=Paper=
{{Paper
|id=Vol-2552/Paper23
|storemode=property
|title= Automatic Identification and Classification of Ambiguous Toponyms
|pdfUrl=https://ceur-ws.org/Vol-2552/Paper23.pdf
|volume=Vol-2552
|authors=Kirill Boyarsky,Dmitry Kanevsky,Darja Butorina
}}
== Automatic Identification and Classification of Ambiguous Toponyms ==
https://ceur-ws.org/Vol-2552/Paper23.pdf
Automatic Identification and Classification of
Ambiguous Toponyms ∗
Kirill Boyarsky 1 Eugeny Kanevsky 2
Boyarin9@yandex.ru eak300@mail.ru
Daria Butorina 1
daanareevna20@gmail.com
1
ITMO University
2
Institute for Regional Economy Studies RAS
Saint Petersburg, Russian Federation
Abstract
Currently, various versions of TextMining technology are widely used, among other
tasks, to allow retrieving various information from documents. In particular, the infor-
mation can refer to certain named entities: personalities, geographical objects, etc. In
this case, there inevitably arise problems associated with the ambiguity of names. The
paper studies the potentialities of disambiguation for Russian texts of political, artistic,
and highly specialized nature using the toponyms. We identified words that form col-
locations with toponyms (marker words). It is shown that the use of short lists of such
words can significantly increase the reliability of identifying a toponym and determining
its type.
Keywords: toponyms, ambiguity, text analysis, parser, marker words
1 Introduction
The mass media produce daily a plethora of data, in particular, the news that describe various
events taking place in one or another region of the globe. More often than not, of importance
in regional news are the so-called named entities. The discovery of named entities is a
key problem of information retrieval (structured data) from unstructured or semistructured
documents [Starostin et al, 2016]. Its main point is to find the names or identifiers of objects
of a certain type in the text. For the first time ever, the problem was stated as far back
as 1996 at the Message Understanding Conference, where the entities under consideration
were organizations, places, people, and some numeric expressions. Later, it was examined
∗ c
Copyright 2019 for this paper by its authors. Use permitted under Creative Commons License Attri-
bution 4.0 International (CC BY 4.0).
at the conferences on computational natural language learning (CoNLL) CoNLL-2002 and
CoNLL-2003 [Tjong, Sang and De Meulder, 2003].
The named entity is today understood to be an object of a certain type, which has
a name or an identifier. Which types (classes) the system distinguishes is defined within a
particular task. Typical for news texts are PERSON, ORGANIZATION, LOCATION, and
MISCELLANEOUS (numeric expressions, dates, events, slogans etc.). In a few cases, the
number of classes may be greater: studies are known to involve 29 [Brunstein, 2002] or even
200 [Shinyama and Sekine, 2004] of those, including, among others, food, books, events,
animals, plants, airports, and so on.
The discovery of named entities can be used for text classification, sentiment analysis,
paraphrase identification etc. The importance of the problems (as well as the complexity
of solving them) is substantiated by holding regular contests of the corresponding software
solutions [Starostin et al, 2016, Loukachevitch and Rubtsova, 2016, Panchenko et al, 2018].
Of great importance in solving the problems is the correct extraction from the texts of
those geographical entity names (toponyms), to which the given message may bear relation.
As noted in [Lieberman snd Samet, 2011], the toponym retrieval problem is rather involved.
Because of the homonymy, one and the same word may designate a geographical entity or that
completely unrelated to toponyms.
These particular features cause the analysis to use more accurate models than just a “bag
of words”, where the word order in the text is disregarded altogether [Barsegyan et al, 2007].
A method of the kind is the construction of a subordination tree [Testelec, 2001]. Yet, even
following this way, one has to resolve issues arising in connection with the high degree of
homonymity of Russian texts.
2 Problem Statement
The means used to analyze texts is parser SemSin [Boyarsky and Kanevsky, 2015] that per-
forms the in-depth syntactic analysis of Russian texts. It employs the expanded version of the
dictionary by V. A. Tuzov [Tuzov, 2004], its size exceeding 194,000 lexemes (about 170,000
words). The classifier has been expanded to 1,700 classes. The text analysis proceeds under
control of production rules [Boyarsky and Kanevsky, 2011a, Boyarsky and Kanevsky, 2011b].
Accomplished along with the syntactic analysis of a sentence are the grammar and POS dis-
ambiguation, phrase segmentation, the construction of syntactic dependency tree, and, fairly
often, the semantic disambiguation as well.
The parser allows adjusting the semantics to a certain subject domain. To do this, it
suffices to pretrain the parser: process several texts for the selected subject domain and
identify a number of homonymic lexemes that distort the parsing semantics. The lexemes
along with their classes are placed into a special file, which ensures the automatic elimination
in all subsequent analysis sessions. As evidenced in practice, the number of such lexemes is
not large: selecting 35 lexemes to adjust the parser to analyze texts from a rather specific
domain of historical shipbuilding proved to be enough [Artemova et al., 2015].
The parser operation result, presented as an xml file, contains the lemma and complete in-
formation on grammatic (part of speech, gender, number, case . . . ), syntactic (relation types),
and semantic (classifier class) properties of words in sentences. The data are a “semi-finished
product” of a kind and can be used for further analysis and retrieval of any information of
interest. This is why making the ambiguity as low as possible as well as raising the parsing
2
accuracy is important.
The present work aims to study the feasibility to establish the following:
− is a capitalized word a toponym?
− which toponym does the word belong to?
− which syntactic relation connects the word to the subordination tree?
When analyzing toponyms, it must be borne in mind that no reference book is capable of
covering all geographic names. So, one should be ready to encounter an unknown name in the
text. It can be a word either absent in the dictionary at all or having another meaning. The
present work disregards the situation when a toponym coincides with a common vocabulary
word (Belukha [the mountain] and Belukha [a white porpoise], Tigr [the river Tigris] and
tigr [a tiger]). It typically suffices in such case to carry an elementary graphematic analysis as
to the word capitalization.
In addition, no task is set to find the geographic location of an object. Apparently, the
problem in the general case has no solution without a detailed analysis of large text fragments.
Now, there is Verblyud [camel] mountain both in Pyatigorsk and in Kamchatka, while there
are at least three Chyornaya rechka [black river] within 30 km distance from St. Petersburg.
In any case, this is not a task for a parser.
About 90 classes in the semantic classifier we use contain words that are proper names,
with 30 of them being allocated for identities of toponyms. Table 1 lists merged classes that
contain toponyms or proper names, homonymic to toponyms, and the number of lexemes in
the classes.
Table 1: Semantic classes of proper names
Semantic classes Number of lexemes
Terrestrial natural objects (mountains, deserts, islands. . . ) 700
Natural water bodies (rivers, sees, gulfs. . . ) 1200
Cities — settlements (cities, towns, streets. . . ) 4600
Counties — administrative units (countries, regions, provinces, states. . . ) 800
Astronomical objects (stars, constellations, planets) 140
Last names 13700
First names 4400
Mythological creatures 120
Institutions (industry, research, sports. . . ) 2700
Enginery (automobiles, ships, weapons. . . ) 300
When parsing, an attempt was made to establish uniquely, if possible, to which class the
entity named with a capitalized word belongs.
To assess the parsing quality, a comparison with the performance of parser ETAP-4 was
undertaken [ETAP-4, 2019].
3
3 Semantic homonymy
The text analysis starts, before anything else, with preprocessing, when the text is split into
sentences and words and the morphological properties of words are found. Then, the text is,
word by word, input to the syntactic-semantic analyzer. The module of rules for processing
toponyms is enabled upon detection of capitalized words. Further on, the word will be called
a target one.
Termed as semantic homonymy is the situation when the target word names objects of
different nature: settlements, rivers, islands, planets etc. Some cases of semantic homonymy
are presented in Table 2. The selection used the Great Encyclopedic Dictionary [BES, 1997].
The coincidences of names are grouped into conditional classes with somewhat greater degree
of detail than in Table 1.
An asterisk in Table 2 marks the cases when not more than 10 coincident names fall into
two classes, two asterisks—from 11 to 20, and three asterisks—more than 20.
Note that one and the same word may mean both a toponym and a “non-geographical” en-
tity (company name, first/last personal name etc.), with graphematic analysis yielding nothing
of use here.
The most frequent situations are those, when a settlement name coincides with a last
personal name, there are more than 100 such cases. Here, the city may be named after
someone (Kirov, Korolyov, Houston), or there can be a mere coincidence (London, Khotyn).
The coincidence city—first name is typically of a random nature (Lyon, Olympia, Milan).
The coincidences of city names with those of rivers are, as a rule, related to the geographical
location of settlements (Aldan, Kabul, Narva).
There are many cases of coinciding names of islands and island states situated on them
(Haiti, Ireland, Cuba). The same goes for a number of city names (Dikson, Kodiak, Zanzibar ),
though the city of Vancouver, e. g., is not situated on the Vancouver island.
To find the target word semantics, the closer context is analyzed.
4
4 Abbreviations
First of all, the presence is checked of a left-contiguous abbreviation, which may mean a
geographical entity (g. [mountain or sity], o. [island], oz. [lake], pos. [settlement], r. [river]),
an address (per. [alley], pl. [square], pr./prosp. [avenue], ul. [street]), or a person (g. [Mr.],
gg. [Messrs.]). If an abbreviation for a geographical entity is detected, the capitalized word is
thought a toponym. For instance, r. Volga [the Volga river] will be expressly a river rather
than a car brand or a sports club.
As typical, the situation is complicated by homonymy, this time, of abbreviations. The
abbreviation “o.” is easy enough to differentiate from a preposition (this has no dot) and
from an initial (this is capitalized), leaving the options “ostrov ” [island] or “otets” [father].
Parser ETAP-4 chooses “father” in all cases, even in the sentence Ya posetil o. Madagaskar [I
visited the Magadascar island]. Our parser treats “o.” as an island, the target word takes on
the island class independently of the presence in the dictionary and its dictionary meaning.
For instance, o. Sumatra [the Sumatra island] (present in the dictionary as an island), o.
Vozneseniya [the Ascension island] (present in the dictionary, but with a different meaning),
o. Baratang [the Baratang island] (absent from the dictionary). A problem remains how to
interpret the abbreviation in combinations of o. Ivan type. It turns out that the islands
named after masculine names are quite few, so one can reasonably assume “o.” to have
meaning “otets” [father] in that case only, if the target word is a dictionary word of the
masculine gender.
Even more confused is the situation with the abbreviation “g.”, which may mean “gora”
[mountain], “gorod ” [city], “gospodin” [mister], “god ” [year]. ETAP-4 produces interpretation
“gorod ” [city] for target words, present in the dictionary, and “god ” [year]—otherwise. Upon
detection of the left-contiguous abbreviation “g.” for the target word, our system goes on
analyzing the left context, and the interpretation “gorod ” [city] is selected only if a numerical
token is detected. The interpretation in the rest of cases is decided by the target word class:
“g.” in the combination g. Novgorod means “gorod ” [city] and takes on the settlement class,
while “g.” in the combination g. Everest is already “gora” [mountain]. All options are retained
for words that are absent from the dictionary. In individual cases, one succeeds to refine the
class during later parsing.
5 Determiners
Determiners will be understood as the words that directly name a toponym class: river, island,
city etc. It should be noted that the number and case of a determiner is exactly the same as
those of a toponym.
If there is a determiner and the toponym is in the dictionary, the semantic disambiguation
presents no difficulty. So, Baikal in combination ozero Baikal [the Baikal lake] is the lake
rather than the drink; the relation is “Name” (its counterpart in ETAP-4 is the appositive
relation).
If there is a determiner, but the target word is not in the dictionary as a toponym (Ya
priyekhal na ostrov Zub [I came to the Zub island]), the parsing is correct both in SemSin and
ETAP-4. Note that the subordination tree, constructed in ETAP-4, has no semantic labels,
so that the correctness of the semantic class cannot be verified. If the word is absent from the
dictionary altogether (ostrov Aogasima [the Aogashima island])—the class and relation type
5
settings in SemSin are correct, while ETAP-4 yields the relation “quasi-agent”, i. e. fails to
identify the word as a name.
If the toponym is in the dictionary, but there is no determiner, the semantic disambiguation
is attended with great difficulties that one succeeds to overcome in individual cases only.
Consider the sentence Ya priyekhal na Lenu [I came to the Lena]. Upon the morphological
parsing, the word form Lenu has four meanings: the dative of lexeme Len (either a land
allotment, or an administrative unit in Sweden) and the accusative of lexeme Lena (either a
river, or a personal name). The two first options are discarded by the results of the graphematic
analysis (they should have been capitalized), the last two remain.
In the course of construction of the full subordination tree, however, the analysis of the
semantic classes that are required by the verb “priyekhal ” [to come] allows disambiguation of
the homonymy, leaving the river name only. ETAP-4 yields label “animate” for lexeme Lena.
Understandably, both options remain in the sentence Ya uvidel Lenu [I saw Lena].
6 Marker Words
In some cases when the determiner is absent, one can try to ascertain the semantics of a proper
name by marker words. The latter will be understood as the words often used with toponyms
of a certain type. For instance, such words can be techeniye [flow], ruslo [bed], ust’ye [mouth]
for rivers, vershina [summit], sklon [slope] for mountains, bereg [shore], ostrov [island] for lakes
and so on. A feature of marker words is that they augment a proper name using the genitive
(this is the relation “quasi-agent” in ETAP-4) rather than as a name.
We have studied the issue how accurately this or that marker word allows unique iden-
tification of the semantic class of a toponym. The results of the expert analysis of several
thousands of sentences are presented in Tables 3–5.
It shall be noted that our study used much greater number of candidates to marker words,
many of which are out of the tables because of their excessive ambiguousness and low percent-
age of co-occurrence with the corresponding toponym class. So, Table 3 contains no words glu-
6
bina [depth], dlina [length], dno [bottom], naberezhnaya [quay], omut [whirlpool], stremnina
[quickwater], and shirina [width], though they are occasionally found in front of a river name
(. . . stol ’ nepokhozhem ni na Ligovku, ni na naberezhnyie Nevy [. . . so unlike either Ligovka, or
the Neva quays]). Table 4 contains no word naberezhnyie [quays], though it may stand in front
of a city name (Mazankovyie i kamennyie dvortsy zapolnili naberezhnyie Peterburga [wattle
and daub houses and stone palaces filled the quays of St. Petersburg]), while Table 5 con-
tains no words dolina [valley], obryv [precipice], peshchera [cave], pik [pinnacle], poverkhnost’
[surface], ushchel’e [ravine], khrebet [ridge] (Prakticheski vse gornyie ushchel’ya Issyk-Kulya
prigodny dlya trekov [Practically all mountain gorges of Issyk Kul are fit for trekking]). First,
selected as marker words were those of class “landscape”, which often occur together with
names of rivers, cities, and mountains. Next, the Russian National Corpus (RNC) [NKRYa,
2019] was sampled under an additional condition of the right-contiguous capitalized word (a
candidate to toponyms). Further, eliminated were the cases when there was a punctuation
mark between the marker and the candidate. The resulting list was appended with sentences,
retrieved under the same rules from news, literary, and sports texts of total size 55 mln. of
words.
Consider the example: On vozglavil ekspeditsiyu v Man’zhuriyu s tsel’yu izyskaniya tropy,
kotoraya spryamlyala by izluchinu Amura [He headed an expedition to Manchuria for the
purpose of finding the path, which would straighten the Amur loop]. The word Amur [Amur]
as a proper name has at least five meanings: a river, a planet, a firm, a sporting club, a god.
With consideration of the marker word izluchina [loop], a single meaning is left, a river.
As seen from Table 3, there is probability 95 % and greater for the co-occurrence of some
markers with river names, the percentage for others being much lower. The latter include, e.
g., the word vodorazdel [watershed]. The word in Russian has a very versatile semantics, and,
as the table shows, there are just 44 % cases of co-occurrence with a river name. The marker
occurs along with other toponyms and proper names in 50 % of cases:
• Akademik Gmelin. . . opredelil vodorazdel Indiyskogo i Ledovitogo okeanov. [Academician
Gmelin. . . identified the watershed of the Indian and Arctic oceans]
• Topograf . . . zasnyal vodorazdel Ural’skogo khrebta. [The surveyor mapped the watershed
of the Ural range]
• “Pamyatnik” zastyl vodorazdelom Pushkina myortvogo i Pushkina zhivogo v nashem soz-
nanii. [“The Monument” froze as a watershed of dead Pushkin and alive Pushkin in our
mind]
However, given that practically all names of mountain systems and sees are found in
dictionaries, one can confidently predict that the words “Baranikhi ” and “Pogindeny” in
the sentence . . . Dolinoy reki spustilis’ vniz i v neskol’kikh verstakh ot khrebta, sluzhashchego
vodorazdelom Baranikhi i Pogindeny, ostanovilis’ na nochleg [We came down by the river
valley and put up for the night in a few miles from the ridge that served as a watershed of
Baranikhi and Pogindeny] are river names. In the course of parsing, this allows predicting the
semantics of the capitalized words that are not in the dictionary.
The word bereg [bank, shore] occurs typically with names of rivers, lakes, islands, sees,
countries, regions. For this reason, a river name rather than a personal one is univocally
left in the sentence Lager’ razbili na beregu Yany [The camp was pitched on the Yana bank].
The parsing can be luckily refined in some cases by expanding the context and analyzing the
7
adjectives, left-contiguous to the marker word. Thus, one succeeds to distinguish mineral’nyie
vody [mineral waters] from vody Volgi [Volga waters], and recognize that what is meant in
the sentence Gorod vyros na oboikh beregakh El’by [The city grew on both Elba banks] is a
river rather than an island. Typical for the marker bereg in the meaning “river bank” are also
the adjectives levyi [left] and pravyi [right]; for the marker basseyn in the meaning “swim-
ming pool”—detskii [wading], krytyi [indoor], otkrytyi [outdoor], plavatel’nyi [swimming], and
sobstvennyi [private]; for the marker techeniye [flow] in the meaning “river flow”—verkhniy
[upper], nizhniy [lower], sredniy [middle].
It shall be noted that even here the parsing can be refined in some cases by expanding the
context and analyzing the adjectives, left-contiguous to the marker words prigorod [suburb] or
predmest’e [outskirt]. If the words are preceded by an adjective derived from a city name, the
target word is a name of a suburb or an outskirt as such: Tak, v pechal’no izvestnom svoyey
tyur’moy bagdadskom prigorode Abu-Graib vplot’ do posledney voyny v Irake raspolagalsya
natsional’nyi bank semyan [So, Bagdad suburb Abu Ghraib, notorious for its prison, hosted
the National Seedbank continuing until the last war in Iraq] (suburb name) vs Po etomu
povodu v prigorode Khabarovska sostoyalas’ ofitsial’naya tseremoniya [This was the occasion
for an official ceremony to be held in a suburb of Khabarovsk] (relation by the genitive). In
both cases, ETAP-4 yields the relation “quasi-agent”. In some cases, one succeeds even to
disambiguate the toponym semantically: Ni doma, ni mashiny zdes’ v prigorode Vankuvera
pochti nye zapirayutsya [Here, in the Vancouver suburb, neither houses, nor cars are hardly
ever locked] (Vankuver [Vancouver] is a city and an island, the analysis leaves a city).
Iz izvestnyakov pochti polnost’yu slozheny zapadnyie sklony Urala [The western slopes of
8
the Urals are almost completely composed of limestone] (Ural is a river, mountains, region,
and a sporting club; the analysis leaves mountains).
In some cases, the correct classification of a toponym is quite difficult. So, the river names
Ona and Talaya in the sentence Za nim v istokakh Ony and Taloy nachinalas’ yego tayga
[Beyond it, in the waterheads of Ona and Talaya, his taiga started] are detected by neither
our parser, nor ETAP-4.
7 Conclusion
The analysis of the usage of toponyms in Russian texts shows that the degree of their
homonymity is rather high. Automatic classification of toponyms is complicated also by the
existence of a vast number of geographical names that are absent from the common vocabulary.
To resolve the problem, it has been proposed to use marker words, which allow inferring
the presence of a toponym circumstantially, in addition to the determiners that indicate the
toponym class directly. So, the markers izluchina [loop], nizov’ye [lower reach], mezhdurech’ye
[interstream], pritok [tributary], verkhovye [upper reach], and ust’ye [mouth] precede a river
name in the overwhelming majority of cases. Their presence secures the accuracy of at least
95 % of interpreting a word that is absent from the dictionary, but is capitalized, as a river
name.
It has been noted that even recognition of a capitalized word as a toponym fails often to
determine the type of a geographical entity unequivocally (Table 2). “Microdictionaries” of
marker words have been compiled for toponyms of the most homonymic classes: rivers, cities,
and mountains. Their use has been shownto allow raising the accuracy of determination of
toponyms and, in some cases, to allow establishing their class.
The results hold out the hope that marker words are capable of disambiguating not only to-
ponyms, but also other semantic classes. For instance, serving as markers for last names could
be such words as mneniye [opinion], vizit [visit], otstavka [retirement], and so on. Compilation
of lists of such words could be the subject of further research.
References
[Starostin et al, 2016] Starostin A. S. et al (2016) Evaluation of Named Entity Recognition
and Fact Extraction Systems for Russian / Starostin A. S., Bocharov V. V., Alexeeva
S. V., Bodrova A. A., Chuchunkov A. S., Dzhumaev S. S., Efimenko I. V., Granovsky
D. V., Khoroshevsky V. F., Krylova I. V., Nikolaeva M. A., Smurov I. M., Toldova S.
Y. // FactRuEval 2016: Komp’yuternaya lingvistika i intellektual’nye tekhnologii: Po
materialam ezhegodnoy mezhdunarodnoy konferentsii “Dialog” (Moskva, 1–4 iyulya 2016
g.). Vyp. 15 (22).—M.: Izd-vo RGGU.
[Tjong, Sang and De Meulder, 2003] Erik F. Tjong, Kim Sang and Fien De Meulder (2003) In-
troduction to the conll-2003 shared task: language-independent named entity recognition.
// In Proceedings of the seventh conference on Natural language learning at HLT-NAACL
2003 - Volume 4, CONLL ’03, Stroudsburg, PA, USA, 2003. P. 142—147.
[Brunstein, 2002] Brunstein A (2002) Annotation Guidelines for Answer Types // BBN tech-
nologies (2002). URL: http://www.ldc.upenn.edu/Catalog/docs/LDC2005T33/BBN-
Types-Subtypes.html.
9
[Shinyama and Sekine, 2004] Yusuke Shinyama and Satoshi Sekine (2004) Named entity dis-
covery using comparable news articles. // In Proceedings of the 20th international con-
ference on Computational Linguistics, COLING ’04, Stroudsburg, PA, USA, 2004. P.
848–853.
[Loukachevitch and Rubtsova, 2016] Loukachevitch N. V., Rubtsova Y. V (2016)
SentiRuEval-2016: Overcoming Time Gap and Data Sparsity in Tweet Sentiment
Analysis. // Komp’yuternaya lingvistika i intellektual’nye tekhnologii: Po materialam
ezhegodnoy mezhdunarodnoy konferentsii “Dialog” (Moskva, 1–4 iyulya 2016 g.). Vyp.
15 (22).—M.: Izd-vo RGGU, 2016. S. 416-426.
[Panchenko A. et al, 2018] Panchenko A. et al (2018) RUSSE2018: a Shared Task on Word
Sense Induction for the Russian Language / Panchenko A., Lopukhina A., Ustalov D.,
Lopukhin K., Arefyev N., Leontyev A., Loukachevitch N. // Komp’yuternaya lingvistika
i intellektual’nye tekhnologii: Po materialam ezhegodnoy mezhdunarodnoy konferentsii
“Dialog” (Moskva, 30 maya–2 iyunya 2018 g.). Vyp. 17 (24), 2018. C. 546-564.
[Lieberman snd Samet, 2011] Michael D. Lieberman, Hanan Samet (2011) Proceedings of the
34th international ACM SIGIR conference on Research and development in Information
- SIGIR ’11. Beijing, China, 2011. P. 1—36.
[[Barsegyan et al, 2007] Barsegyan A.A., Kupriyanov M.S., Stepanenko V.V., Holod I.I.
(2007) Tekhnologii analiza dannyh: Data Mining, Visual Mining, Text Mining, OLAP.
2-e izd., pererab. i dop. SPb.: BHV-Peterburg, 2007. (In Russian).
[Testelec, 2001] Testelec YA.G. (2001) Vvedenie v obshchij sintaksis. M.: RGGU, 2001. (In
Russian).
[Boyarsky and Kanevsky, 2015] Boyarsky K.K., Kanevsky E.A. (2015) Semantiko-
sintaksicheskiy parser SEMSIN // Nauchno-tekhnicheskiy vestnik informatsionnyh
tekhnologiy, mekhaniki i optiki. 2015, T. 15, 5. – S. 869-–876. (In Russian).
[Tuzov, 2004] Tuzov V.A. (2004) Komp’yuternaya semantika russkogo yazyka. SPb. Izd-vo
S.-Peterb. un-ta, 2004. (In Russian). “”
[Boyarsky and Kanevsky, 2011a] Boyarsky K.K., Kanevsky E.A. (2011) Yazyk pravil dlya
postroeniya sintaksicheskogo dereva // Internet i sovremennoe obschestvo: Materialy XIV
Vserossiyskoy ob”edinennoy konferentsii “Internet i sovremennoe obschestvo”. – SPb.:
OOO “Mul’tiProdzhektSistemServis”, 2011. S. 233–237. (In Russian).
[Boyarsky and Kanevsky, 2011b] Boyarsky K.K., Kanevsky E.A. (2011) Sistema produktsion-
nyh pravil dlya postroeniya sintaksicheskogo dereva predlozheniya. Prikladna lingvistika
ta lingvistichni tekhnologii: MegaLing-2011. K.: Dovira, 2012. S. 73–80. (In Russian).
[Artemova et al., 2015] Artemova G. et al. (2015) Text Categorization for Generation of His-
torical Shipbuilding Ontology / G. Artemova, K. Boyarsky, D. Gouzévitch, N. Gusarova,
N. Dobrenko, E. Kanevsky, D. Petrova. // Communications in Computer and Information
Science, 2015, v. 468. P. 1—14.
10
[ETAP-4, 2019] Lingvisticheskiy protsessor ETAP-4 [Elektronnyy resurs] // URL:
http://www.http://proling.iitp.ru/ru/etap4 (data obrascheniya: 22.02.2019). (In
Russian).
[BES, 1997] Bol’shoy entsiklopedicheskiy slovar’ (1997) / Gl. red. A.M. Prohorov. – 2-e izd.,
pererab. i dop. – M.: Bol’shaya Rossiyskaya Entsikl., 1997. (In Russian).
[NKRYa, 2019] Natsional’nyy corpus russkogo yazyka [Elektronnyy resurs]
//URL:http://www.ruscorpora.ru/ (data obrascheniya: 19.01.2019). (In Russian).
11