Lemmatization of Historical Old Literary Finnish Texts in Modern Orthography

Abstract

Texts written in Old Literary Finnish represent the first literary work ever written in Finnish starting from the 16th century. There have been several projects in Finland that have digitized old publications and made them available for research use. However, using modern NLP methods in such data poses great challenges. In this paper we propose an approach for simultaneously normalizing and lemmatizing Old Literary Finnish into modern spelling. Our best model reaches to 96.3% accuracy in texts written by Agricola and 87.7% accuracy in other contemporary out-of-domain text. Our method has been made freely available on Zenodo and Github.

Full text

Lemmatization of Historical Old Literary Finnish Texts in
Modern Orthography
Mika Hämäläinen1Niko Partanen2Khalid Alnajjar1
(1) Department of Digital Humanities
(2) Department of Finnish, Finno-Ugrian and Scandinavian Studies
(1,2) University of Helsinki, Finland
firstname.lastname@helsinki.fi
RÉSUMÉ
Les textes écrits en vieux finnois littéraire représentent la première œuvre littéraire jamais écrite en
finnois à partir du XVIe siècle. Il y a eu plusieurs projets en Finlande qui ont numérisé des anciennes
collections de textes et qui les ont rendues disponibles pour la recherche. Cependant, l’utilisation de
méthodes TAL modernes avec telles données pose de grands défis. Dans cet article, nous proposons
une approche pour normaliser et lemmatiser simultanément des textes écrits en vieux finnois littéraire
à l’orthographe moderne. Notre meilleur modèle donne une précision de 96,3 % avec les textes écrits
par Agricola et de 87,7 % avec d’autres textes contemporains hors du domaine. Notre méthode est
publiée gratuitement sur Zenodo et Github.
ABSTRACT
Texts written in Old Literary Finnish represent the first literary work ever written in Finnish starting
from the 16th century. There have been several projects in Finland that have digitized old publications
and made them available for research use. However, using modern NLP methods in such data
poses great challenges. In this paper we propose an approach for simultaneously normalizing and
lemmatizing Old Literary Finnish into modern spelling. Our best model reaches to 96.3% accuracy in
texts written by Agricola and 87.7% accuracy in other contemporary out-of-domain text. Our method
has been made freely available on Zenodo and Github.
MOTS-CLÉS :données historiques, normalisation, lemmatisation.
KEYWORDS:historical data, normalization, lemmatization.
1 Introduction
Finnish language has a long literary history starting from the 16th history. A large portion of the
books printed in Finnish is currenlty openly available for research, and especially the bibliographic
record they form has already been studied. (Lahti et al.,2019) investigated the document size and
language in the bibliographic metadata records, and (Tolonen et al.,2019, 58) took into account a
number of other metadata fields including the titles, and recognize the lack of full-text documents as a
downside of investigation that is possible. Although the whole body of books printed in Finland is not
available as full-text, there are numerous smaller corpora that can already be used in various ways.
Actes de la 28e Conférence sur le Traitement Automatique des Langues Naturelles
Lille, France, 28 juin au 2 juillet 2021
Volume 1 : conférence principale, pages 189–198.
Cette œuvre est mise à disposition sous licence Attribution 4.0 International.

Both the National Library of Finland and Institute for the Languages of Finland have produced a
large number of digitized material from the era of the Old Literary Finnish. Former has made a large
amount of scanned and text recognized publications available (National Library of Finland,2021),
and latter has created large plain text corpora from selected works (Institute for the Languages of
Finland,2014). The Agricola corpus used in this study has been morpho-syntactically annotated,
and is to our knowledge the only annotated resource in the Old Literary Finnish (Institute for the
Languages of Finland & University of Turku,2020). Another very central resources is the Dictionary
of Old Literary Finnish.
1
Thus far the Old Literary Finnish has largely eluded any attempts of NLP
research as the historical written form cannot be processed easily with currently available NLP tools
for Finnish, as they are designed for modern Finnish. We present our approach for normalizing and
lemmatizing Old Literary Finnish automatically to modern Finnish orthography. As the resources for
processing historical Finnish text are scarce, we have released the models presented in this paper on
Zenodo 2and through an easy-to-use Python library 3.
The use of existing NLP tools targeted for modern Finnish on historical materials can be made
possible through normalization. Previous work conducted on English data indicates that normalization
is a viable way of improving the accuracy of NLP methods such as POS tagging (van der Goot et al.,
2017). Another direction of digital humanities study has benefited from normalization of historical
data in studying the use of neologisms in old letters (Säily et al.,2018;Säily et al.,2021). In their
approach, without normalization, they would have been able to cover only a small subset of the corpus.
The same corpus has also been studied without NLP tools (Nevalainen,2021).
The history of printed written Finnish starts from the 16th century with the works of Mikael Agricola.
His primer and religious works were followed by a continuous increase in the amount of the written
Finnish materials. The language form used by Agricola is known as Old Literary Finnish (vanha
kirjasuomi). The majority of the early Finnish publications included religious materials, although
the text types started to diversify already in the 18th century. The majority of the oldest texts are
translations. The period of Old Literary Finnish is often estimated to have lasted until 1810, after
which the written Finnish started to transform into Early Modern Finnish. Eventual changes in printing
laws and printing technology, which expanded the amount and variation of the printed materials, and
the creation of regular Finnish newspapers, contributed to the stabilization of the written standard.
Linguistically one of the exceptional features of Old Literary Finnish is the variation it displays. The
orthography was not yet entirely established, and there was extensive spelling variation. The age of
these materials adds also a historical dimension, as there are linguistic features that are not present in
the modern Finnish, or exist currently only in the dialects.
Our main contributions in this work are :
— Building the first artificial neural network model for normalizing historical Finnish.
—
Conducting an evaluation for assessing the performance of the model on historical data from
1) the same source of the data used in building the model and 2) external out-of-domain
historical data. In both cases, a high accuracy is achieved by the model.
— Publishing a user-friendly Python library that permits an instant usability of our model.
The target language form in our work is modern Finnish. Our model primarily lemmatizes, but since
the output is harmonized into contemporary orthographic forms, the work is closely connected to the
normalization task as well, and the output can be considered as one type of a normalization. The exact
1. https ://kaino.kotus.fi/vks/
2. https ://zenodo.org/record/4734143
3. https ://github.com/mikahama/murre
190

lemmatization choices and conventions were decided at the level of the original morpho-syntactic
database, and we followed those closely also when our own additional test material was created. In
later research also the morpho-syntactic annotations present in the corpus could be taken into account
to further enrich the analysis. However, at the current stage a successful lemmatization is already a
large improvement in available NLP methods.
This paper is structured as follows. We begin by describing the related work. Thereafter, the details of
the data used to build the neural model are given, followed by the architecture and hyperparameters of
the neural model. Section 5 presents the results and evaluation where we explain the different training
strategies we experimented with and their performance against a baseline (historical Omorfi). Lastly,
we discuss and conclude our work while highlighting future directions with a potentially great impact
on humanities research such as automatic analysis of historical Finnish.
2 Related Work
Historical text normalization has been studied in the past for other languages than Finnish. A recent
literature review (Bollmann,2019) finds that there are five categories in which modern normalization
approaches can be divided : substitution lists like VARD (Rayson et al.,2005) and Norma (Bollmann,
2012), rule-based methods (Baron & Rayson,2008;Porta et al.,2013), edit distance based approaches
(Hauser & Schulz,2007;Amoia & Martinez,2013), statistical methods and most recently neural
methods (Partanen et al.,2019;Duong et al.,2020).
Statistical machine translation (SMT) based methods have been the most successful ones in the
past in terms of statistical methods. The key idea behind these methods is to approach the task as
a character-level machine translation problem, where a word is translated character by character to
its normalized form. These methods have been applied to historical text (Pettersson et al.,2013;
Hämäläinen et al.,2018) and dialect normalization (Samardzic et al.,2015).
In the recent years, normalization has been approached as a character-level neural machine translation
(NMT) problem similarly to the previous SMT approaches. The additional advantage is that a neural
model does not need a separate language model like SMT does. Bollmann & Søgaard (2016) have
shown that a bi-directional long short-term memory (bi-LSTM) can be used to normalize historical
German texts. The paper presents a so-called multi-task learning setting where auxiliary data is added
to improve the performance of the model. Multi-task learning setting generally improved the results.
Their system outperformed the existing conditional random fields and Norma based approaches in
terms of accuracy.
Text written in Uyghur language has been normalized with an LSTM and a noisy channel model
(NCM) (Tursun & Cakici,2017). They use a relatively small set of gold annotated data for training
(around 200 hand normalized social media sentences). They augment this data by synthetically
generating non-normalized text by introducing random changes in normalized text. In the same
fashion, another research has used an LSTM model to normalize code-mixed data (Mandal &
Nanmaran,2018).
Recently Hämäläinen et al. (2019) have shown that bi-directional recurrent neural networks (BRNN)
outperform regular unidirectional recurrent neural networks (RNN) when normalizing historical
English data. Interestingly, additional layers and different attention models do not improve the results.
Additional data such as time period, social metadata or pronunciation information in IPA characters
191

makes the results worse. According to them, post-processing can boost the accuracy of a character
level NMT model more than changing the network structure. A simple dictionary filtering method
improved the results.
Omorfi (Pirinen,2015) is a popular rule-based tool used to do morphological analysis and lemmatiza-
tion of modern Finnish. While Omorfi itself is not relevant for our work, there is a GitHub fork of
the project known as Historical Omorfi
4
. The fork introduces several improvements to better cater
for historical Finnish text. Currently, this tool is the only tool available for lemmatizing historical
Finnish. Thereby we compare the results of our model to this tool.
3 Dataset of Historical Finnish
In order to use machine learning methods, data is needed to train a model. The data needs to have text
written in Old Literary Finnish and its normalized lemmas. The lemmas should be aligned on a word
level with the historical data in order to train the normalization more accurately. Fortunately, such
a dataset exists. The corpus we use is The Morpho-Syntactic Database of Mikael Agricola’s Works
(Institute for the Languages of Finland & University of Turku,2020) that contains 522,237 tokens
and 38,222 sentences. The corpus includes all nine Finnish books translated by Mikael Agricola. The
corpus is openly available in the Language Bank of Finland.
5
In our testing we also use the Dictionary
of Old Literary Finnish
6
(Institute for the Languages of Finland,2014), from which a small number
of sentences has been sampled and manually normalized and lemmatized. Both resources are available
under Creative Commons licenses. Since the The Morpho-Syntactic Database of Mikael Agricola’s
Works is licensed under CC BY-ND 4.0 (CLARIN PUB), we do not redistribute the training material
ourselves, but it can be accessed in the Language Bank of Finland’s concordance service
7
. Naturally,
since this data is so old, it is already in Public Domain, and many of the original works are entirely
openly accessible. For example, the Agricola’s prayer book is available as high quality scans by the
National Library of Finland 8.
Although there has not been prior use of this dataset in the computational linguistics, the corpus has
been used in the linguistic studies. To illustrate this with few recent examples, Toropainen (2018)
investigated the compounds in this variety, and Toropainen (2015) studied nouns that contain an
initial adjective. Salmi (2020) discussed recently the German influence in the Agricola’s language.
Also annotating the corpus into it’s current stage has been a long undertaking, and Inaba (2015)
investigated the use of two Finnish cases with a prototype of the current database. It is beyond doubt
that the materials of Old Literary Finnish still have much to contribute to the linguistic research. We
believe that by creating new tools for natural language processing of these and similar materials we
can further expand toward these goals. The research concerning Agricola’s language and Old Literary
Finnish in general is naturally much wider and has a long research history, especially in Finland, and
we primarily wanted to illustrate in this section some of the previous studies where the same database
was used.
4. https://github.com/jiemakel/omorfi/
5. urn:nbn:fi:lb-2019121804
6. https://kaino.kotus.fi/vks
7. https://korp.csc.fi
8. https://www.doria.fi/handle/10024/43445
192

4 Neural Normalization
In this section, we describe our artificial neural network model for normalizing and lemmatizing
historical Finnish into standard Finnish. We begin by explaining how the dataset is used and how it is
preprocessed. Thereafter, we elucidate the architecture of the model along with the technical details
of its hyperparameters.
The corpus contains nine distinct works. In order to test the model’s accuracy realistically, we selected
seven books into the training data, and left the remaining two into the test set. Thereby the training
data was 393,779 tokens and the test set 128,294 tokens. The books in the test set were specifically
‘Messu eli Herran echtolinen’ from 1549 and ‘Rucouskiria’ from 1544. Additionally 15% of the
training data was used in the validation set. We considered it important not to select the test set from
the entire corpus, as this would not give a clear picture about how much the model generalizes into
new use cases, which would be the other books written in the same language variety. With the current
sparsity of manually annotated data, the Agricola works in the currently used corpus, but which we
kept unseen for the model, were the best option we had. We also extended the testing into other
examples of Old Literary Finnish, but in a smaller scale. The data has the original written form of
each sentence and a token level normalized lemma for each word. We use this parallel data to train
our models.
We model the problem as a character level NMT problem. In practice, we split words into characters
separated by white space and mark actual spaces between words with an under score (_). This allows
to pass word boundary information to the model while the characters themselves are separated by
spaces. We train the model to predict from the Old Literary Finnish word forms to the lemmas.
As previous research (Partanen et al.,2019) has found that using chunks of words instead of full
sentences at a time improves the results, we train different models with different chunk sizes. This
allows direct comparison of different chunk sizes. This way we train the models to predict one word
at a time, two words at a time all the way to five words at a time. An example of the data can be seen
in Table 1. This example is within the sentence A-II-369 in the corpus, which is located in the New
Testament translation.
Size Source Target
Chunk 1 sydhemen sydän
Chunk 3 paluellen_herra_caiken palvella_herra_kaikki
TABLE 1 – An example of the training data for chunk sizes 1 and 3.
We train all models using a bi-directional long short-term memory (LSTM) based model (Hochreiter &
Schmidhuber,1997) by using OpenNMT-py (Klein et al.,2017) with the default settings except for the
encoder where we use a BRNN (bi-directional recurrent neural network) (Schuster & Paliwal,1997)
instead of the default RNN (recurrent neural network), since BRNN based models have been shown
to provide better results in a variety of tasks. We use the default of two layers for both the encoder and
the decoder and the default attention model, which is the general global attention presented by Luong
et al. 2015. The models are trained for the default of 100,000 steps. All models are trained with the
same random seed (3435) to ensure reproducibility and to make their intercomparison possible.
193

5 Results and Evaluation
Our initial evaluation results were very good as seen in Table 2where the accuracies are reported
on a word level. The best model was the chunk of 3. The quality we reach is very high and on par
with other comparable normalization tasks, such as in (Partanen et al.,2019), which also corroborate
the best performance at the chunk of three tokens. However, we are also interested in seeing how
well our model works with other Old Literary Finnish texts. At any rate, we can see that our models
outperform Historical Omorfi, which is the only tool publicly available for historical Finnish. As
Omorfi produces all the possible lemmas for a given word, we count the accuracy based on if the
correct lemma is in the list of the lemmas Omorfi produced for each word.
Chunk 1 Chunk 2 Chunk 3 Chunk 4 Chunk 5 Omorfi
Accuracy 96.1% 96.2% 96.3% 96.2% 95.9% 40.5%
TABLE 2 – Token level accuracy of each model in the test data.
Because there is no other dataset freely available that would both be written in Old Literary Finnish
and lemmatized to modern orthography, we take randomly 50 sentences from the example sentences
of the dictionary of Old Literary Finnish
9
that is available online. Altogether these sentences have
562 words (excluding punctuation). We lemmatize these sentences with the model that has been
trained with chunks of 3 as it worked the best out of the models and verify the lemmatization by hand.
The results of this experiment are seen in Table 3.
Chunk 3 Omorfi
Accuracy 87.7% 47.9%
TABLE 3 – Accuracy in the Old Literary Finnish dictionary sample.
The results drop in this evaluation, but it is only to be expected given that out-of-domain performance
is typically lower for neural models. Nevertheless, we see clearly that the model does well in out-of-
domain data and beats the current state of the art. The fact that Historical Omorfi gets better results in
this dataset is a good indication that the text is very different from what is in the Agricola dataset.
However, it clearly is not entirely distinct when we consider how well the model still performed.
If we look at the results more closely, analyzing the errors, we can see that the model usually does
not predict non-words but rather words that are a part of the Finnish vocabulary. This indicates that
the model has learned a good target representation. There are several errors in which the model
has normalized the historical word correctly, but it has not lemmatized it, for example runsast was
normalized to runsaasti although the lemma would be runsas ‘plenty’. Another example is ulosteon
that was already written as in the modern orthography was normalized unchanged to ulosteon, while
the lemma is ulosteko. This example also illustrates how there is variation and historical change in
spelling, as we find in the Agricola corpus comparable words spelled with the initial v, whereas in
the later materials we find the variant above that is spelled closely the current standard. Analysing
how this relates to the changes in characters used in different centuries is beyond the scope of our
study, but it illustrates well the kind of variation we can find in the historical texts and their digitized
versions.
9. https ://kaino.kotus.fi/vks
194

To analyze the errors further, the most typical source of problems are verbs that get lemmatized into
nouns that look similar and vice versa. Thereby olan was lemmatized as olla ‘to be’ while it should
have been olka ‘shoulder’. Also, nuole was lemmatized as nuoli ‘arrow’, while the correct lemma is
nuolla ‘to lick’. Normalization to a wrong lemma within the same part-of-speech is also possible e.g.
kaipanne to kaivaa ‘to dig’ instead of kaivata ‘to miss’. Improving the recognition of such instances
is a very important task for the future work, but the current accuracy also appears to be useful and
satisfactory for many tasks, and is without doubt an improvement to the existing methods.
6 Conclusions and Future Work
Our results have a clear indication, both with in-domain and out-of-domain test data, of working
successfully in lemmatizing Old Literary Finnish in the modern orthography. The models have been
released on Zenodo and in a Python library
10
. By sharing the models we are making NLP research on
historical Finnish data more widely accessible for the research community, as the currently available
Historical Omorfi does not work well for texts that are this old. Our study also creates a benchmark
into which the further work can easily be compared.
Having lower accuracy in another dataset is a reminder of the importance of evaluating normalization
models on data that comes from a different distribution. This is something seldom seen in the previous
work on historical spelling normalization. Despite this, the accuracy has remained relatively high and
we have identified several possibly problematic phenomena in the test data that are more prone to
errors. This error analysis helps in understanding the biases that using our model might introduce in
historical data when it is used to lemmatize a corpus completely new to the model. Further research is
needed to evaluate how the error rate varies when the distance grows to the materials of Agricola. It is
beyond doubt that more diverse training material is needed to successfully process the entire corpus
of Old Literary Finnish, but our study certainly has improved the position to initiate and continue
such work.
The work we presented here also makes it possible for the current research on analyzing historical
Finnish newspapers, such as (Jean-Caurant & Doucet,2020;Kettunen et al.,2020), to standardize
post-OCR historical Finnish, which in turn permits employing state-of-the-art Finnish NLP methods
and tools on such data (e.g. sentiment and semantic analysis (Hämäläinen & Alnajjar,2019)).
When we work with the currently available resources, we must also remain aware that the digital
versions have been edited in various ways, and do not contain necessarily all features of the original
printed text (Toropainen,2016, 175). Now when we increasingly have access also to the original
prints as high quality scans, it is important to think how these different resources can be connected to
one another. This will need a combination of both text recognition and NLP tools.
In the future, we are interested in conducting work on semantic change on historical data. This should
be greatly facilitated by the fact that we can now considerably reliably lemmatize historical text. This
means that training word embeddings models will become more accurate as the model is trained on
lemmas instead of inflectional forms.
10. https ://github.com/mikahama/murre
195

Références
AMOIA M. & MARTINEZ J. M. (2013). Using comparable collections of historical texts for building
a diachronic dictionary for spelling normalization. In Proceedings of the 7th workshop on language
technology for cultural heritage, social sciences, and humanities, p. 84–89.
BARON A. & RAYS O N P. (2008). VARD2 : A tool for dealing with spelling variation in historical
corpora. In Postgraduate conference in corpus linguistics.
BOLLMANN M. (2012). (Semi-)automatic normalization of historical texts using distance measures
and the Norma tool. In Proceedings of the Second Workshop on Annotation of Corpora for Research
in the Humanities (ACRH-2), Lisbon, Portugal.
BOLLMANN M. (2019). A large-scale comparison of historical text normalization systems. In Pro-
ceedings of the 2019 Conference of the North American Chapter of the Association for Computational
Linguistics : Human Language Technologies, Volume 1 (Long and Short Papers), p. 3885–3898,
Minneapolis, Minnesota : Association for Computational Linguistics.
DOI
:10.18653/v1/N19-1389.
BOLLMANN M. & SØGAARD A. (2016). Improving historical spelling normalization with bi-
directional LSTMs and multi-task learning. In Proceedings of COLING 2016, the 26th International
Conference on Computational Linguistics : Technical Papers, p. 131–139, Osaka, Japan : The
COLING 2016 Organizing Committee.
DUONG Q., HÄMÄLÄINEN M. & HENGCHEN S. (2020). An unsupervised method for OCR
post-correction and spelling normalisation for Finnish. arXiv preprint arXiv :2011.03502.
HÄMÄLÄINEN M. & ALNAJJAR K. (2019). Let’s FACE it. Finnish poetry generation with aesthetics
and framing. In Proceedings of the 12th International Conference on Natural Language Generation,
p. 290–300, Tokyo, Japan : Association for Computational Linguistics.
DOI
:10.18653/v1/W19-
8637.
HÄMÄLÄINEN M., SÄILY T., RUETER J., TIEDEMANN J. & MÄKELÄ E. (2018). Normalizing
early English letters to present-day English spelling. In Proceedings of the Second Joint SIGHUM
Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and
Literature, p. 87–96.
HÄMÄLÄINEN M., SÄILY T., RUETER J., TIEDEMANN J. & MÄKELÄ E. (2019). Revisiting NMT
for normalization of early English letters. In Proceedings of the 3rd Joint SIGHUM Workshop on
Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, p. 71–
75, Minneapolis, USA : Association for Computational Linguistics.
DOI
:10.18653/v1/W19-2509.
HAUSER A. W. & SCHULZ K. U. (2007). Unsupervised learning of edit distance weights for
retrieving historical spelling variations. In Proceedings of the First Workshop on Finite-State
Techniques and Approximate Search, p. 1–6.
HOCHREITER S. & SCHMIDHUBER J. (1997). Long short-term memory. Neural computation,
9
(8),
1735–1780.
INABA N. (2015). Suomen datiivigenetiivin juuret vertailevan menetelmän valossa, volume 272 de
Mémoires de la Société Finno-Ougrienne. Société Finno-Ougrienne.
INSTITUTE FOR THE LANGUAGES OF FINLAND (2014). Corpus of Old Literary Finnish.
http ://urn.fi/urn :nbn :fi :lb-201407165.
INSTITUTE FOR THE LANGUAGES OF FINLAND &UNIVERSITY OF TURKU (2020). The Morpho-
Syntactic Database of Mikael Agricola’s Works version 1.1, Korp.
196

JEAN-CAURANT A. & DOUCET A. (2020). Accessing and investigating large collections of
historical newspapers with the NewsEye platform. In Proceedings of the ACM/IEEE Joint Conference
on Digital Libraries in 2020, JCDL ’20, p. 531–532, New York, NY, USA : Association for
Computing Machinery. DOI :10.1145/3383583.3398627.
KETTUNEN K., KOISTINEN M. & KERVINEN J. (2020). Ground truth OCR sample data of Finnish
historical newspapers and journals in data improvement validation of a re-OCRing process. Liber
quarterly.DOI :10.18352/lq.10322.
KLEIN G., KIM Y., D ENG Y. , SENELLART J. & RUSH A. M. (2017). OpenNMT : Open-Source
Toolkit for Neural Machine Translation. In Proc. ACL.DOI :10.18653/v1/P17-4012.
LAHTI L., MARJANEN J., ROIVAINEN H. & TOLONEN M. (2019). Bibliographic data science and
the history of the book (c. 1500–1800). Cataloging & Classification Quarterly,57(1), 5–23.
LUONG M.-T., PHAM H. & MANNING C. D. (2015). Effective approaches to attention-based
neural machine translation. arXiv preprint arXiv :1508.04025.
MANDAL S. & NANMARAN K. (2018). Normalization of transliterated words in code-mixed
data using Seq2Seq model & Levenshtein distance. In Proceedings of the 2018 EMNLP Workshop
W-NUT : The 4th Workshop on Noisy User-generated Text, p. 49–53, Brussels, Belgium : Association
for Computational Linguistics. DOI :10.18653/v1/W18-6107.
NATIO NA L LIBRARY OF FINLAND (2021). Digital Collections. https ://digi.kansalliskirjasto.fi.
NEVALAINEN T. (2021). Time’s arrow reversed? the (a)symmetry of language change. In M.
HÄMÄLÄINEN,N.PARTANEN &K.ALNAJJAR, Éds., Multilingual Facilitation. Rootroo Ltd.
PARTANEN N., HÄMÄLÄINEN M. & ALNAJJAR K. (2019). Dialect text normalization to normative
standard Finnish. In W. XU,A.RITTER,T.BALDWIN &A.RAHIMI, Éds., The Fifth Workshop
on Noisy User-generated Text (W-NUT 2019), p. 141–146, United States : The Association for
Computational Linguistics.
PETTERSSON E., MEGYESI B. & TIEDEMANN J. (2013). An SMT approach to automatic annota-
tion of historical text. In Proceedings of the workshop on computational historical linguistics at
NODALIDA 2013; May 22-24 ; 2013; Oslo ; Norway. NEALT Proceedings Series 18, volume 087,
p. 54–69 : Linköping University Electronic Press.
PIRINEN T. A. (2015). Development and use of computational morphology of Finnish in the open
source and open science era : Notes on experiences with Omorfi development. SKY Journal of
Linguistics,28, 381–393.
PORTA J., SANCHO J.-L. & GÓMEZ J. (2013). Edit transducers for spelling variation in Old
Spanish. In Proceedings of the workshop on computational historical linguistics at NODALIDA
2013; May 22-24 ; 2013; Oslo ; Norway. NEALT Proceedings Series 18, volume 087, p. 70–79 :
Linköping University Electronic Press.
RAYSON P. , ARCHER D. & SMITH N. (2005). VARD versus WORD : a comparison of the UCREL
variant detector and modern spellcheckers on english historical corpora. Corpus Linguistics 2005.
SÄILY T., MÄKELÄ E. & HÄMÄLÄINEN M. (2018). Explorations into the social contexts of
neologism use in early English correspondence. Pragmatics & Cognition,
25
(1), 30–49.
DOI
:
10.1075/pc.18001.sai.
SALMI H. (2020). German influence on the Finnish in Mikael Agricola. Finnish-German Yearbook
of Political Economy, Volume 2, p. 135.
197

SAMARDZIC T., SCHERRER Y. & GLASER E. (2015). Normalising orthographic and dialectal
variants for the automatic processing of Swiss German. In Proceedings of the 7th Language and
Technology Conference. ID : unige :82397.
SCHUSTER M. & PALIWAL K. K. (1997). Bidirectional recurrent neural networks. IEEE transac-
tions on Signal Processing,45(11), 2673–2681.
SÄILY T., MÄKELÄ E. & HÄMÄLÄINEN M. (2021). From plenipotentiary to puddingless :
Users and uses of new words in early english letters. In M. HÄMÄLÄINEN,N.PARTANEN &K.
ALNAJJAR, Éds., Multilingual Facilitation. Rootroo Ltd.
TOLONEN M., LAHTI L., ROIVAINEN H. & MARJANEN J. (2019). A quantitative approach to
book-printing in Sweden and Finland, 1640–1828. Historical methods : a journal of quantitative
and interdisciplinary history,52(1), 57–78.
TOROPAINEN T. (2015). Adjektiivialkuiset yhdyssubstantiivit Mikael Agricolan teoksissa. Sanan-
jalka,57(1), 54–85.
TOROPAINEN T. (2016). Typografian vaikutus yhdyssubstantiivien oikeinkirjoitukseen agricolan
teoksissa. Sananjalka,58(1), 175–198.
TOROPAINEN T. (2018). Yhdyssanat ja yhdyssanamaiset rakenteet Mikael Agricolan teoksissa.
Thèse de doctorat, University of Turku. Turun yliopiston julkaisuja. Sarja C, Scripta lingua Fennica
edita.
TURSUN O. & CAKICI R. (2017). Noisy Uyghur text normalization. In Proceedings of the
3rd Workshop on Noisy User-generated Text, p. 85–93, Copenhagen, Denmark : Association for
Computational Linguistics. DOI :10.18653/v1/W17-4412.
VAN D ER GOOT R., PLANK B. & NISSIM M. (2017). To normalize, or not to normalize : The
impact of normalization on Part-of-Speech tagging. In Proceedings of the 3rd Workshop on Noisy
User-generated Text, p. 31–39.
198