=Paper=
{{Paper
|id=Vol-2084/shortplus6
|storemode=property
|title=FSvReader – Exploring Old Swedish Cultural Heritage Texts
|pdfUrl=https://ceur-ws.org/Vol-2084/shortplus6.pdf
|volume=Vol-2084
|authors=Yvonne Adesam,Malin Ahlberg,Gerlof Bouma
|dblpUrl=https://dblp.org/rec/conf/dhn/AdesamAB18
}}
==FSvReader – Exploring Old Swedish Cultural Heritage Texts==
https://ceur-ws.org/Vol-2084/shortplus6.pdf
FSvReader – Exploring Old Swedish Cultural
Heritage Texts
Yvonne Adesam, Malin Ahlberg, and Gerlof Bouma
Språkbanken
Department of Swedish
University of Gothenburg
firstname.lastname@gu.se
Abstract. This paper describes FSvReader, a tool for easier access
to Old Swedish (13th–16th century) texts. Through automatic fuzzy
linking of words in a text to a dictionary describing the language of the
time, the reader has direct access to dictionary pop-up definitions, in
spite of the large amount of morphological and spelling variation. The
linked dictionary entries can also be used for simple searches in the text,
highlighting possible further instances of the same entry.
1 Introduction
When exploring cultural heritage texts, different types of textual data require
different types of tools for accessing the contents. For larger amounts of data,
with millions or even billions of words, powerful query tools are necessary, in
combination with annotations of the text, linguistic information about words and
sentences in the texts. Although the added linguistic information, or annotation,
may appear to be of interest only to someone exploring linguistic questions, it is
in effect necessary for any user. For example, without direct access to the lemma
for each word in a text, we would have to search for all the different forms of,
e.g., searching, searches, search, searched to find mentions of people searching for
something – for a language a bit more morphologically rich than English, this
can be a true challenge.
For older historical text, the amount of text may not be as large, but we still
require tools to glean insights from these sources as the variation in this material
is typically larger than in modern material. For instance, a language’s morphology
may have simplified much over time (the case for, e.g., English and Swedish),
and increased language standardization means less variation in vocabulary and
the way words are written.
The intricacies of a language variety and of a text written in that language
variety need not present serious problems for the historical linguist who is an
expert of that variety. However, historical texts are also of interest to researchers
from other disciplines, with different areas of expertise. In the study of a histor-
ical text, such researchers might be helped by a tool that provides additional
information like dictionary definitions, and lets them quickly find related passages
in the text.
We present a tool for exploring Old Swedish (Swedish: fornsvenska) texts
(c1225–1526) – the FSvReader.1 The tool links text words to a dictionary of Old
Swedish, giving the reader quick access to dictionary definitions while reading, as
well as simple search facilities of words linked to the same dictionary entry. The
links are automatically created, which means that we can make large amounts of
text available through the FSvReader.
2 Some Peculiarities of Historical Language: Old Swedish
Exploring historical text mostly entails close reading, and depending on the
distance between the historical language and the contemporary language, this
may be cumbersome. Additional knowledge is necessary to guide and assist access
to the content. In the case of Old Swedish, the language has changed quite
considerably in certain aspects. First, there will be word forms unfamiliar to the
modern reader because of changes in morphology. As an example, the earlier Old
Swedish material still uses a system with three genders and four cases, compared
to modern Swedish two genders and two cases. Incidentally, these changes were
already set in progress during the Old Swedish period which makes the picture
even more muddled.
Secondly, some words may have disappeared from the language or may have
acquired a different meaning because of semantic change. For example, Old
Swedish aptanbakka or affton backe is replaced by fladdermus ‘bat/Chiroptera’ in
modern Swedish, while the concept bakvaþi ‘(the result of) accidentally stabbing
someone behind you’ has altogether disappeared from the Swedish language.
The Old Swedish mot ‘meeting’ developed from the noun into a preposition
meaning ‘towards’. As further examples of this problem we can consider jargon
that is particular to (a body of) historical text, such as legal or medical terms,
or standard formulations that may be assumed to have a particular meaning in
some praxis described in a text. Without access to a lexicon, such cases make
the text more difficult to understand. For Old Swedish, we have access to three
dictionaries: the more general Söderwall and Söderwall’s supplement [16] with
44,000 entries, and Schlyter [14] with 10,000 entries, focusing on law texts. These
dictionaries substantially overlap in their inventory, but crucially supplement
each other in their descriptions and attestations.
Thirdly and finally, some words can be difficult to understand because of
spelling variation. For example, the (di)graphs t, th, þ2 were used interchangeably
in certain contexts, as were d, dh, þ, which creates a lot of uncertainty surrounding
the realization of dental stops and fricatives alone.
Due to the combined sources of variation – morphological change, changes in
lexical inventory, and spelling – even identifying occurrences of one and the same
1
It has previously been briefly introduced in [2] and [1].
2
Neither the dental fricative, nor the corresponding graph þ, are present in contempo-
rary Swedish.
word can be difficult. For instance, Old Swedish bokstaver,3 cf contemporary
Swedish bokstav, ‘letter’, can be found in the following variants, differing both in
morphology and spelling, in the editions we have digitally available: bogstaffwa
bokstaf bokstaff bokstaffua bokstaffwa bokstaffwane bokstaffwinor 4 bokstaffwom
bokstafuom bokstafwa bokstaua bokstawa bokstawane bokstawin bokstawom. Finding
those quickly and/or automatically is a challenge.
Although a researcher working with these texts will learn to understand them,
they are not easily accessible for non-expert researchers, for instance those who
are just becoming acquainted with the material, let alone for the general public.
The lack of an orthography is not only a problem for a modern reader of the texts,
but also a major obstacle for looking up words in the dictionary. The dictionary
user must both decode the form encountered in the text (requiring a passive
knowledge of the spelling conventions in a material) as well as learn how the
dictionary maker chose to write the base form for the expected lemma (requiring
an active knowledge of the dictionary maker’s spelling and/or lemmatization
conventions). In some cases, the dictionary maker caters for this need, by putting
alternate base forms as entries referring to the main entry describing the lemma.
Yet in general, learning to use the different dictionaries in itself presents a hurdle.
As an example, one of our texts contains the multi-word unit vkuaþins ord
‘insulting address’. Knowing a few possible character variations from just looking
at the text, we can guess that v can also be written u or o. However, without
knowing the dictionary authors’ spelling conventions used for main entries – and
these vary between the two main lexica available for Old Swedish – we do not
know for certain where in the dictionary to look. As it turns out, our method
for linking text words to dictionary entries presents the entries oqväþins orþ,
oqväþis ordh, and oqväþi from Söderwall. Note that even though these entries
come from the same dictionary, ord (‘word’) is once standardized to þ and once
to dh. (See [9] for specifics on the spelling standardization in the Söderwall
dictionary.) Automatic linking thus facilitates access to the historical sources,
without the reader getting lost in the dictionary – even though this is a nice way
to spend many hours.
3 The FSvReader
In the vein of dictionary look-up tools where you can read the text and get a
pop-up dictionary definition, the FSvReader5 contains a simple interface which
allows the reader to explore the texts with the help of additional information
about the words. The important difference between most such tools for modern
3
For the rest of the paper, we will follow the convention that dictionary entries and
also characters in dictionary entries are written in boldface, whereas italics refer to
actual tokens and their spelling.
4
The form bokstaffwinor is most likely a misspelling for bokstaffwinom, but we do not
know where in the manuscript-edition(s)-digitalization chain this misspelling was
introduced.
5
https://spraakbanken.gu.se/fsvreader
Top 1 Top 3 Top 10 # tokens
Late 13th century legal prose 66.0 89.9 95.6 1228
Mid 14th century biblical prose 60.5 78.6 83.6 7051
Mid 14th century legal prose 67.9 91.2 96.2 22107
Mid 15th century satirical prose 66.9 90.2 94.8 540
Mid 15th century fictional prose 64.9 89.3 96.2 2813
Average over texts 65.2 87.8 93.3
Table 1. Results in percent of correctly found lexicon entries for the lexical linking
method, among the top 1, 3, and 10 lexical links.
language and the FSvReader, is that we apply fuzzy matching to link words in a
text to entries in the historical dictionaries for Old Swedish.
3.1 Lexical Links
Fuzzy matching allows us to link text tokens to dictionary entries without knowing
the exact spelling of the lemma beforehand. The fuzzy matcher has access to a
large collection of mapping rules that relate character sequences in the lemma
string (the dictionary entry) to character sequences in the token string (the
text word). So, for instance, we know that a dictionary o may be realized as a
token o or u, or even v or w; a þ may show up as a þ, t, d,th, or dh; or that a
lemma ending in er, in principle could occur in a text with a large number of
endings capturing both the different suffixes of the strong masculine paradigms
and their possible spellings, etc. Each of these rules has an associated weight,
which encodes how likely the mapping is. The sum of the weights of all possible
ways to relate a dictionary entry to a token forms a score and for a given token,
we consider the top scoring dictionary entries as matches.
The weighted mapping rules are automatically induced from variants that are
listed in Söderwall’s dictionary. Our current fuzzy matcher knows about 171.000
rules of the kind described above. Not all rules have the convincing quality of the
given examples, however, this is to be expected of rules produced by an automatic
induction method. Note that a user of the end result of fuzzy matching is never
confronted with these rules directly.
The fuzzy matcher is an application of spelling error correction techniques
described in [6], and has been outlined in [3]. To give an idea of the quality of
the resulting lemmatization, Table 1 shows the effectiveness of our method when
considering just the highest scoring entry, the top 3 best entries and the top 10
entries, on five different corpora. The gold standard lemmata were annotated in
the context of the morphosyntactic annotation project described in [10].
As shown in the table, the correct dictionary entry is found among the top 10
entries for more than 90% of the tokens. This average is pulled down considerably
by the biblical prose, a paraphrase of the books of Moses. This text contains
many proper names, which will be linked erroneously to the dictionary. In fact,
Fig. 1. The FSvReader presents the closest lexicon entries to a word in the text, here
mothor (‘mother’) towards the end of the Church Act in Younger Västgötalagen
two out of three words for which we cannot find the correct match are proper
names. Handling them separately would thus boost performance. Looking at the
other texts, we see that considering only the top 3 matches suffices to retrieve
the correct lemma in 90% of the cases.
3.2 Exploring Text with the FSvReader
The FSvReader shows a whole text to the reader. When clicking on a word in the
text, the sidebar presents the closest lexicon entries to that word. An example
can be found in Figure 1, where part of a law text is shown, and the word mothor
(‘mother’ in singular, non-nominative case, underlined) on the second line of text
has been clicked. Words that do not have matches are marked in grey in the text,
mainly excluded items such as punctuation and numeral section headings, e.g. 4.
or §.2.
We have chosen to show up to three best scoring entries for each graphic token
in the FSvReader, to strike a balance between finding the correct entry and not
loosing the sense of overview. In addition to the best entries from the Söderwall
dictionary, we also show related entries from Söderwall’s supplement and the
Schlyter dictionary. By clicking on the abbreviation marker ‘. . . ’ we get access
to the full lexicon entry. In the example, the first match is modhir, which is
the correct entry for ‘mother’ in Söderwall’s supplement. The second suggestion
moþer ‘keen/anxious’ is an example of an incorrect match. The selection is
however typically small enough for a reader to quickly figure out what the correct
entry is.
Fig. 2. The FSvReader highlights all text words linked to the same lexicon entry, in
this case ok (‘and’) in the start passages of the Younger Västgötalagen
If appropriate, we also list entries for combinations of two graphic tokens.
For example, the two words kirkmæssu dagh are separately linked to kirkmässa
‘church mass’ (i.e., the anniversary of a church’s inauguration) and dagher ‘day’,
and together to kirkmässo dagher ‘church mass day’, which is a separate entry
in the Söderwall dictionary.
If the reader clicks on one of the lexicon entries, all words in the text linked
to this entry are highlighted. In Figure 2 we can see how one instance of the
word oc (’and’, underlined) has been clicked. Upon clicking on the (correct) entry
ok we see several different spellings – oc, och, and ok – of the same word, even
within the same paragraph. This is thus a simple way of finding other possible
instances of the same word, regardless of their spelling.
4 Related and Future Work
There has been quite a lot of work recently on automatic methods for analy-
sis of historical texts (see also the overview in [13]). Most approaches handle
spelling variation by normalizing text words to modern spelling (e.g., [4, 5, 11,
12]), although there are approaches where standardization is done to a historical
or artificial standard form (e.g., [8]). The main difference to our approach is that
we do not normalize spelling, but perform lemmatization against a lexicon. The
latter neutralizes inflection, and not just spelling.
Although there are many tools to analyze historical texts, there are few tools
where such analysis is used to display the text and help the reader access it for
close reading. A project very close in spirit to FSvReader is the digital ‘reading
edition’ of Hrafnkels saga [15],6 which presents the Old Icelandic text with links
to a dictionary, a grammar, and morphological information for words in the text.
In contrast to our approach, these reading editions appear to start from existing
manually annotated data, which means that the quality versus quantity balance
is very different from FSvReader’s.
For future work, we aim to improve the quality of the automatic dictionary
linking by incorporating further linguistic knowledge, such as part-of-speech [3],
which may help weed out irrelevant links. In addition, we aim to link to other
resources, in particular to an onomasticon, so that we may be able to correctly
recognize known proper names and also supply information about these names
and their bearers. As shown in [7] for Latin, the inclusion of an onomasticon can
drastically improve lemmatization quality. The treatment of lemmata that span
more than two graphic tokens, and, conversely, the treatment of graphic tokens
that combine several lemmata, is also planned for the future.
Further text processing, such as recognition of named entities or events, and
richer semantic mark up, are very interesting, especially from a distant reading
perspective. At the moment, however, we choose to concentrate our efforts for
the FSvReader on the lexical level.
5 Conclusions
Reading historical text can often be difficult for someone without a detailed
knowledge of the language in the text, regarding morphological and lexical
differences from the modern language, as well as spelling variation. We describe
a tool for exploring historical texts, which facilitates reading, by providing
additional information about the words. The text words have been connected
to dictionary entries with a fuzzy linking method, which gives the reader easy
access to definitions, attestations, related words, etc.
The method, while also giving us erroneous links, shows us the correct entry
for nine out of ten words in an Old Swedish text. In spite of the errors, this is
very helpful, since manually trying to find the correct entry in the dictionary can
also be difficult. We therefore hope that such a tool can help make our textual
cultural heritage more accessible.
Acknowledgments
The FSvReader and its associated resources and code are freely available and ac-
cessible through https://spraakbanken.gu.se/fsvreader. The work of Gerlof
Bouma and Yvonne Adesam was supported by the Marcus and Amalia Wallenberg
foundation (MAW 2012.0146: MAÞiR).
6
http://ecenter.uni-tuebingen.de/hrafnkels-saga/start.html
References
1. Adesam, Y., Ahlberg, M., Andersson, P., Borin, L., Bouma, G., Forsberg, M.:
Språkteknologi för svenska språket genom tiderna. Studier i svensk språkhistoria
13, 65–87 (2016), http://urn.kb.se/resolve?urn=urn:nbn:se:umu:diva-121383
2. Adesam, Y., Ahlberg, M., Bouma, G.: Processing spelling variation in historical text.
In: Proceedings of SLTC. pp. 1–2. Lund (2012), http://lup.lub.lu.se/record/
3191709
3. Adesam, Y., Bouma, G.: Old Swedish part-of-speech tagging between variation and
external knowledge. In: Proceedings of the 10th SIGHUM Workshop on Language
Technology for Cultural Heritage, Social Sciences, and Humanities, Berlin, Germany,
August 11, 2016. Association for Computational Linguistics (2016), https://aclweb.
org/anthology/W/W16/W16-2104.pdf
4. Baron, A.: Dealing with spelling variation in Early Modern English texts. Ph.D. the-
sis, Lancaster University (2011), http://eprints.lancs.ac.uk/id/eprint/84887
5. Bollmann, M., Petran, F., Dipper, S.: Applying rule-based normalization to different
types of historical texts—an evaluation. In: Vetulani, Z., Mariani, J. (eds.) Human
Language Technology Challenges for Computer Science and Linguistics. pp. 166–177.
Springer International Publishing, Cham (2014)
6. Brill, E., Moore, R.C.: An improved error model for noisy channel spelling correction.
In: Proceedings of the 38th Annual Meeting on Association for Computational
Linguistics. pp. 286–293. ACL ’00, Association for Computational Linguistics,
Stroudsburg, PA, USA (2000), https://doi.org/10.3115/1075218.1075255
7. Budassi, M., Passarotti, M.: Nomen Omen. Enhancing the Latin morphological anal-
yser Lemlat with an onomasticon. In: Proceedings of the 10th SIGHUM Workshop
on Language Technology for Cultural Heritage, Social Sciences, and Humanities,
Berlin, Germany, August 11, 2016. Association for Computational Linguistics (2016),
http://aclweb.org/anthology/W/W16/W16-2110.pdf
8. Dipper, S.: Morphological and part-of-speech tagging of historical language data:
A comparison. Journal for Language Technology and Computational Linguistics,
Special Issue: Proceedings of the TLT-Workshop on Annotation of Corpora for
Research in the Humanities 26(2), 25–37 (2011), http://dblp.org/rec/journals/
ldvf/Dipper11a
9. Djärv, U.: Fornsvenskans lexikala kodifiering i Söderwalls medeltidsordbok. No. 91
in Samlingar utgivna av Svenska fornskriftsällskapet. Serie 1. Svenska skrifter,
Svenska fornskriftsällskapet, Uppsala (2009), http://urn.kb.se/resolve?urn=
urn:nbn:se:su:diva-26468
10. Eckhoff, H., Bech, K., Bouma, G., Eide, K., Haug, D., Haugen, O.E., Jøhndal, M.:
The PROIEL treebank family: a standard for early attestations of Indo-European
languages. Accepted for publication in Language Resources and Evaluation (to
appear), https://doi.org/10.1007/s10579-017-9388-5
11. Jurish, B.: More than Words: Using Token Context to Improve Canonicalization of
Historical German. Journal for Language Technology and Computational Linguistics
25(1), 23–40 (2010), http://dblp.org/rec/journals/ldvf/Jurish10
12. Pettersson, E.: Spelling Normalisation and Linguistic Analysis of Historical Text
for Information Extraction. Ph.D. thesis, Uppsala University (2016), http://urn.
kb.se/resolve?urn=urn:nbn:se:uu:diva-269753
13. Piotrowski, M.: Natural Language Processing for Historical Texts. Morgan &
Claypool (2012), https://doi.org/10.2200/S00436ED1V01Y201207HLT017
14. Schlyter, C.J.: Ordbok till Samlingen af Sweriges Gamla Lagar, Saml. af Sweriges
Gamla Lagar, vol. 13. Lund (1887)
15. Schwabe, F.: Im stillen Kämmerlein. Zur digitalen Leseausgabe der ›Hrafnkels saga
freysgoða‹. Magazin für digitale Editionswissenschaften 4, 21–34 (2017), https:
//www.mde.fau.de/?page_id=487
16. Söderwall, K.F.: Ordbok öfver svenska medeltids-språket/Ordbok över svenska
medeltids-språket. Supplement. No. 27/54 in Samlingar utgivna av Svenska forn-
skriftsällskapet. Serie 1. Svenska skrifter., Svenska fornskriftsällskapet, Lund &
Uppsala (1884–1918/1925–1973)