MiMe-MeMo/Corpus-v1.1

MeMo corpus v1.1

Jens Bjerring-Hansen, Philip Diderichsen, Dorte Haltrup Hansen, June 2023

This is data release version 1.1 of the MeMo corpus comprising almost all Danish novels from the period 1870-1899, known as the Modern Breakthrough.

The current version of the corpus is publicly viewable and searchable at https://alf.hum.ku.dk/korp/?mode=memo_all.

The corpus has been enhanced since version 1.0 with the following 19 titles that have been reprocessed or added to the corpus.

1. Vilhelm Bergsøe: Bruden fra Rørvig (1872)
2. Johanne Schjørring: Rige Dage (1877)
3. Anonymous: Tante Jacobine (1878)
4. Jonas Lie: Rutland (1880)
5. Vilhelm Malling: Fra Kjøbstadlivet i gamle Dage (1882)
6. Adda Ravnkilde: To Fortællinger (1884)
7. Henrik Pontoppidan: Ung Elskov (1885)
8. Therese Brummer: Som man gifter sig (1888)
9. Henrik Pontoppidan: Natur (1890)
10. R.H.: En Kjøbenhavners Livshistorie eller Lykkens Omskiftelser (1891)
11. Henrik Pontoppidan: Minder (1893)
12. Johannes Jørgensen: Hjemvee (1894)
13. Henrik Pontoppidan: Nattevagt (1894)
14. Jonas Lie: Naar Sol gaar ned (1895)
15. Gustav Wied: Ungdomshistorier (1895)
16. Herman Bang: Ludvigsbakke (1896)
17. Cornelia Levetzow: Havemanden (1896)
18. Karl Larsen: Kresjan Vesterbro (1897)
19. Christian Christensen: Kærlighedens Mysterier (1899)

The release contains the following files:

File	Contents
texts	Text files of the now 558 novels in the corpus. The text has a newline at line breaks in the book, and two newlines at page breaks. Some of the texts (the ones originally set in Fraktur) have been post-OCR-corrected using a procedure described in Bjerring-Hansen et al. (2022). The rest have been post-OCR-corrected. Error types were identified manually and implementet with look-up in the dictionary (Sprogteknologisk Ordbase, STO) to awoid the creation of new errors. This cautious method has the consequence that not all error were corrected.
normalised	Orthographically normalized versions of the 558 texts. Same format as the files in "texts", normalized to Danish standard spelling. Nouns were lower cased, aa changed to å and frequent character patterns changed to obey the Danish orthography norm from 1948. Like the error corrected version of the corpus, character patterns were identified manually and mainly implementet with look-up in the dictionary (Sprogteknologisk Ordbase, STO) to awoid overgeneration. The method has the consequence that not all words were normalized.
memo_all.vrt	VRT file (vertical format) of MeMo corpus v1.1 for indexing in Corpus Workbench (CWB). Format: One token per line delimited by <corpus>, <text>, and <sentence> XML elements. The XML elements contain attributes with metadata. The tokens are annotated with various categories separated by tabs. For more information about the metadata, see the metadata excel file. For more information about the token annotations, see below.
MeMo-corpus-metadata-v1.1-2023-06-20.xlsx	Excel file with metadata about the novels in the corpus. See the "info" tab for information about the metadata categories.

Token annotations and metadata in VRT file

There are nine columns of tokens and annotations in the corpus VRT file:

Column 1	Column 2	Column 3	Column 4	Column 5	Column 6	Column 7	Column 8	Column 9
Token	Normalized	Lemma form	Part of speech	Word no. in sentence	Word no. in line	Word no. in book	Line no. on page	Page no.

For information about the metadata also contained in the VRT file, se the file MeMo-corpus-metadata-v1.1-2023-06-20.xlsx.

References

Bjerring-Hansen, Jens, et al. "Mending Fractured Texts. A heuristic procedure for correcting OCR data." (2022). https://ceur-ws.org/Vol-3232/paper14.pdf

Data Statement

1. Header

1. Dataset Title

MeMo Corpus 

2. Dataset Curator(s) [name, affiliation]

Jens Bjerring-Hansen, University of Copenhagen; Philip Diderichsen, University of Copenhagen; Dorte Haltrup Hansen, University of Copenhagen  

3. Dataset Version [version, date] 

Version 1.1, August 15, 2023

4. Dataset Citation and, if available, 

####



5. DOI Data Statement 

####

6. Author(s) [name, affiliation] 

Jens Bjerring-Hansen, University of Copenhagen; Philip Diderichsen, University of Copenhagen

7. Data Statement Version [version, date] 

Version 1, September 25, 2023

8. Data Statement Citation 

####

2. Executive summary

The MeMo corpus is established to investigate literary and cultural change in a seminal epoch of Scandinavian cultural and social history (known as 'the modern breakthrough') using natural language processing and other computational methods. The corpus consists of original novels by Norwegian and Danish authors printed in Denmark in the period 1870-99. It includes 858 volumes, totaling 4.5 million sentences and 65 million words.

3. Text characteristics

The corpus consists of novels, i.e. long works of narrative fiction, usually written in prose and published as a book. The novels contain both dialogue and description. As instances of imaginative literature they are infused with ambiguity, interpretational confounding, rhetorical sophistication, and narrative layerings between author, narrator, and characters.

The cultural diversity of the texts in the corpus is pronounced. From a genre perspective, we have contemporary novels as well as historical novels and other forms of genre fiction such as romance, crime, and war stories (cf. Bjerring-Hansen and Rasmussen, 2023). And from an aesthetic perspective we have both avant-garde forms of realism, including instances of naturalism and impressionism, and more traditional prose with a preference for abstract or generalized over concrete specification (cf. Bjerring-Hansen and Wilkens, 2023).

Bjerring-Hansen, Jens, and Sebastian Ørntoft Rasmussen. 2023. “Litteratursociologi og kvantitative litteraturstudier Den historiske roman i det moderne gennembrud som case”. In Passage 89: 171–189.

Bjerring-Hansen, Jens, and Matt Wilkens. 2023. “Deep distant reading: The rise of realism in Scandinavian literature as a case study”. Orbis Litterarum. doi:10.1111/oli.12396

4. Curation Rationale

The MeMo Corpus was created as the basis for a research project, MeMo – Measuring Modernity: Literary and Social Change in Scandinavia 1870-1900, investigating how processes of social change in late nineteenth century Scandinavia were reflected and discussed in the novels from the period (project page: https://nors.ku.dk/english/research/projects/measuring-modernity/). As opposed to traditional historiography on the period, which has focused on selected texts by a few prominent, male authors, our digital corpus, with rich metadata on texts and authors, allows for the capturing of robust literary and sociological trends and for new insights into the processes of modernization in this formative period in the literary and social history of Scandinavia. To this corpus we thus ask questions such as: How did this breakthrough of new ways of thinking and writing actually unfold? Who were the actors? And to what extent did newness relate to literature at large?

Also, the corpus acts as the empirical foundation of an interrelated methodological project, Mining the Meaning, which aims to develop state-of-the-art computational semantic methods and training large language models towards written late 19th-century Danish and Norwegian (project page: https://mime-memo.github.io/).

Included in the corpus are all original (i.e. newly written) novels by Danish and Norwegian authors published in Denmark 1870-99. The list of texts was compiled on the basis of _Dansk Bogfortegnelse _(a continuous list of books published in Denmark since 1841; from 1861 published annually) supplemented with literary handbooks and special bibliographies.

Not included (mainly due to pragmatic reasons and for the sake of coherence) in the corpus are:

• reprints
• translations
• serializations (i.e. serialized novels from newspapers and magazines)
• diasporic literature (i.e. novels by Danish emigrant authors in the U.S.)

Around 20% of the novels are produced by female authors. Thus, highlighting and exploring the often overlooked female literary production of the period is a distinctive ambition of the corpus and the explorations based on it.

5. Language Varieties

The language of the novels in the corpus is late nineteenth century Danish (BCP-47: da). On the whole, we are dealing with a more or less linguistically coherent body of texts. However, the following circumstances must be acknowledged:

• The texts contain a pronounced spelling variation, partly on an individual level, partly explained by an ongoing orthographic standardization, which is most clearly expressed in the Spelling Reform of 1892. Here, forms such as 'Kjøbenhavn' and 'Familje' became 'København' and 'Familie'.
• Some books are written in dialect (e.g. Jutlandic or West Norwegian) or contain dialectal features to create psychological individualism in the dialogue.
• Approximately 16% of the books are written by Norwegian authors. In this regard it should be noted that, until 1907, written Norwegian was practically identical to written Danish. ‘Norvagisms’ (i.e. distinct Norwegian words, not used by Danes) do appear.

6. Preprocessing and data formatting

OCR scans: The book volumes were scanned with optical character recognition (OCR) by the Royal Danish Library’s Digitization on Demand (DoD) team. The data were delivered as full volume PDF files with the OCR’ed text as an invisible searchable, copyable text layer, as full volume text files, and as single page text files (one text file per page for each volume).

OCR correction: The text files were automatically post-corrected for OCR errors. This involved two different processes, one for texts originally typeset in Antikva (Roman) typefaces, one in Fraktur (Gothic) typefaces. The Antikva files were corrected using a set of hand-crafted substitution patterns, with look-up in the dictionary Sprogteknologisk Ordbase, STO (Eng. ‘Word database for language technology’). The Fraktur files were corrected using a correction procedure involving a combination of spelling correction, hand-crafted pattern substitution, and improved OCR using the pretrained “Fraktur” Tesseract data plus an alternative OCR layer from the pretrained “dan” Tesseract data, which was used as a corrective to problems with the Danish characters “æ” and “ø” in particular. This procedure improved the word error rate of the Fraktur data from 10.46% to 2.84% (cf. Bjerring-Hansen et al. 2022).

Bjerring-Hansen, Jens, Philip Diderichsen, Dorte Haltrup Hansen, and Ross D. Kristensen-McLachlan. 2022. “Mending fractured texts. A heuristic procedure for correcting OCR.” Proceedings of the 6th Digital Humanities in the Nordic and Baltic Countries Conference, Uppsala, Sweden, March 15-18, 2022 (DHNB 2022): 177–186.

Token-level annotation: The corrected data were annotated with grammatical information using the pipeline orchestration tool Text Tonsorium available at https://cst.dk/texton/, provided by the Danish CLARIN node. The particular pipeline used included the LaPos part of speech tagger, the CSTLemma lemmatizer, and an implementation of the Brill tagger. Grammatical information included lemma and part of speech, plus sentence and paragraph segmentation (which are of course not strictly speaking token-level annotations). In addition to the grammatical annotations, convenience annotations with various counters were also added: word number in sentence, word number on line, word number in book volume, line number on page, page number in book volume.

Text normalization: After OCR correction, all texts were normalized to modern Danish spelling using hand-crafted substitution patterns and lookup in STO (see above). Nouns were lower cased, “aa” changed to “å”, and frequent character patterns changed to obey modern Danish orthography.

VRT transformation: After annotation with token-level categories and metadata, the data were transformed to a VRT file (vertical format) for indexing in Corpus Workbench (CWB). Format: One token per line delimited by <corpus>, <text>, and <sentence> XML elements. The XML elements contain attributes with metadata. The tokens are annotated with the above-mentioned token-level annotations, separated by tabs. For more information about the metadata, see below.

The data are available as:

• OCR-corrected full volume text files
• Normalized full volume versions of these text files
• A single VRT file containing the whole corpus.

7. Limitations

A standard limitation of data preprocessed and annotated using automatic natural language processing tools and procedures is that the results are not perfect. Thus, basically all the layers of the data can be assumed to be flawed:

• Text data: The raw texts come from OCR scans of the physical book volumes. This process is not perfect, and although we have taken steps to mitigate errors, the basic text layer of the data can still be expected to have OCR errors (or wrong corrections) in 2-3% of tokens.
• Normalized data: The normalization to modern Danish spelling as such should not be expected to be perfect either. We currently do not have estimates of the error rate in the normalized data.
• Grammatical annotations: These are also added using automatic tools which cannot be expected to yield perfect results. We currently do not have estimates of error rates in the grammatical annotations.
• Metadata: The metadata are hand-curated by literary scholars and should be close to perfect. However, the occasional human error can of course not be ruled out.

8. Metadata

The metadata was curated with the help of students (Lasse Stein Holst, Lene Thanning Andersen, and Kirstine Nielsen Degn) on the basis of Dansk Bogfortegnelse (1861-), https://www.litteraturpriser.dk/, Ehrencron-Müller: Anonym- og Pseudonym-Lexikon (1940) as well as additional literary and bibliographical handbooks.

Among the metadata categories are the following:

• file_id
• filename
• [author] firstname
• [author] surname
• [author] pseudonym
• [author] gender [m/f/unknown]
• [author] nationality [da/no/unknown]
• title
• subtitle
• volume
• year [of publication]
• pages [in total]
• illustrations [y/n]
• typeface [gothic/roman]
• publisher
• price

9. Disclosure and Ethical Review

Funding for the creation and curation is supplied by The Carlsberg Foundation through a Young Researcher Fellowship awarded to Jens Bjerring-Hansen, University of Copenhagen.

In terms of data management, the project data (novels from 1870-1900) consist of imaginative texts by non-living authors. The texts are out-of-copyright. From a GDPR perspective, the biographical, bibliographical and demographic data are historical as well as non-sensitive.