly to the range of digital library resources. Other studies
are also available on the wide scope of this special part of the lexicon, which makes the work of, for
instance, lexicographers, more difficult [10].
There is no unified database for research of Latvian names for organisms, but Ventspils University
of Applied Sciences, in cooperation with the Institute of Horticulture, is currently developing a special
database for the research of the names of organisms (hereinafter IMDS) for this purpose. The new
database will be with specific emphasis on the Latvian language and for research purposes, providing
both historical information of the usage of lexemes, as well as terminological data source for a wide
range of interested parties – translators, linguists and lexicographers, terminologists, authors. The
terminological part of the database will be most important, and it will contain relations among scientific
names and equivalent Latvian names. The structure, data retrieval and results of the database will be
reported in future studies by the authors. Since the project will terminate at the end of 2023, the current
scope of research is related to the automation of data entry. Most of the data (scientific names of
organisms and their equivalents in other languages) are entered manually, but solutions for automated
data addition are also considered. One of the solutions is the present study that proposes the processing
of material with a structure of the dictionary and adding the data extraction to the database. The data
retrieved by using the tool developed in this study shall be stored in this database, thus enabling the
creation of a specific set of data for research purposes faster. The following chapters of this article
provide an analysis of the developed tool set and a description of the methodology.
2. Research methodology
2.1. Characteristics of the Problem and the Tasks of the Study
The National Library of Latvia (hereinafter LNB) provides an opportunity access books not only to
in printed version, but also in electronic form using the Latvian National Digital Library (hereinafter
LNDL) tool (http://gramatas.lndb.lv/). The use of the above mentioned tool allows to create an IT
solution that automatically processes LNB resources to supplement the database of organism names
that is currently developed. The LNDL tool uses the Optical Character Recognition (hereinafter OCR)
[11] algorithm to digitize scanned books, but often the used OCR algorithm has failed to convert the
scanned image into a 100% correct text (see Fig. 1). Figure 1 shows fragments of the book “Augstāko
augu sistemātika” [“Systematics of Higher Plants”] [12], which will be used further on to describe the
methodology.
Figure 1: (a) Fragment of a scanned page of the book “Systematics of Higher Plants”; (b) result
retrieved by LNDL tool.
Considering the inaccuracy of the results returned by the LNDL tool, own set of algorithms was
developed for the study in order to automatically identify, process and store the names of organisms in
the created database of organism names (IMDS) mentioned in the books provided by the LNDL tool.
Following tasks were carried out in order to achieve the defined objective:
• Automated data retrieval i. e. retrieval of scanned book pages from the LNDL
(http://gramatas.lndb.lv/);
• Preparation of scanned images of books for processing by an OCR algorithm;
• Processing images with an OCR algorithm that allows to retrieve information in text
form;
• Evaluation and comparison of the results obtained and the results offered by the LNDL
tool;
• Identification of the Latvian and Latin organism names in the obtained text.
Since the terminology of the field is not explicitly marked in the specialized literature, as is the case,
for example, in the general translation dictionaries, where the terms are accompanied by the
abbreviation of the industry designation: bot. – botany, zool. – zoology, etc., the developed solution
should identify the scientific or Latin names of organisms that are often added to the respective
organism names in texts from the field of biology. It is important to retrieve organism names in local
languages (Latvian, Russian, German, etc.) by identifying them together with the given Latin names.
2.2. Used Technologies and Workflow of the Developed Solution
The developed methodology consists of five stages (see Fig 2):
(1) Book pre-processing stage, where the book is scanned as a pdf file and automatically converted
to png files where each page is a unique png file. To retrieve LNDL book page representation in image
format, Selenium architecture was used through Python library.
(2) scanned page processing stage, where multiple image processing methods are applied to the png
images, for example, the background of the image is filtered using Otsu thresholding, then the small
details of the image are retrieved using erosion and expansion operations, later the image quality is
improved using erosion, reconstruction and closing morphological operations. For this stage, a pre-
processing tool was developed in Python including Skimage, OpenCV2 and numpy libraries. Google
Tesseract OCR technology which is based on LSTM (Long short-term memory) artificial recurrent
neural network is used to process obtained png files and to digitize information from scanned png files
into text document files where processed results are organized in multiple columns related to specific
language. In this research, Tesseract OCR versions 3 and 4 were used. To combine all operations under
one language, the PyTesseract and tesserocr libraries were used to enable the use of Tesseract using the
Python environment.
(3) Language recognition stage, where obtained results of all names from the stage 2 are classified
by Language identificator LangId. In this research only three data sets of languages are used – Latvian
(LV), Latin (LA) and Russian (RU). Developed script in Python manages the language identificator
process for each organism name in the text document, identifies unclear results and stores all obtained
results in the excel data file, where information about organism names are stored with information about
language of organism name and commemoration page number.
(4) Result post-processing stage, where the user manually post-processes all results in the excel file.
In this stage a developed tool in Python is used to integrate obtained results in the IMDS database.
(5) Data storage and displaying stage (IMDS), where users can view and process the integrated data.
IMDS system consists of three main parts – (1) the database where MySQL technology is used, (2) the
backend where Spring Framework technologies are used, and (3) the frontend where React JS
technology is used.
Figure 2. Used technologies and workflow of the developed solution.
2.3. Processing and Digitization of Scanned Images
For more accurate results, primary processing of the images had to be carried out before performing
the OCR operation. Depending on the problem, the actions described further on might vary, but the
objective outlined in this study is to obtain a binary image with as little light gradient effect as possible.
Ideally, the pixels belonging to the text in the image array would be marked with true values, and the
background – with false values.
Given that scanned books do not always have a homogeneous background and there are various
noises from lighting and material, it is necessary to use additional processing algorithms to create a
binary image. A popular solution to these problems is the use of morphological methods of images
[13] – moving the structural element through the image matrix, performing dilation or erosion
operations at each step. These basic operations can be combined to create new ones, such as opening
and closing operations. The methodology provides an opportunity to improve the quality of letters,
retrieve the small details of the images, as well as to improve the possibility of higher quality image
thresholding.
Thresholding in the context of image processing means evaluating pixels to determine if they meet
criteria specified by the user. Since colour information is not as important for this task, it is more
appropriate to use the intensities of grey tone values. In simple cases, it is possible to specify a threshold
chosen by the user, but given that the scanned pages have different lighting, it is necessary to
automatically calculate the threshold values. The algorithm uses the Otsu [14] thresholding method,
which calculates the threshold based on keeping the sum of foreground and background intensity
dispersion to a minimum.
Figure 3 (a) Example of a scanned page; (b) result obtained by image processing algorithm.
Examples in the Figure 3 and the first results obtained in the study lead to the conclusion that the
result obtained by the algorithm is clearer and that the text contains clearly visible words and language
units to be retrieved (see Fig. 3).
For the OCR algorithm, the Google Tesseract technology was used because of its promising results
in the similar use cases related to text digitization from books [15] [16][17]. Obtained results showed
that Tesseract technology usage was good enough to identify the names of the organisms also in this
case, so no other OCR tools were used for further research, however Keras-OCR, Amazon Textract,
Google Document AI and EasyOCR should be mentioned as options.
After the image processing stage, the obtained result is a binary image, which is then automatically
processed with the Tesseract OCR tool to retrieve textual information that can be automatically
processed retrieving the organism names mentioned in the text. Figure 4 below shows the result
obtained by the LNDL tool and the result of the algorithm developed in this study.
(b)
(a)
Figure 4: (a) Example fragment of a scanned page; (b) result obtained by image processing and OCR
algorithm: above – LNDL tool, below – algorithm developed in this study.
Figure 4 shows that the developed algorithm makes a good distinction between graphemes, whereas
the results of the LNDL tool are converted, not relevant and cannot be identified.
2.4. Recognition of Language for Retrieved Words
Next step is to process this obtained research material with a language identification algorithm. For
this step multiple tools for language detection/recognition were used, for example, (1) Natural
Language Toolkit with Language ID module using TextCat algorithm; (2) pycld3 package which
contains Python bindings (via Cython) to Google’s CLD3 library; (3) Polyglot which depends on
pycld2 library and on cld2 library for detecting language(s) used in plain text; (4) TextBlob that
provides a consistent API for diving into common natural language processing (NLP) tasks such as
part-of-speech tagging, noun phrase extraction, sentiment analysis; (5) LangId [18] that contains a
trained neural network and a 97-language trainee dataset. LangId showed the best result for Latvian and
Latin languages recognition and it was used for future applications. In this research, methods that do
not require additional training were chosen, making them more versatile / use-case independent.
The LangId tool can be retrained according to the needed precision in the execution of a specific
task. To identify the conformity of the specific organism names in Latvian, Russian or Latin included
in the text fragment, multiple data sets were used to examine the methodology – the text in Latvian, the
text in Latin, the text in Russian, the text containing Latvian and Latin words, the text containing all
three languages – Latvian, Latin and Russian. The identification algorithm itself is based on the neural
network which is trained using corpus data. Various field specific texts have been used in Latvian,
Russian and Latin to train the algorithm.
The methodology is represented in the examples in Figures 5, 6, 7 and 8, showing instances of the
text fragments, thus reflecting the effectiveness of the methodology and the processing results obtained
for specific test data. The results provide a probability of matching each word in Latvian, Russian and
Latin in each of the data sets. Probability is shown in the 0–1 scale where 1 corresponds to 100%
confidence matching.
Figure 5: Application of the methodology developed in the study data set with text in Latvian: (a)
Fragment of a scanned page; (b) Probability of compliance of each digitized language unit in Latvian
and Latin.
Figure 6: Application of the methodology developed in the study to a data set containing only Latin
text: (a) Fragment of scanned page; (b) Probability of conformity of each digitized language unit with
Latvian and Latin.
Figure 7: Application of the methodology developed in the study to a data set from P. Galenieks’s work
“Botanical Dictionary”: (a) Fragment of scanned page where plant names organized in three columns
– Latvian, Latin and Russian; (b) Probability of conformity of each digitized language unit with Latvian,
Latin and Russian.
Figure 8: Application of the methodology developed in the study to a data set containing text in
multiple languages: (a) Fragment of a scanned page where data set containing text in Latvian and
Latin; (b) Probability of conformity of each digitized language unit with Latvian and Latin; (c) Fragment
of a scanned page where data set containing text in Latvian, Russian and Latin; d) Probability of
conformity of each digitized language unit with Latvian, Russian and Latin.
As shown in Figure 5, when working with a data set containing words only in Latvian, the algorithm
provides a 100% correct match in Latvian, thus it would be possible to process Latvian language
dictionaries and automatically save the retrieved language units in a developed database. Results
obtained in the example displayed in Figure 6 show that only 80% of Latin words are correctly
identified, so it would not be possible to save this data to the database automatically, since verification
and manual editing of the results is required. For all the data that has a probability of compliance below
a certain threshold (e. g. 80%) a manual check would be applied as a post-editing technique. In the
example in Figure 6, the first word anaplasma (anaplasma) is identified as a Latvian word with a
probability of 93%, although anaplasma is a Latin origin word. It is possible that the algorithm has
linked it to the Latvian plazma (plasma), which differs from the Latin word only by one different letter
– z. When testing the algorithm with other text fragments in Latin, the accuracy in several cases was
even lower than 80%. Another reason for the algorithm’s inaccuracy could be that a Latin data set
containing few texts on biology and botany was used to train the neural network.
Working with a data set that includes Latvian, Russian and Latin languages, plant names in Russian
were identified with approximately 100% precision as it is shown in Figure 7. This is how the Zinnia
example described in the introduction of the article is also represented in the book. However, the Latin
name has been incorrectly determined by the algorithm, even though there is a 4 character difference
between the Latin and Latvian names. Despite the algorithm identifying plant names in Latvian with
high precision, it also identified some of the plant names in Latin as plant names in Latvian, and it can
be trained using biology data sets in Latvian and Latin, however this specific set of biology data may
not be sufficient enough for neural network training. [19] The developed algorithm works correctly
with plant names in Russian and Latvian, but since Latin and Latvian use the Latin alphabet, the
algorithm needs improvements for the distinction of both languages.
Figure 8 (a) and (b) shows an example of text that contains both Latvian and Latin words, and the
results are similar to those of the previous two cases – the algorithm has identified Latvian words with
100% accuracy, while the Latin words have a 50% match accuracy. Obtained result is also similar when
processing other equivalent text fragments – the word match accuracy for Latvian is close to 100%, but
the word matching accuracy for Latin is 30% to 50%. After processing the third data set, which contains
words in both Latvian and Latin, it must be concluded that it is necessary to retrain the neural network
with a larger set of training data in Latin, which would include more texts from scientific literature in
the field of biology.
Figure 8 (c) and (b) shows an example of the text that contains Latvian, Russian and Latin words. If
the OCR algorithm recognized the scanned image into its correct digitized version, words in Russian
were identified correctly. Language identification algorithm working in the Russian language is
promising because the probability of errors is low in all cases. It can be explained by the Russian
alphabet being unique among all used languages in this study and it is easy to differentiate because of
the Cyrillic symbols. However, words in Latin and Latvian could be confused. Currently the language
identification algorithm works best when the alphabets are different.
3. Research Results
The methodology developed in the study and described above was tested using the book
“Systematics of Higher Plants” by V. Langenfelds, E. Ozoliņa and G. Ābele, and the test results show
that the methodology can successfully digitize poor quality scans and identify the conformity of the text
in Latvian, but there are problems with the identification of Latin text. Given the specific features of
this methodology, the application of the methodology could work better when digitizing dictionaries,
because words are arranged in columns there and matching could be done for the respective languages.
The dataset in the books selected before was too small to draw convincing conclusions for testing the
methodology, so another book was analysed for the study.
Data processing from the P. Galenieks’s work “Botanical Dictionary” was carried out according to
the methodology described before. It should be noted that this dictionary is not available in LNB’s
catalogue of digitized books, nor is its data added, for example, in the Latvian National Terminology
portal (https://termini.gov.lv/). First, each page of the book was scanned, then data from chapter 1 of
the dictionary, p. 5–70, was retrieved and analysed. Figure 9 in the article shows the level of certainty
about the correctness of the plant names language recognition in each page. It is possible to retrieve
precision data from the algorithm for each analysed element, however, it may be inaccurate, as the
analysed element may not be a word, but a blank space or so called “noises” (e. g., dots, lines, needless
symbols, etc.). Therefore, the image (see Figure 9) highlights the average certainty of the page. This
figure also shows linear regression curves for each graphic (dotted red, purple and orange lines).
When processing data from Chapter 1 of the book with a language identifier, data on 2204 plant
names in Latvian, 2135 in Russian, 2287 in Latin were obtained. Russian plant names gave 100%
correctness. It is not surprising, because Russian alphabet differs the Latin and Latvian alphabets and it
is easy to identify Cyrillic symbols. When the processing algorithm had to choose between Latvian and
Latin, identification success of Latvian was 90.6%, however only 39.74% of Latin words were
identified correctly.
Figure 9: Results of data reading with the algorithm developed in the study from P. Galenieks’s
“Botanical Dictionary”, p. 5–70.
After choosing a certain section from P. Galenieks’s “Botanical Dictionary”, a total of 65 pages of
text were processed with the developed solution. That is a main part of the dictionary, and plant names
are arranged in three columns like in dictionaries. In the processed 65 pages, overall 6626 plant names
were organized in three columns – Latvian, Latin and Russian. This format was chosen for checking
the accuracy of the language identification algorithm. The results of recognizing the text from the
images are so good that they can be used for retrieval of the data and addition to the database. Achieving
full precision of the data retrieved from the study requires human involvement, as the data should be
carefully reviewed, compared with the printed work and corrected if necessary. Manual processing of
data from Galenieks’s dictionary (text review and collation, correcting errors, adding page numbers)
after extraction requires 0.5 staff-hour for 10 pages or 0.0024 man-hour per entry consisting of one
scientific plant name and 2 local language names, in total 0.0008 staff-hour per one plant name. To
compare our developed software solution with data entry processes done by persons, manual entry
actions of six people were analysed during September–December 2021. During this period six people
stored 20704 organism names in the developed IMDS database. Taking into account the staff-hours
spent to enter organisms names, obtained results showed overall 0.0263 staff-hour per one organism
name entered manually, which is 32 times slower than using our developed solution.
4. Conclusions
The study described in the article is part of a larger study developing a database of organism names
(IMDS) whose main task is to address the dispersion of data across different digitized and non-digitized
resources, as well as to produce data with accurate metadata that would be useful for research purposes.
These data processing operations have been performed using existing tools in order to
develop a simplified processing workflow for this type of dictionaries. The methodology
developed in this study could be useful for digitizing other books (including those with a dictionary
structure), where the local language names of plants, animals (also terms from other fields, e. g. medical
terms in local language and Latin) are linked to their scientific or Latin name. The methodology
developed in the study should be a new contribution to the further digitization and processing of
specialized literature for the study of texts in this field as well.
It can be concluded that the Latvian language and Russian language models are sufficiently well
trained to correctly identify the conformity of language units with the Latvian and Russian languages.
The existing Latin model, on the other hand, is not sufficiently well trained to identify terms in the field
of botany. The Latin names of the taxa of organisms often are given in italics in the printed books. This
is also the case in P. Galenieks’s work “Botanical Dictionary”, where all Latin names are in italics
making data reading from images more difficult. This means worse quality input data for language
identification. Also, similar alphabets in Latvian and Latin could make it more difficult to identify
words in Latin.
In perspective, the approach described in this study could be applied to other work in similar fields,
such as the digitization of dictionaries of zoological terms. Similarly, medicine is one of the fields where
Latin terms are often used in addition to local terms.
The data prepared by the method described in the study will be further collected in the IMDS
database. However, the data acquisition methodology described and developed in the study also relieves
the manual work of the data entry personnel when entering data in the intended database. The overall
process takes less time than entering all organism names manually. Obtained results of calculations
showed that using developed software and manually checking is 32 times faster than manual entry in
the IMDS database.
5. Acknowledgements
This research has been funded by the Latvian Council of Science, project “Smart complex of
information systems of specialized biology lexis for the research and preservation of linguistic
diversity”, No. lzp-2020/1-0179.
6. References
[1] Līdzsvarotais mūsdienu latviešu valodas tekstu korpuss (LVK2018 (beta). [The Balanced Corpus
of Modern Latvian.] University of Latvia – Institute of Mathematics and Computer Science, 2018.
URL: http://www.korpuss.lv.
[2] K. Levāne-Petrova, Līdzsvarotais mūsdienu latviešu valodas tekstu korpuss, tā nozīme gramatikas
pētījumos [The Balanced Corpus of Modern Latvian, its role in grammar studies.], in A. Kalnača
(Ed.), Valoda: nozīme un forma, 10. Latvijas gramatiskā doma gadsimta gaitā [Language: Meaning
and Form, 10: A century of Latvian grammatical thought.], Riga: Latvijas Universitātes
Akadēmiskais apgāds, 2019, pp. 131–146.
[3] A. Spektors, The dictionary and thesaurus of Latvian Tēzaurs.lv. URL: https://tezaurs.lv/cīnija.
[4] Latvijas Nacionālais terminoloģijas portāls Termini.gov.lv. [National Terminology Portal
Termini.gov.lv.] URL: https://termini.gov.lv/atrast/cīnija.
[5] Kurzemes sēklas, SIA. Sēklu piedāvājuma elektroniskais katalogs. [Electronic catalogue of seed
supply]. URL: https://www.kurzemesseklas.lv/lv/product/759-cinnija-aztec-sunset.
[6] I. Birulis, 400 puķes Latvijā no pavasara līdz rudenim. [400 flowers in Latvia from spring to
autumn], Riga: AS “Lauku Avīze”, 2007.
[7] P. Galenieks, Botāniskā vārdnīca. Rīga: Latvijas Valsts izdevniecība, 1950.
[8] Latviešu valodas seno tekstu korpuss SENIE. [The Corpus of Ancient Latvian texts SENIE], URL:
http://www.korpuss.lv/id/Senie
[9] I. Ēdelmane, Ārstniecības augu terminoloģija, in Latviešu valodas kultūras jautājumi. Volume 4.
Rīga: Liesma, 1968, pp. 103–107.
[10] F. Dornseiff, Der deutsche Wortschatz nach Sachgruppen. 8. völlig neu bearb. Aufl. von Uwe
Quasthoff, Berlin, New York: de Gruyter, 2004.
[11] L. Eikvil, Optical Character Recognition, Norsk Regnesentral, Oslo, Norway, Rep. 876, 1993.
[12] V. Langenfelds, E. Ozoliņa, G. Ābele, Augstāko augu sistemātika. [Systematics of Higher Plants],
Rīga: izdevniecība “Zvaigzne”, 1973.
[13] N. Efford, Digital Image Processing: A Practical Introduction Using JavaTM. Pearson Education,
2000.
[14] S. L. Bangare, A. Dubal, P. S. Bangare, S. T. Patil, Reviewing Otsu’s Method For Image
Thresholding, International Journal of Applied Engineering Research 10(9), 2015/1, pp. 21777–
21783.
[15] T. Hegghammer, OCR with Tesseract, Amazon Textract, and Google Document AI: a
benchmarking experiment, Journal of Computational Social Science, 2021.
[16] C. Clausner, A. Antonacopoulos, S. Pletschacher, Efficient and effective OCR engine training,
International Journal on Document Analysis and Recognition (IJDAR), Volume 23, 2020, pp. 73–
88.
[17] J. Martínek, L. Lenc, P. Král, Building an efficient OCR system for historical documents with little
training data, Neural Computing and Applications. Neural Computing and Applications, Volume
32, 2020, pp. 17209–17227.
[18] M. Lui, T. Baldwin, langid.py: An Off-the-shelf Language Identification Tool, Proceedings of the
ACL 2012 System Demonstrations, 2012, pp. 25–30.
[19] R. Darģis, K. Levāne-Petrova, I. Poikāns, Lessons Learned from Creating a Balanced Corpus from
Online Data, Human Language Technologies – The Baltic Perspective, Volume 328, 2020,
pp. 127–134.