Estonian Wordnet started in 1998 and since then at least once a year a new version has been released. The paper gives an overview of the main indicators for every year, as well as highlighting some problems that have been solved or as yet have to be solved.
There is an old legend about the city of Tallinn, the capital of Estonia, that is situated on the northern coast of the country between the sea and a lake called Ulemiste. Once upon a time, a legend tells, there lived an old gray man, the gremlin of the lake. He used to ask the guards of the city whether the city was nished. Nobody was allowed to answer \yes, it is nished," as then the gremlin would have opened the gates of the waters and ushed the city down into the sea.
This old legend gives us a hint why creators of large and complicated language resources such as dictionaries, lexicons and thesauri, not to mention wordnets, tend to answer about the state of their product \well, under construction," instead of \yes, it is completed".
Whatever the reason, the users of such resources are not willing to wait forever for a completed resource. Half a loaf is better than no bread. Thus, there is a pressure from the users for a (at least intermediate) release of the resource. Should a grant be involved, the creators would be under nancial pressure too.
Here we are going to present the twenty year long history of Estonian Wordnet (EstWN), and try to answer the question whether it is ready or not.
The history of EstWN started back in 1998 when the Estonian team joined
the EuroWordNet-2 project.1 Back in that time, the only available database
following principles of WordNet structure was Princeton WordNet. The
EuroWordNet project followed the same principles, but additionally had the possibility to
link each synset in di erent languages to a central InterLingualIndex (ILI) that
was then based on Princeton Wordnet version 1.5. The EuroWordNet database
and software are described in the rst place by
EuroWordNet/ of
EuroWordNet project, see https://www.illc.uva.nl/
In construction of EstWN we used the expand method as a starting point.
Base concepts from English were translated into Estonian as a rst basis for
a monolingual extension. The extensions have been compiled manually from
Estonian monolingual dictionaries2 ( rst of all Explanatory Dictionary of the
Estonian Language) and other monolingual resources like terminological
collections. In this sense, EstWN applies a hybrid method including both expand and
monolingual techniques. EstWN includes nouns, verbs, adjectives and adverbs,
and some multiword units among them. The EstWN data is available under a
CC-BY-SA license and can be accessed via WordTies3 and TEKsaurus query
tool4. For details, see
For EuroWordNet project a new tool was developed. It was made by Novell
(member of Groupwise software) and used the Flaim database format. It was a
desktop tool: there were neither shared les nor a central server. For exporting
and importing data, a plain text format based on GEDCOM was used. This le
structure was originally intended for genealogical use
The le is structured hierarchically, consisting of level numbers, elds, and
eld values. A wordnet of a particular language consists of synsets. At every
synset, there were descriptions of synset members, internal relations between
synsets, equal relations (the links to ILI) between synsets in di erent languages,
and properties or property values. The le syntax used to export a wordnet is
described by
As no client-server model was used, each collaborator worked individually and had to export her results via this export text le.
The Estonian team did not use any version control software back in these days, and in order to maintain con ict-free development, certain strict rules were applied and some helper scripts were used.
Each member was working on a speci c topic, and added only synsets and
literals strictly connected to the topic that was assigned to him. However, it is
not possible to work in complete isolation and there were frequent discussions
on some concept or word. It was not always possible to reach common ground,
but most of the arguments were settled
The individual les from each team member were sent to the project manager, who carefully imported each version individually and checked manually whether there were no overlaps in the vocabulary or synsets. There were always some lost and found synsets, relations or synset members after each manual upgrade. To decrease possibilities for such mistakes, some strict rules and some bash scripts
were implemented in fresh data collected from team members. It was a laborious e ort, and as we see in section 3, some inconsistencies still have slipped in.
When a certain stage in development of EstWN was completed, or when it was time to report some results, the individual les were merged, checked for duplicates and inconsistencies, provided with a version number and released. 3
There are 27 letters in the Estonian alphabet. Besides the letters found in
USASCII, there are a, o, u with umlauts (a, o, u), o with a tilde (~o) and s and
z with a caron (s, z). See
Although best attempts were made to use only one single code page for EstWN, some encoding bugs have sneaked in. Since the Polaris tool was too old to support any Unicode encodings, so the work les had to be in some extended ASCII code page. As the number of lexicographers grew along with the number of computers used, it was practically impossible to keep track of the code pages actually in use. In fact, this problem only emerged around 2010 when we started to convert the export les to UTF-8.
The code pages that were in use are rather similar, there is no problem with the letters with umlauts. The rst problems rose with O~ /~o (letter O with the diacritic mark tilde) that shares common code point with O} /}o (letter O with double acute accent) in Windows 1252 and Windows 1250, respectively. These characters can be hard to separate, depending on the font design. Characters S/s (letter S with a caron) and Z/z (letter Z with a caron) have di erent coding points in these code pages. If one lexicographer used Windows 1252 and another Windows 1250, after merging of the two les it could happen that synsets containing S/s and/or Z/z were duplicated, one of them having nice literals and the other garbage. That led to synsets that di ered in literals when really they should have been the same .
The \umlauts" and \carons" were not the only causes of trouble. In the process of creating the lexicon, lexicographers at times copied parts of text from other sources. The copied parts of text occasionally included typographic characters (quoting marks, em and/or en dash), mathematical symbols ( degree symbol), foreign characters from other languages (a, ).
Finally we included only those characters that are in the extended Estonian alphabet. By doing this we lost some information, mostly foreign characters and typographic symbols, but the conversion was possible without any encoding errors. 5 27 is o cial number of characters in Estonian alphabet. This number does not include some letters found in US-ASCII: c, q, w, x, y. With all these included we will get the Estonian extended alphabet that contains 32 letters. If we will omit all \foreign" letters, we will get the number 23. This number includes the umlauts and o~; in addition to earlier mentioned letters, \carons" and f are excluded too.
The numeration of EstWN releases has been very simple: the number of each subsequent version is incremented by one. The last number is 73, but this common ground does not mean, that we have exactly 73 releases of EstWN, and that all of them are numbered in this manner. There are 69 versions of EstWN at the Gitlab code repository6 that is hosted at the Center of Estonian Language Resources (CELR)7. Most of them have version numbers, but two of them are identi ed by date. Those two we denoted as derivations of kb53 in Table 1.
For each new version of EstWN the summary of statistical measures has been made.
For the present overview we made an excerpt of the table, including the very rst version (kb01 released on June 5, 1998), and thereafter the last version from each year. The date of the \last" version of the year varies a lot, from August to mid-December. The number of versions per year varies also. There were 13 releases in 1998, in 2006{2008 there was only one release per year.
The capacity of a wordnet can be measured by various indicators. The most prominent ones are the number of synsets, and the number of lexical items. A more precise indicator is number of synsets by part-of-speech. As a wordnet is also about semantic (or lexical, or lexico-semantic) relations, the number of relations is informative as well. In order to link to other lexicons, external references are important. In EuroWordNet project the equal relations or ILI relations served this purpose .
For EuroWordNet, the number of the Base Concepts (see Vossen et al. 1998) served as one of the indicators. It is not listed in this paper, but according to our raw data the Base Concepts were included in EstWN kb14 ( rst release of 1999) already, and their number has remained greater than 2,300 since then.
Following the running statistics has in uenced the process of extending EstWN. This can give insight to features, in which the data is missing. One of the aims to complement EstWN has been adding semantic relations, the decreasing number of synsets without semantic relations is a good indicator of this process.
The last row in Table 1 entitled \Synsets with many hyperonyms" lists the
number of synsets that have more than one hyperonym. Although more than
one hyperonym per synset is a common practice in wordnets, their existence
may cause anomalies in the wordnet hierarchy, as
Better overview about the dynamics of the number of synsets by part of speech is given in Figure 1. The chart shows the number of noun, verb, adjective and adverb synsets cumulatively. Instead of a linear growth there is a S-shaped curve, which has a plateu at the beginning of the Millenium and a sharp burst in 2010.
According to Figure 1 there is a increase in number of synsets from 2004 to 2005, but at the same time the number of synsets without relations, both internal and ILI, has changes too. This may be the co-e ect of automatic addition of new synsets via derivations. For some derivations it was possible to make semantic relations in the same manner, but not for all of them. The method and problems are described by Kahusk, Kerner, and Vider (2010).
As for kb73, currently the the last version, the number of synsets without
relations is zero. EstWN kb73 will probably remain the last version to be linked
to Princeton WordNet ver. 1.5, the next version will be linked to to 3.0 at least,
and perhaps we will use a new numbering scheme as well. Kb 73 is also browsable
by the current version of TEKsaurus tool
As EstWN is compiled according to Estonian language, the meaning of words in Estonian is considered instead of translating synsets from English wordnet. This practice has lead to a rather small number of eq synonym relations in EstWN (see Figure 2). There was a rather broad spectrum of equal relations used in EuroWordNet project, 15 equal relation types (besides eq synonym) are used in EstWN. That explains the relatively high number of other ILI relations on Figure 2.
The most numerous of the other ILI relations are eq has hyperonym and eq near synonym. Both of them are relations that are used to cope di erent granularity and coverage of synsets in Estonian and many other languages. There are ve versions of EstWN (see table 2) described in the CELR language resources metadata registry META-SHARE8. They have been viewed more than 600 times, more than 330 views imputed to the oldest version. As for downloads, the number is 113, the latest version has been downloaded most often.
The metadata from our repository is harvested by CLARIN9 to the Virtual Language Observatory (VLO)10, and META-SHARE11. As for the last registry, search for \wordnet" gives 74 answers12, if Estonian is speci ed as the language, 7 of them will remain in results. In VLO 's case, the search for \wordnet" gives 106 answers, after limiting language to Estonian, 10 of the results remain. Although the numbers are not very large, it may take for a while to nd out the last version. Probably version 73 will not remain last version of the EstWN, and every next version will add some more confusion.
We may assume, that generally the user would be interested in nding the last version of the wordnet, but it may not always be so. A speci c version of the wordnet or several (more than one) legacy versions may be needed in order to replicate a particular experiment or to develop a resource. If we poured all the more than 60 legacy versions of EstWN into our repository, then it would over ow not only our metadata registry, but CLARIN VLO and www.metashare.org too.
An elegant solution for this problem would involve an elaborated CMDI pro le with references to older and newer versions of the resource.13 However, an upgrade of harvesting sites would be required as well. This would probably work for VLO, but not for www.meta-share.org or other harvesting services (our registry is harvested by Estonian E-Repository too).14
As a compromise, we are going to describe metadata of all the legacy versions as single collection of resources and upload it to our repository. 6
Development of EstWN is supported by the National Programme for Estonian Language Technology during 2006{2017.15 9 https://www.clarin.eu/ 10 https://vlo.clarin.eu/ 11 http://www.meta-share.org 12 The search site redirects to http://metashare.ilsp.gr:8080/ 13 https://www.clarin.eu/cmdi1.2 14 https://www.e-varamu.ee 15 https://www.keeletehnoloogia.ee/en?set_language=en Vossen, Piek et al. (1998). Set of Common Base Concepts in EuroWordNet-2.
Tech. rep. url: https://www.illc.uva.nl/EuroWordNet/docs.html.