Digitalisation of cultural heritage in Estonia has been in progress during recent years, and we will see expansive mass digitalisation of printed books and handwritten documents in the very near future. This situation and potential actualises questions of the usage of the literary heritage. In our paper, we consider bene ts for digital literary research that arise from representing a literary text collection as an annotated language resource. We discuss the pilot project of creating a text corpus based on private letters between two Estonian avantgarde writers in the beginning of the 20th century. The advantages and possibilities of corpus query system KORP that we have chosen for representing and searching literary heritage DH corpora as a language resource are described. Challenges that the application of Natural Language Processing and Text and Data Mining imposes on the preparation and representation of texts are discussed along with bene ts for the research.
Application of Natural Language Processing (NLP) and Text and Data Mining
(TDM) methods and tools is gradually increasing and becoming more relevant in
DH research in Estonia. Similarly to the Nordic and Baltic countries, Estonia can
expect an explosive growth in digital heritage and text resources: preparations
for massive digitisation of cultural heritage started as a national programme
in 2019. The creation of these new digital resources will be the priority for all
Estonian leading memory institutions and the scope of this huge project includes
di erent types of cultural heritage (printed books, archival documents, photo and
lm heritage, ethnographic and ne art objects) [
What are the new challenges and the new knowledge o ered by using the
tools of NLP in analysing the digital literary heritage? Our interdisciplinary
project \Literary Studies Meet Corpus Linguistics" concentrates on using
cultural heritage, especially literary history sources in research with the help of the
NLP&TDM methods. The primary question is how to bridge the gap between
the research possibilities o ered by the contemporary NLP&TDM and the ever
increasing amount of digitised texts and other digital data, produced by memory
institutions. This has proved to be a complicated task and international
practice has shown that literary scholars are slower to embrace new practices than
linguists for whom corpus-based research is already a professional standard. In
general, literary scholars are used to working with texts using traditional
methods, analysing them as undivided poetical and semantic entities [
NLP&TDM methods require treating literary works or texts as data, which
can be analysed and processed with computer programmes. To process data for
statistics and for nding trends, a literary scholar needs either to do a close
reading of literary texts as huge data amounts, or to develop new professional skills
in data analysis methods. This imposes rethinking the approach to empirical
object in literary studies in general and posing new and di erent research
questions. The possibility to compare text strategies, rhetorical and stylistic patterns
in literary, religious and political text corpora might give us new insights into
intertwining ideology, rhetoric and identity presentations. One of the important
trends of automated textual analysis in DH is Sentiment Analysis which, for
instance, allows to measure emotion in parliamentary debates [
The empirical foundation of our pilot project of private letters in KORP is based on the archival ego-documents. This is a rare collection of handwritten private letters: the correspondence between two Estonian writers Johannes Semper (1892{1970) and Johannes Barbarus (1890{1946). The authors were well-known avant-garde writers in Estonian literature, best friends and school classmates. They shared the same intellectual attitudes, views, and values. For example, they were Francophiles and during several decades they both reviewed and translated French literature into Estonian.
During the years covered by the correspondence, in 1911{1939, Barbarus published the total of 17 collections of poetry, and Semper published four collections of poetry and four books of short stories and novels. The letters were written mainly at di erent locations in Estonia, and also during travels: they both travelled extensively in Europe (and other parts of the world) and in their letters they described to each other the sights they saw as well as the literary events at home. They also wrote about their work, everyday life, health and prescriptions, hobbies (hunting, swimming, skating, etc.), and guests.
The temporal and contextual frames of this correspondence are the period in Europe between two World Wars and, according to the French philosopher Pierre Bourdieu, the creation of the Estonian cultural eld in the Estonian Republic in years 1918{1940. As such, the correspondence o ers semantically rich and multidimensional content for literary scholars.
Generally, we have been inspired by the question how this kind of sources
(private letters, correspondences, manuscripts, etc.) in di erent kinds of
machinereadable formats (pdf, docx, txt etc.) can be used for creating new knowledge
in the interdisciplinary eld of literary studies, including biographies and life
writing studies [
The private correspondence of two Estonian writers during 28 years, as empirical material, is extremely rich in the themes and possibilities to pose the traditional as well as the new research questions: the factors of subjectivity and emotionality, verbal and poetic creativeness of the both authors, etc.
The original letters are held at di erent archives: the letters of Barbarus to Semper belong to the Estonian Cultural History Archives of the Estonian Literary Museum and the letters of Semper to Barbarus belong to the Estonian National Archives due to the unique literary and historical value of this correspondence. Their correspondence consists of 670 letters with, all in all, more than 1,100 pages and more than 310 000 tokens or about 250 000 words. The number of tokens and words is approximate, as we will discuss later. 3
KORP is a corpus query analysis system that allows to nd concordances and
build various statistics from di erently annotated corpora using text metadata
(author, year of publishing, text type etc) and linguistic annotation (splitting
into sentences and words, punctuation, morphology, syntax and semantics).
Technically, KORP is a frontend tool that uses much of IMS Open Corpus
Workbench [
Estonian KORP is hosted by the Centre of Estonian Language Resources9. The Estonian data available in KORP currently consists of more than 850 million tokens. As Estonian is a highly in ective language with a rich morphology and free word order, almost all texts are automatically annotated and disambiguated on the morphological level to enable form- and baseform-based search.
As a pilot project for the needs of literary studies in addition to the corpora designed for linguistic research, we have created Correspondence Corpus of private letters (Semper and Barbarus, 1911{1940). The original handwritten manuscripts where already typed in. They were to transform to machine-readable text les by manually adding metadata and automatically analysing and disambiguating morphological categories.
KORP as a corpus query system was chosen because of its open source value, search exibility and easiness, graphical overview of search results in subcorpora, easy switching between concordances and broader context (as well as between frequency statistics and concordances), broad possibilities to group absolute frequencies of tokens or occurrences and automatically calculated relative frequencies (occurrences per million tokens). Search results and statistics from KORP are exportable into Comma Separated Values (CSV) and JSon format.
Text fragments cited in KORP are limited to a sentence or a passage, so that KORP does not infringe copyrights. Metadata cited in the query results allow exact pinpointing of the source of the sentence, and it is possible to link whole texts hosted elsewhere to the search metadata for close reading of a broader context when needed.
KORP helps to analyse and objectively verify or refute di erent research arguments, for example, such as presented by Estonian literary scholars in the 1980s: (1) The correspondence of Semper and Barbarus is subjective and emotional, the letters reveal the character and state of mind of the authors at the moment of writing them. (2) The letters demonstrate the authors awareness of topical problems in Estonia and in Europe. (3) The subject range of the letters includes everyday life, health, hobbies, visits and visitors, literary work, books and reading, and the literary, economic and political life in Estonia and Europe. 4
Which distinctive features of older correspondences need to be considered in preparing them to be used as a corpus? Several steps are necessary to convert the 3 https://spraakbanken.gu.se/korp/ 4 http://korp.csc. 5 http://gtweb.uit.no/korp/ 6 http://korp.keeleressursid.ee 7 http://alf.hum.ku.dk/korp/ 8 http://malheildir.arnastofnun.is/ 9 https://www.keeleressursid.ee digitised text collection into a morphologically analysed corpus, and to provide it with essential metadata about the sources. 1. The rst step in digitising di erent kinds of manuscripts, incl. private letters, is retyping the handwritten originals in order to transform them into machine-readable format. Digitised typed texts can be subjected to the character recognition (OCR). 2. In order to enable statistics and to nd di erent linguistic phenomena in the texts, it is important to tokenise at least sentences and words, and to preserve meaningful units (chapters, articles, verses, letters). 3. The process of annotation was carried out semi-automatically. At the rst stage, sentence boundaries were determined, then words were tokenised. Tokenised words were lemmatised and tagged by part-of-speech and other grammatical features.
Di erent levels of annotation make it possible to search for di erent
linguistic or other phenomena. In a tokenised text we can search by word forms (or
their parts), but not by lemmas. In a lemmatised text we can search also by
lemmas, but not by morphological characteristics (e.g., the case, the number).
When the text has been morphologically analysed and disambiguated, we can
search also by morphological characteristics and parts of speech. The automatic
morphological analysis and automatic disambiguation that enable searching for
di erent morphological forms of a word based on the lemma, seriously improve
the usability and quality of the corpus. The accuracy of automatic
morphological disambiguation in the Estonian standard language corpora reaches up to
93{98% [
Meaningful units, e.g., temporal expressions or named entities, found in the text can also be used in searches. If all the meanings of words were tagged, it would be possible to search for a speci c meaning of a word and discard all other meanings. 5
The private letters of writers, the correspondence as whole was not addressed for a wider audience. Creators of the corpus met serious di culties because often, both correspondents were very familiar with the subjects mentioned in the letters, so that sometimes they only hinted at the events or persons and used plenty of abbreviations known only to themselves. Authors of letters re ect both the historical period and the personal context and idiolects. Being avantgarde poets, Semper & Barbarus both experimented with the language a lot.
In the following sections we are going to present the distinctive features of the texts that would be of great interest for both linguists and literary scholars. See Fig. 1 for a passage from a letter by Barbarus to Semper with annotations and English translation.
In this corpus, there are a lot of nonstandard abbreviations that were comprehensible to the authors of the letters. The abbreviations pose a challenge to the morphological analysis and disambiguation of abbreviations itself, but also of the neighbouring words, the disambiguated analysis of which often relies on the context (for instance, \is. Linde" should be analysed as \isand (i.e master) Linde" and not as a sentence ending and a starting of a new sentence although there is no common abbreviation \is." in standard Estonian). 5.2
The date format of the manuscript letters varies greatly: \11. jaanuar 1927",
\19/XII.37", \2. 2. 1934", \J~oulu 3. puhal 1934" (on the 3rd day of Christmas
1934). The date in the format chosen by the letter authors needs to be saved
as a part of the text, but in addition to that, all the date formats need to be
normalised so that they could be present in the letter metadata in a standardised
format (i.e., YYYY-MM-DD) whenever possible. When it is not possible, we can
use the category \unde ned" for the analysis transparency. Temporal expressions
in the text of letters can be automatically recognised with the help of EstNLTK
library for Python [
Special interest for the literary scholars lies in the proper names used in the letters: their usage combined with the metadata (time period, author) might lead to recognising the important patterns and tendencies.
Named Entity Recognition (NER) annotations can be produced with the EstNLTK toolkit and include the types of entities: person (PER), organisation (ORG), and location (LOC). 5.4
Semper & Barbarus were both promoters of French literature and culture in Estonia. They both worked as translators and travelled a lot. They scattered their letters with words, phrases, sentences and longer citations in many other languages than Estonian that were familiar to them (Latin, French, Russian, German). This is a challenge for the automatic language recognition and annotation tools which are currently oriented only to the analysis of standard Estonian. These tools do not help much in case of excerpts from other languages, but if such code-switching pieces were precisely recognised, lemmatised and analysed, that information would be useful both to linguists and literary scholars. \ Bifur "i kiri saabus Sinu kirjaga uhel paeval . Palutakse kohe luule ule artikkel ara saata ja kedagi paluda, kes teisi kusimusi ( sur la vie en general ) kasitaks. Neil olla to~lkija, nii siis vo~ivat eesti keeleski kirjutada. Et mul artikkel juba valmis oli, siis saatsin ta tana minema. Kui Sul lusti midagi saata, siis lakita kohe , | ehk novelli to~lge (maksavad 50 fr. lehekuljest), ehk siis mahutavad neljandamasse nr -isse; ehk viskad proosa & teaatri ulegi artikli. Aadress : \ Bifur " , Editions du Carrefour , 199 , boul . St.-Germain , Paris (VIe) ( M{eur le redacteur en chef Ribemont Dessaignes ). Kusisin kirjas, kas nende to~lkija luuletisi to~lkida vo~iks, siis vo~iksime valiku teha, ehk Suitsi ilmuva antoloogia neile saata. \ Bifur " harrastab kull rohkem proosat & informatsioonilaadilisi ulevaateid, nii siis vaevalt nad luule liimile lahevad, aga eks ole ju veel teisi zurnaale paale \ Bifur "i, kus avaldada saaks, kui aga to~lkija leiduks. Mis teeb see E.K.L. propagandakomitee? Kas peab viimaks nende liikmete seas propagandeerima hakkama? Visnapuu kirjutab \ V.-M aas" Igori potserduse puhul, et vaja propagandeerida, aga, kui ma ei eksi, oli ta ise selles komitees? Legend: green
| Named Entity: Person, Organisation, Title yellow
| Named Entity: Location, Address blue orange | Abbreviations
| other languages than Estonian magenta
| temporal expression \The letter from Bifur arrived on the same day as yours. They asked to send immediately the article about poetry, and to nd somebody who could treat other questions (sur la vie en general). They are supposed to have a translator, thus it could be written in Estonian. Since I had already nished my article, I sent it to them today. If you think youd like to send something, do it at once, | maybe a translation of some short story (they pay 50 francs a page), and perhaps they can t it into the fourth issue; or maybe you will even rush an article about prose and theatre. Address: Bifur, Editions du Carrefour, 199, boul. St.-Germain, Paris (VIe) (M|eur le redacteur en chef Ribemont Dessaignes). I asked them in my letter whether their translator can translate poems, in this case we could select something, or we can send them the anthology by Suits, which will come out soon. Bifur is more into prose and information-like reviews, so it's unlikely that they can be cajoled into accepting poetry, but there are other magazines besides Bifur where we could publish, if only we found a translator. What is the E.W.U.'s [Estonian Writers' Union] propaganda committee up to now? Should we by any chance start propagandising among their members? Visnapuu writes in the V.-Maa about Igor's handiwork that it is necessary to propagandise, but if I'm not mistaken, wasn't he on the committee himself?" Fig. 1. Excerpt from letter (from Barbarus to Semper, Nov 8., 1929) with di erent annotation types. See Table 1 for metadata associated to the letter the passage comes from.
Text Attribute Value author Barbarus recipient Semper catalogue no. 333 original date 8. november 1929 category [empty] date 1929-11-08 year 1929 location Parnu notes [empty]
The creation of the text corpus of the Semper & Barbarus correspondence had two wider objectives.
From the perspective of corpus linguistics, to study the e ect of the distinctive characteristics of private letters in creating a text corpus: what can and what cannot be achieved? What di culties may arise in the course of such work? How can a digitised collection be converted into a language resource to study the linguistic phenomena in their literary and cultural contexts?
From the perspective of literary studies, to test the suitability of the methods of corpus linguistics in solving the problems that literary scholars face.
One of the corpus usage examples is veri cation of the index of proper names mentioned in the correspondence. The index was created manually more than 30 years ago and it lists the foreign writers mentioned in the correspondence. From the visualisation of index statistics (Fig 2) we can only see that Gide is the most discussed French writer in the whole correspondence.
Query of KORP allows us to see the concordances of the Gide mentions, their absolute and relative frequency in the corpus, but not only. KORP allows us to organise the statistics by all the categories used in the corpus, including metadata categories. To know who of the authors and when mentioned Andre Gide, we only have to add those metadata categories (\sender" and \date") to the statistics criteria; other metadata and linguistic categories can also be used for more sophisticated statistics that is connected to the concordances and broader context text blocks.
From the statistics we see that it was mainly Semper who talked about Gide and the time interval is 1926 to 1930. It aligns well with the literary history: in 1928 Semper graduated as Master of Arts (magister artium ) from the University of Tartu (with a masters diploma in literary studies, \The structure of the literary style of Andre Gide") and continued his work at the university as an aesthetics and stylistics lecturer.
At present stage we have applied the standard procedures for converting text into KORP-compatible corpus to the Correspondence corpus as well with some minor additions. The standard procedure is a pipeline consisting of tokeniser, sentence splitter, morphological analyser (including POS-tagger and lemmatiser) and morphological disambiguator. Sentence splitter was modi ed according to non-standard abbreviations found from the Correspondence texts, but morphological analyser was used with standard options: guesser and forced proper-name nder.
Sentence splitter and tokeniser are tools that determine the quality of the corpus. They can be adjusted by properties of corpus and annotation aims, e.g. we can annotate a date as one token, or as a series of several tokens. This would give di erent results of number of tokens and number of sentences.
Guesser makes the analyser nd answers for words not in analyser's lexicon, and proper-name nder gives precedence to proper names, if analysis is ambiguous and word starts with a capital letter. The corpus is not properly evaluated, but quick overview revealed at least two problems in POS tagging and lemmatisation. First, the POS tagging seem to be too optimistic about proper names, and second, foreign words and phrases are analysed as Estonian.
We compared the number of proper names with manually disambiguated corpus, and found the general number to be rather normal. Although being far more than in ction, it fell between relative frequency of proper nouns in informational texts and journal texts (newspapers). The over-estimation of proper nouns showed up at the introductory parts of the letters, where capitalised word \Armas" stands before name of the recipient. The English equivalent would be \Dear" at the very beginning of letter. Well, \Armas" is a rst name used in Estonia, but it is a extremely rare one, and it was really never used as such in this correspondence. We adjusted the work ow accordingly by switching o forced proper-name nder, and with some decline of the number of proper nouns, got rid of analysis of \Armas" as name. How much did it a ect the recall of proper names, needs some further examination.
As it was mentioned before, both of the authors used foreign language in their letters as well. Just now it is di cult to tell, how much, because the Estonian morphological analyser was too eager to nd suitable analyses. It would be possible to apply some language detecting tool before morphological analysis, but it is possible that foreign phrases are not so numerous and this strategy would generate too much noise. 8
The number of words and sentences in this corpus is rather big to annotate it entirely manually, but at least by some extent it would be necessary. Without manually annotated data it would be di cult, if not impossible, to evaluate automatic annotation that is trained on nowadays texts. We are going to annotate at least some parts of the texts manually to nd better methods to improve automatic analysis.
Such archival materials as a private correspondences of writers have a great cultural value, but promise an incredible linguistic importance as well. The elaborated search possibilities of KORP may be used to verify objectively, via linguistic analysis, the conclusions which have so far been drawn by using traditional methods of literary research, mainly using the well known method of slow close reading of the texts.
Application of corpus linguistic methods in literary studies based on archival sources requires meticulous preparation of the material and transforming it into a text corpus. The text corpus of the correspondence of the writers, translators and friends Johannes Semper and Johannes Barbarus allows to go further with principally new type of research questions, about the networks of the writers, or the ideas, that have in uenced the original works of both authors, leading more abstract semantic research. This could be the next phrase of the work with KORP.