=Paper=
{{Paper
|id=Vol-1899/CfWNs_2017_proc6-paper_7
|storemode=property
|title=Towards Revised System of Verb Wordnet Relations for Polish
|pdfUrl=https://ceur-ws.org/Vol-1899/CfWNs_2017_proc6-paper_7.pdf
|volume=Vol-1899
|authors=Agnieszka Dziob,Maciej Piasecki,Marek Maziarz,Justyna Wieczorek,Marta Dobrowolska-Pigoń
|dblpUrl=https://dblp.org/rec/conf/ldk/DziobPMWD17
}}
==Towards Revised System of Verb Wordnet Relations for Polish ==
https://ceur-ws.org/Vol-1899/CfWNs_2017_proc6-paper_7.pdf
Towards Revised System of Verb Wordnet Relations for
Polish
Agnieszka Dziob, Maciej Piasecki, Marek Maziarz,
Justyna Wieczorek, and Marta Dobrowolska-Pigoń
G4.19 Research Group, Wrocław University of Science and Technology, Wrocław, Poland
{agnieszka.dziob,maciej.piasecki}@pwr.edu.pl
Abstract. A revised model for the description of verbs in plWordNet - a large
wordnet of Polish is presented in the paper. The model has originated from the
previous one (the versions 2.0 and 3.0) as a combination of simplifications and
expansions. The simplifications were motivated by the collected experience of
the previous model implementation for almost 18 thousand Polish verbs. The
proposed expansions are meant to improve the description of verbs and support
provided by plWordNet for various applications.
1 Underestimated Wordnet Part
Verbs are mostly described by fewer relations in wordnets and the verb subnetwork
expresses lower density providing less information. In this paper we claim that:
● a richer structure of verb relations can be introduced into a wordnet,
● rich verb relations in a wordnet can supplement valency-based model in a
way which is attractive for many applications.
Our goal is to outline a model for the description of verbs in plWordNet which
consists of: features (e.g. verb class, aspect) and lexico-semantic relations and is
strongly based on the plWordNet general model. We are going to confine the
discussion only to the Polish language and plWordNet1 (plWN - a large wordnet for
Polish [30]), but in a perspective enabling more general conclusions. The presented
work builds on the previous model proposed for plWN 2.0 [28] and used also in the
version 3.0 [30], as well as the experience collected from its implementation for
17,391 verb lemmas. During this work it became clear that the model must be refined
in order to improve its applicability in daily wordnet editing practice and consistency
of the wordnet editor decisions.
Six verb relations were proposed for Princeton WordNet (PWN) [34]: synonymy,
antonymy, inclusive entailment (proper inclusion), troponymy (coextensiveness),
cause and presupposition. Later, since PWN 1.5, inclusion and presupposition have
been combined into entailment.
GermaNet identified troponymy with hyponymy [22]. PWN’s inclusive entailment
was renamed subevent in EWN [1], and ‘entailment’ is only a label for presupposition
(backward along the timeline) [15]. Causality is cross-categorial [23]. EWN’s set is
similar to GermaNet’s [48]. Hyponymy is PWN’s troponymy. Cause includes PWN’s
1
http://plwordnet.pwr.edu.pl
presupposition [47]. EWN also introduced near-synonymy (for semantically close co-
hyponyms), near-antonymy, as well as cross-categorial synonymy, antonymy and
hypernymy.
There are 19 verb relations in plWN 3.0, 6 of them have derivational origin. Table 1
contrasts our selected verb relations with those of PWN, EWN and GermaNet. A list
of plWN 4.0 verb relations is presented in Table 2 and discussed in Sec. 3-4.
2 Constitutive Relations and Features
2.1 Synset Definition
The plWordNet model, cf [29; 30] is based on lexical units (LUs, i.e. triples: lemma,
part of speech, sense identifier) as basic building blocks. LUs are grouped into synsets
on the basis of sharing constitutive relations and features, i.e. two LUs belong to one
synset if and only if they share targets of the constitutive relations and values of the
constitutive features. Constitutive relations are selected lexico-semantic relations that
are enough frequent, express relatively high sharing factor and originate from the
tradition of lexicography and wordnets and are relatively well understood, e.g.
hyper/hyponymy, holo/meronymy, iterativity, distributivity. Constitutive features
constrain the shape of the system of lexico-semantic relations (e.g. aspect or adjective
and verb classes) or express some general pragmatic conditions on the use of lexical
meanings (e.g. stylistic register). Constitutive features are often referred to in the
definitions of the lexico-semantic relations, e.g. hyper/hyponymy is defined only for
adjectives of the same class or inter-register synonymy links LUs of non-compatible
lexical registers, cf [30].
Summing up this very brief plWN model description, see also [29], lexico-semantic
relations are the main means of description. They all can be traced back to the
language data. Synset relations are a kind of abbreviations representing sets of
constitutive relations held between members of two synsets. The synonymy relation is
to some extent derived from the constitutive relations and features that define synsets.
2.2 Aspect
Aspect is one of the fundamental grammatical verb categories and has gained a lot of
attention in the literature, cf. e.g. [4; 8; 47]. Comrie describes it as “the internal
temporal constituency of the situation” [4]. In Slavic languages it is marked
morphologically by affixes and - in some cases - by ablaut, cf. e.g. [3; 4; 19; 24; 35;
49]. According to the most widespread opinion among slavists, grammatical aspect is
a way in which a speaker envisages the situation he is speaking about. Thus the
imperfective aspect classifies the situation represented by a verb as ongoing, while the
perfective aspect informs that the represented situation is delimited in time and
described as complete in relation to its elements [24]. Every Polish verb may be either
perfective, or imperfective, if only we ignore ca 150 bi-aspectual verbs [31].2
Concerning lexical aspect (Aktionsarten) there are many verb pairs with opposition
2
The bi-aspectual verbs are ambiguous with respect to the morphological form,
and the actual aspect is determined by the context of use [41].
extended beyond imperfective/perfective by additional features of the represented
situation, e.g., the impf. verb robić (‘~to do, to be doing some activity’) gives birth
not only to its perf. counterpart zrobić, but also to many other derivatives formed with
non-empty prefixes, including the delimitative derivative po-robić ‘to have been
doing some activity for some time, and to have it finished’ or accumulative derivative
na-robić się ‘to work until being exhausted’. plWN covers the grammatical and
lexical aspects jointly in the aspectuality relation. Pure aspectual pairs differs only in
the grammatical aspect, while secondary aspectual pairs embody also the lexical
aspect [28].
In plWN 3.0 we adopted the theory of grammatical aspect that binds it with the
concepts of telicity/atelicity in such a way that only telic verbs were allowed to form
pure aspectuality pairs. Since telicity depends heavily on syntactic context [46], we
took into account and compared only primary uses of verbs [29], cf [24,35]. However,
the identification of the verb pairs was very laborious, because it required corpus-
based analysis of many examples. Moreover, it was difficult to maintain the
consistency of editors’ decisions. For instance, the pair gubićimpf. / zgubićperf. ‘to lose
something’ looks like an ordinary pure aspectual pair, but the close inspection of a
corpus reveals that gubić has the restriction that it rarely occurs with direct objects
expressed with countable NPs in singular (in cursive meaning), which is a clear signal
of atelicity [46; 8]. For many verbs such semantic-syntactic patterns are not so clear
with either telic or atelic contexts being dominant, e.g., the primarily atelic verb
jeśćimpf. ‘to eat’ (being a couple of zjeśćperf) occurs in many telic contexts (like jeść
jabłko ‘to eat the apple’). Thus, in the model proposed here, we have resigned from
the telicity condition and based the category on the analysis of definitions and
paraphrases. Thus secondary aspectuality was restricted mainly to those verb pairs
that are overtly linked derivationally by a non-empty affix, mostly a prefix.
Aspect is an intriguing problem for Slavic wordnet designers [13; 20; 21; 36; 37].
Some put pure aspectual pairs into one synset, e.g. BulNet [20] or CzechWordNet
[37], following the model of PWN (in English the similar distinction of progr./perf. is
inflectionally marked, and not derivational in nature). We treat the aspect in an
opposite manner as an important constitutive feature that shapes our verb relations
[29].
The rationale for this decision is following: if we took a closer look into Polish
dictionaries, we could notice that verb definitions (paraphrases) almost always include
superordinate LU (genus proximum) in the same aspect as the verb being defined
(definiendum). For example, the impf. robić ‘to do’ gets the impf. wykonywać ‘~ to
execute’ as a genus, while the perf. zrobićperf. ‘to do’ - the perf. wykonać ‘~ to
accomplish’. This is, of course, not a coincidence: you cannot freely substitute in a
context a verb of a given aspect with a superordinate verb in an opposite aspect. This
is due to the fact, that although a perf. verb implies its impf. counterpart, the reverse
implication is rarely true (what is in progress not necessarily would be completed
[42]).3 This suggests natural restriction of the hyper-/hyponymy relation only to LUs
of the same aspect. Consequently, if two LUs differ in their constitutive relations they
cannot belong to the same synset, see Sec. 2.1. Bi-aspectual verbs have mostly two
hypernyms: imperfective and perfective and the verb hypernymy tree forks, e.g.
3
The exception are parametric pairs, e.g., rosnąćimpf. - urosnąćperf. ‘to grow up’.
mianować `to nominate’ can mean both nadać [tytuł] `to grant [a title]’ but also
nadawać [tytuł] `to be grating [a title]’.
Pure aspectual pairs that we disunite into separate synsets are interlinked solely by
pure aspectuality relation. The meaning differences in the case of secondary
aspectuality are even more significant and results also in additional lexico-semantic
relations between the pair elements.
Aspect is also correlated with other verb relations, e.g. multiplicativity:iterativity is
observed only for imperfective verbs and tests for meronymy are constrained by
aspect, see Sec. 3.
2.3 Verb Classes
In plWN 3.0 verbs were organised into a hierarchical system of classes derived from
the proposal of Vendler [45] and Laskowski [24] for Polish. Classes and subclasses
were represented by the top level verb hypernym synsets. However, the classification
was fine-grained, multi-level and appeared to be difficult in consistent use. Moreover,
the representation of classes as top synsets made shaping the structure a challenging
task, e.g. all co-hyponyms had to belong to one single fine-grained class. For plWN
4.0 we decided to significantly simplify the classification and its representation.
While still being close to Vendler [45], we use only two top classes: state and
dynamic verbs.
State verbs are imperfective atelic and durative verbs that represent situations which
are stable and do not include any change during their time span. State verbs can
represent [24]: 1) localisation or position in space, (e.g. leżeć `to lie’, stać `to stand’,
znajdować się ‘~to be in some place’, otaczać ‘to surround’), 2) possession of
permanent material features like mass, temperature, state of matter (e.g. ważyć ‘to
weigh’, mierzyć ‘~to be of particular size’, świecić ‘~to produce light’), 3)
relationships between entities, both material and non-material, like to be a part of
something, to belong to someone/something, to be in a space (e.g. ograniczać ‘to
delimit’, równać się ‘to equal to’, składać się ‘to consist of’), 4) mental states,
emotional, sense experience (e.g. odczuwać, czuć both meaning ‘to feel’, wiedzieć ‘to
know’, doświadczać ‘to experience’, kochać ‘to love’, widzieć ‘to know’), 5) other
static situations (e.g. żyć ‘to live’, spać ‘to sleep’, spodziewać się dziecka ‘to be
pregnant, lit. to expect a child’).
State verbs are all imperfect and are linked with their perfective counterparts -
dynamic verbs - by the secondary aspectuality relation.
Dynamic verbs include all former auxiliary classes of plWN 3.0 [28]: 1) distributive
verbs (to do something by many agents or in relation to many objects, e.g.
powykańczać się `~to get exhausted, finished, plural sub., about many things’,
powyłamywać ‘~to be breaking off many things/parts’,), 2) accumulative (to do
something to such an extent that it is enough, e.g. najeść się ‘to be eating until an
agent is full’, spracować się ‘~to be working so long that it is enough’), 3)
perdurative (to be doing something during particular or limited time, e.g.
przepracować ‘to be working during some time’, przewegetować ‘~to be vegetating
during a particular time’), 4) delimitative (to be doing or to happening for some time
or to some extent, e.g. pobiegać ‘~to be running a little’, popadać ‘~to be raining a
little’).
All other dynamic verbs are grouped in 5) action verbs: a) all perfective and bi-
aspectual, b) imperfective derivatives of accumulative, delimitative, perdurative, and
distributive verbs (representing changing situations), c) imperfective derivatives of
semelfactive verbs4 (i.e. representing punctual or instantaneous events), e.g.
mrugaćimpf (from mrugnąćperf) ‘to flicker’, representing multiple changes, d)
imperfective causative verbs (expressing occurrence of a new situation), e) processive
(expressing gradual transition to a new state), f) inchoative (representing introductory
phase of a new situation) and g) limitative (representing ending of some situation),
and also g) all other imperfective verbs that represent situation changing due to
actions of entities involved or with respect to them, e.g. iść ‘to walkimp’, biegać `to
runimp’, płakać `to cryimp’.
The proposed system is much less fine-grained and more shallow - two main classes
and 5 subtypes - than the one of plWN 3.0 (9 main classes, 4 auxiliary classes). In
plWN 3.0 only some of these verb classes had direct influence on the system of verb
relations. Thus the former auxiliary classes became subclasses of the dynamic verbs.
Their direct influence on the relation structure was limited only to distributive verbs
related to the distributivity relation, see Sec. 4.3.
The main reasons for distinguishing state verbs were the cross-categorial synset
relation called state [28], see Sec. 3.4, imperfective aspect of state verbs and also the
secondary aspectuality which links state verbs with their dynamic counterparts.
3 Synset Relations
3.1 Hypernymy Hierarchical Structures
In PWN troponymy relation was used instead hyponymy for verbs [9], and described
as “to V1 is to V2 in some particular manner”. It is called also “a manner relation”
and links temporally co-extensive verbs. However, EWN has opted for verb
hyponymy which can be adapted to the manner relation [48]. GermaNet identified
troponymy with hyponymy [22]. Maziarz et al. [29] compared the ways of defining
nouns and verbs in dictionaries and argued for the use of hyper/hyponymy relations
for verbs in plWN 3.0. It links verbs of the same semantic class, aspect and
compatible register:
Derwojedowa et al. [5] argued that there is place for both relations: troponymy and
hyponymy in plWN due to the specific morphology of Polish verbs. Many verbs are
derived by prefixes from their derivational bases that seem to be their hypernyms in
the same time. Troponymy could be a relation signalling a derivational association
between a more specific and more general verb. However, starting from plWN 2.0
such association started to be described by secondary aspectuality relation.
Derwojedowa et al. [5] and Piasecki et al. [38] noticed the necessity of having two
verb relations expressing specification of the meaning as a result of the rich semantics
of verb prefixes in Polish. In GermaNet these two tendencies are combined [15; 16]
by putting on the upper levels of the verb hierarchy synsets that exemplify the
meaning of verb prefixes. In Polish, the semantics of the verb prefixes is more
4
Semelfactive verbs as perfective verbs are included into actions, see the point a).
complex and its more in depth description is planned for further research.
plWN 3.0 verb hierarchy resembles the one from PWN 3.1 where top levels are
populated by artificial synsets (non-lexicalised)5 that are similar to 40 “top-level
synsets” in PWN [11]. We plan to rebuild it in such a way that the upper levels of the
hierarchy will be also synsets grouping hyponymy branches, but constructed in a
bottom-up direction. The reconstruction work will be started from the most specific
verbs in the bottom levels and proceed to upper levels. The top synsets will result
from the language data and potential needs for more comprehensible organisation of
the tree in the case of artificial synsets. As in the case of nouns, we will not combine
all verb synsets into one single-rooted tree.
Semantic class and aspect are kept as constitutive features that determine the
hypernymic trees for two reasons. First, the simplified set of classes is easier in use
for wordnet editors and next the class assignment facilitates identification of
hypernyms. Second, state verbs are all imperfective, while their perfective derivates
are in fact secondary aspectuality pairs and express additional semantic difference.
Division into classes supports consistent organisation of the verb synsets.
3.2 Associations between Situations
Besides hyper/hyponymy, the proposed verb constitutive relations (i.e. synset
relations) can be divided into 3 groups. The first group includes lexico-semantic
relations expressing associations between situations that are related to the timeline to
some extent. It includes: presupposition and preceding (backward relations),
meronymy (co-occurrence of two situations), inchoativity (beginning of a situation),
causality (resulting in a situation) and processuality (a state resulting). They are only
slightly amended since plWN 3.0. If we observe a situation represented by a verb
from the perspective of a point t0 in the present time, then causality, processuality and
inchoativity refers to a point t1 in the future resulting somehow from the verb
situation, presupposition and preceding refer to a point t-1 in the past, as the verb
situation is influenced by what had happened around t-1, while meronymy refers to the
present time period and a co-occurring situation, cf [28].
In plWN 3.0 model all these relations had many subtypes related to properties of LUs
linked (e.g. aspect). According to our experience, definitions and substitution tests
have been simplified by parameterizing them by the feature values. WordnetLoom
[39] application has been extended in such a way that on the basis of the aspect
attribute of a LU, an appropriate version of a test is generated and presented to the
editor.
Table 1. Selected verb lexico-semantic relations in four wordnets
PWN GermaNet EWN plWN
synonymy (V-V) synonymy (V-V) synonymy (V-V) synonymy (V-V)
antonymy (V-V, N,
antonymy (V-V) antonymy (V-V) antonymy (V-V)
Adj, Adv)
troponymy (V-V) hyponymy (V-V) hyponymy (V-V) hyponymy (V-V)
5
Representing very fine-grained division into semantic classes.
entailment - proper
subevent (V-V) subevent (V-V) meronymy (V-V)
inclusion (V-V)
entailment - backward cause factial (V-V, N, presupposition (V-V,
entailment
presupposition (V-V) Adj) N, Adj, Adv)
cause non-factial (V- preceding (V-V, N,
entailment (V-V) entailment
V, N, Adj) Adj, Adv)
cause (V-V, N, causality (V-V, N, Adj,
cause (V-V) cause (V-V, N, Adj)
Adj) Adv)
troponymy (V-V) ____________ manner (V-Adv) manner (V-Adv)
selectional
______________ _______________ subject (V-N)
restrictions (V-N)
selectional
______________ _______________ object (V-N)
restrictions (V-N)
_______________ _______________ role inclusion (V-N) role inclusion (V-N)
Except meronymy (V-V) and inchoativity (V-V,N) all other relations can link verbs to
any other part of speech. The system of 6 situation association relations can seem to
be complex. In PWN, most of them are covered by entailment with cause as the only
exception [9]. A more fine grained description was proposed in EWN [48], see Table
1. However, some relations are signalled derivationally, more fine-grained types helps
to write more precise definitions and substitution tests and all of them are relatively
frequent, see statistics from plWN 3.0 in Tab. 2. Due to the above mentioned
problems with the model of plWN 3.0 we can expect that the presented numbers are
not complete.
Inchoativity [28] links: V-V,N, see Tab. 2, where the first verb represents an initial
phase of a situation represented by the second element [24]. Inchoativity is signalled
derivationally, i.e. the second verb LU is a derivational basis for the first, e.g.
rozpłakać się `~to start crying’ -inch.→płakać `to be crying’, but due to its semantic
regularity was expanded to LU pairs not linked derivationally, e.g. {aktywować się 1
`to activate itself’} -inch.→ {funkcjonować 1 `to functionimp’, działać 3 `to workimp’,
chodzić 4 `~to workimp’, ... , pracować 3 `to workimp’, ...}, in all these cases
aktywować się 1 means `to start X-ing’. In pairs with nouns, the noun names a
situation whose beginning is represented by the verb, e.g. rozpłakać się `~to start
crying’ -inch.→ płacznoun `cry’.
Processuality [28] can be described as ‘to become or to be becoming’. It is often
signalled by a derivational association, but it was also expanded to a synset relation,
e.g. czerwienieć 1= `to be becoming’ czerwony 1 `red’ or zmieniać się 1 `to be
changing itself/yourself’ = to be becoming ‘inny 1 `different’.
Causality is used in PWN [33] to link only verbs, in EWN is cross-categorial. plWN
causality is defined for dynamic verbs (as in EWN), and the effect, intentional or
unintentional, can be represented by a dynamic or static verb, as well as any other PoS
[28]. For instance ogrzać 1 `to warm up’ means: 1) ‘to cause smth’ zagrzać się 1 `to
have become warm by itself’, 2) ‘to cause heat [ciepło 1]’, 3) ‘to cause that smth. is
warm [ciepły 1]’, 4) ‘to cause that it is warm’ [spowodować, że jest ciepło 1]’.
EWN cause relation encomapasses also cause non-factial relations. In plWN this type
of relationship is expressed by preceding relation and the stricter cases of necessity by
presupposition [28]. Presupposition is close to logical presupposition and informs
about the necessity of earlier occurrence of some situation, e.g. żyć 1 `to live’
←pres.- umrzeć 1 `to die’, because if someone has died, he had to live earlier, the
same if he has not died; thus umrzeć 1 presupposes żyć 1.
Preceding relation represents a possibility that one situation happens before the other
one, e.g. siedzieć `to sit’, stać `to stand’ ←prec- położyć się `to have laid down’, if
someone has laid down, he could earlier sit or stand, but both precede the former.
Because during construction of plWN 3.0 we could observe problems with consistent
selection between causality and presupposition/preceding, we propose here to
introduce two subtypes of both: identical (e.g. martwy `dead’←pres- zmartwychwstać
`to resurrect’; żona `a wife’, mąż `a husband’ ←prec.- rozwieść się `to divorce’) and
non-identical subjects (e.g. zabić `to kill’←pres.- żyć `to live’; minimum `a
minimum’, maksimum `a maximum’←prec.- przekroczyć `to overcome’). The latter
subtype of presupposition and preceding resembles causality, but in opposite
direction, and we assume that both can be use in parallel but in opposite directions.
Table 2.Verb relations in plWordNet 4.0.
Relation POSs Example No(3.0
inter-register pieprzyć [vulgar] ‘~to speak nonsense’ →
V-V 2016
synonymy mówić ‘to speak’
aspectuality
napisać ‘to writeperf’ - pisać ‘to writeimperf’
- pure V-V 26558
popisać ‘~to writeperf a little’ - pisać ‘to
- secondary
writeimperf’
płynąć ‘to swim’ → przemieszczać się ‘to
hyponymy V-V 22680
move’
rozpłakać się ‘to start crying’ → płakać `to
inchoativity V-V, N 403
cry’
owdowieć ‘to become a widow or widower’
processuality V-N, Adj, Adv 729
→ wdowa ‘a widow’, wdowiec ‘a widower’
suszyć ‘to cause that smth. is drying’ →
causality V-V, N, Adj, Adv 1773
schnąć ‘to dry’
presupposition V-V, N, Adj, Adv umrzeć ‘to die’ presupposes żyć ‘to live’ 167
siedzieć ‘to sit’ precedes wstać ‘to stand up’
preceding V-V, N, Adj, Adv 169
(but wstać does not presuppose siedzieć)
przełykać ‘to swallow’ is an integral part of
meronymy V-V 2829
situation jeść ‘to eat’
jeść ‘to eat’ is a typical situation including
holonymy V-V 2888
przełykać ‘to swallow’
multiplicativity
grywać ‘~to play a little from time to time’
- iterativity V-V 542
→ grać ‘to play’
- distributivity
pokraść ‘~to steal much’ → ukraść ‘to steal’
state V-V, N, Adj, Adv czekać ‘to wait’ → gotowy ‘ready’ 93
antonymy
oddać ‘to give back’ - zatrzymać ‘to keep’
-
V-V zwiększać ‘to increase’ - zmniejszać ‘to 2348
complementary
reduce’
- proper
converseness V-V sprzedać ‘to sell’ - kupić ‘to buy’ 112
role inclusion
V-N 1357
- subject ocieniać ‘to shade’ ← cień ‘a shade’
- instrument solić ‘to salt’ ← sól ‘salt’
- result portretować ‘to make portrait’ ← portret ‘a
- location portrait’
- object garażować ‘to keep in garage’ ← garaż ‘a
- time garage’
- indefinite bębnić ‘play drums’ ← bęben ‘drum’
nocować ‘to spend a night’ ← noc ‘a night’
akumulować ‘to accumulate’ ← akumulacja
‘an accumulation’
subject V-N rżeć ‘to neigh’ → koniowate ‘equine’ 0
object V-N wzuwać ‘to put on shoes’ → but ‘a shoe’ 0
manner V-Adv podsmażyć ‘to fry a little’ → trochę ‘a little’ 0
circumstance V-N dobijać ‘to reach a shore’ → brzeg ‘a shore’ 0
uwznioślić ‘to get exalted’→ wzniosły
derivationality V-V, N, Adj, Adv 281
‘exalted’
Meronymy (holonymy) was defined in PWN and EWN only for nouns. Because
plWN verb hyponymy does not cover cases in which a situation is an element of a
larger, more general situation, verb meronymy was introduced to express such
associations [28] with two subtypes of sub-situation and accompanying situation. As
this distinction was not very clear in many cases, for plWN 4.0, we have removed
both subtypes and put emphasis on the necessary simultaneous co-occurrence of two
situations. A verb meronym represents a more narrow situation, e.g. chrapać 2 `to
snore’ -mero→ spać 1 `to sleep’. Holonymy is a reverse relation, but is not
automatically defined.
3.3 Multiplicativity
Multiplicativity relation describes repetition of some state or activity, and performing
an activity by many subjects or on many objects. The first subtype is called iterativity,
and the second distributivity, cf [28; 49]. Iterativity links imperfective verbs, while
distributivity perfective. Both are expanded also to verbs pairs that come from the
same root, but which are not direct derivates. Multiplicativity relation is more specific
than hyponymy and excludes it. Multiplicative verbs are linked with their derivational
bases of different aspects by the lexical relation (non-synset) of secondary
aspectuality (e.g. nakupować 1 `to buyperf many things’ -sec. asp.- kupować 1 `to
buyimp’).
3.4 Cross-categorial relations
Constitutive relations are the primary means of defining synsets and we are still
looking for possibilities to increase the number of their instances and to increase the
network’s density. Moreover, cross-categorial relations are valuable for many
wordnet applications, including WSD. State relation, introduced in plWN 2.0, in
plWN 4.0 is restricted to the state verbs. It represents being in some state, e.g.
ograniczać 3 `to delimitimp’ means ‘to be a border [granica 1]’, czerwienić się 1 `to be
red’ -state→ czerwony 1 `red’, czuwać `to stayimp awake’-state→ świadomie
`consciously’.
We propose four new cross-categorial relations in plWN 4.0. Manner relation
describes a verb by a link to an adverb describing a manner in which the activity is
performed:
Jeżeli ktoś lub coś X-wał(o), to znaczy, że Z-wał(o) Y-owo. (gdzie X jest hiponimem
Z).
`If smn/smth has X, it means that he/it has Z Y[ADV]’, where X is a hyponym of Z’
e.g. podgotować ‘to cookperf a little’ → trochę [Adv] ‘little [Adv]’;
Circumstance describes a verb by referring to an adverbial realised by a simple
prepositional phrase and links the verb to the noun which is the semantic head of the
prepositional phrase:
Jeżeli ktoś lub coś X-wał(o), to znaczy, że Z-wał(o) Y[Prep]. (gdzie X jest hipo. Z.)
`If smn/smth has X, it means that has Z Y[Prep]’
e.g., dopłynąć ‘~to swimperf to some point/place’ -circum.→ brzeg `a bank’, where the
phrase do brzegu `to the bank’ is an unexpressed element.
The prepositional phrase is an unexpressed element of the relation.
Subject and object are conceptually generalised from the morpho-semantic role
relation (see Sec. 5.2) It was motivated by functional grammar associating semantics
with syntax [18], e.g. subject and object relations in [6]. They characterise semantic
roles of the semantic subject and object [43]. Inclusion of predicate-argument roles as
definition elements can result in their proliferation and subjectivity in adding them.
That is why, we limited the number of instances per verb to 3. However, we aim at
adding only very characteristic links, so smaller number, better. Such links must
originate from meaning definitions.
Subject: Jeżeli mowa o tym, że ktoś lub coś X-ował(o), to ten ktoś lub to coś pewnie
był(o) Y-kiem. `If we say that smn/smth X, then this smn/smth is certainly Y’
e.g. muczeć ‘to moo’→ krowa ‘a cow’,
Object: Jeżeli mowa o tym, że X-owano kogoś lub coś, to ten ktoś lub to coś pewnie
był(o) Y-kiem. `If we say that smn/smth was X, then this smn/smth had to be certainly
Y’ e.g. wzuwać `~to put shoes on’→ but `a shoe’.
4 Lexical Relations
Relations that do not fulfill requirements for constitutive relations are described on
the level of LUs, not synsets. This type includes first of all semantic oppositions (like
in PWN and most other wordnets), but also a large number of derivationally
motivated relations (that is common for Slavic wordnets), called also morpho-
semantic relations.
4.1 Oppositions
Opposition relations have not changed since plWN 3.0, cf [28; 38]: pure antonymy,
complementary antonymy, and converseness. All are restricted to verbs.
Converseness, considered in PWN, but not included, cf [9] links verbs representing
the same situation, but from different viewpoints and with the reverse positions of
arguments in the semantic valency frame, e.g. sprzedać 1 `to sell’ - kupić 1 `to buy’.
4.2 Morpho-semantic relations
Derivational relations mostly express some lexical meaning change, but derivational
morphemes are often very ambiguous, cf [40]. The idea of morpho-semantic relations
was proposed, e.g. [12] to overtly describe these semantic links, and used in wordnets,
e.g. [48]. Morpho-semantic relations are especially numerous in Slavic wordnets, e.g.
[20]. In EWN relations of this type does not need to be strictly derivationally
motivated that is similar to our expansion of some morpho-semantic relations,
namely: inchoativity, multiplicativity, processuality, and state to synset relations, see
Sec. 4. In addition a lexico-semantic relation of role cf. [14] was introduced in plWN
2.0 [28], and has been preserved unchanged. Role is described only if a noun is
derived from a verb (role inclusion describes derivation of a verb from a noun), and
has 7 subtypes (see Tab. 2). In plWN 4.0 role will be registered only for imperfective
verbs, as most perfective verbs are derived from imperfective. Such a solution should
help to improve consistency in adding role instances and limits its artificial
proliferation.
5 Potential Applications and Conclusions
The modified and extended system of verb relations will lead to a denser network and
richer description of verbs. Cross-categorial links are especially important for WSD in
relation to all PoSs. WSD methods based on activation spreading across the wordnet
network express increasing performance with the increasing number of words from
the given text context that are connected in the network.
Relations describing associations between situations in Sec. 4.2 form a kind of a
system. On the basis of our former experience, we are convinced that they can be
defined with good consistency and they are relatively frequent. Moreover, the current
description, see Tab. 2, is definitely not complete. So, such relations should facilitate
calculating graph-based text similarity, recognition of textual entailment, text
semantic relations or paraphrases, as well as even some forms of reasoning.
The former, plWN 3.0, verb model was complex, and some problems have
appeared in its application in wordnet development. Some elements of the model (e.g.
classes and relations) were not interpreted in a way consistent enough, many relation
subtypes have not been described with good coverage. Our applications of plWN (e.g.
WSD, text similarity) revealed a need for further increase of the verb network density,
especially in relation to cross-categorial links, e.g. there are significant differences in
WSD quality depending on the density of the local wordnet subnetworks around
different verbs.
The proposed modifications in the plWN verb model are more an evolution than
revolution. In most cases we try to simplify the model, but in a careful way, from the
perspective of the wordnet development process. We try to firmly base new proposed
relations on the analysis of the language data. References to the intuition of linguists
are inevitable, but not the core of the definitions. We try to perceive plWN as a multi-
purpose, semi-formal description of the Polish lexical semantics systems, but
necessarily built on a large scale and focused on range of applications in NLP.
Acknowledgments
This work was co-financed as a part of the investment in the CLARIN-PL research
infrastructure (www.clarin-pl.eu) funded by the Polish Ministry of Science and
Higher Education.
References
1. Alonge, A.: Definition of the links and subsets for verbs, EuroWordNet Project LE4003,
Deliverable D006. Pisa (1996), https://www.illc.uva.nl/EuroWordNet/docs/D006.ai.
2. Apresjan, J.: Semantyka leksykalna: synonimiczne środki języka. Zakład Narodowy im.
Ossolińskich, Wrocław (2000).
3. Cockiewicz, W. (2007). Na peryferiach aspektu. LingVaria 2(4), 9-25 (2007).
4. Comrie, B.: Aspect: An introduction to the study of verbal aspect and related problems
(Vol. 2). Cambridge university press, Cambridge (1976).
5. Derwojedowa, M., Zawisławska, M.: Relacje leksykalne w polskiej i czeskiej bazie
WordNet. In: Rudnik-Karwatowa, Z. (eds.), Z polskich studiów slawistycznych, seria XI,
Językoznawstwo, Warszawa (2007).
6. Dik, S. C.: Functional grammar (Vol. 7). Foris, Dordrecht (1981).
7. Dimitrova, T., Tarpomanova, E., and Rizov, B. Coping with derivation in the Bulgarian
WordNet. In Proceedings of the Seventh Global Wordnet Conference (GWC 2014), pages
109–117, (2014).
8. Dowty, D.: The Effects of Aspectual Class on the Temporal Structure of Discourse:
Semantics or Pragmatics? Linguistics and Philosophy 9(1), 37-61 (1986).
9. Fellbaum, C.: A semantic network of English verbs. In: Fellbaum (eds.), WordNet: An
electronic lexical database, MIT Press, Cambridge (1998).
10. Fellbaum, C.: WordNet. Blackwell Publishing Ltd., Hoboken (1998a).
11. Fellbaum, C.:WordNet. In: Poli, R., Healy, M., Kameas, A. (eds.), Theory and
applications
of ontology: Computer applications. Springer, Heidelberg (2010).
12. Fillmore, C. J. (1967). The case for case. In Bach, E. and Harms, R.T. (eds.). Universals in
Linguistic Theory, pp. 1-91, New York (1967).
13. Fišer, D.: A multilingual approach to building slovene wordnet. In Proceedings of the
workshop on A Common Natural Language Processing Paradigm for Balkan Languages
held within the Recent Advances in Natural Language Processing Conference RANLP
(Vol. 7, p. 26) (2007).
14. Grzegorczykowa, R., Puzynina, J.: Rzeczownik. In: Grzegorczykowa, R., Wróbel H.,
Laskowski R. (eds.), Gramatyka współczesnego jezyka polskiego. Morfologia 2,
Warszawa (1998).
15. Hamp, B., Feldweg, H.: Germanet-a lexical-semantic net for german. In: Vosen, P.,
Adriaens, G., Calzolari, N., Sanfilippo, A., Wilks, Y., Proceedings of ACL workshop
Automatic Information Extraction and Building of Lexical Semantic Resources for NLP
Applications, pp. 9-15, Madrid (1997).
16. Hoppermann, C., Hinrichs, E.: Modeling prefix and particle verbs in GermaNet.
www.aclweb.org/anthology/W14-0107 (2017/04/09).
17. Karolak, S.: Arguments sémantiques contre la distinction: aspect / modalité d’action, In:
Studi Italiani di Linguistica Teorica e Applicata, anno XXII, 1993, No 2, pp. 255-284
(2003).
18. Kiklewicz, A.: Podstawy składni funkcjonalnej. Wydawnictwo UWM. Olsztyn (2004).
19. Kiklewicz, A.: Finitywny (teleologiczny) model aspektualności: założenia teoretyczne.
Prace Filologiczne L, pp. 59-82 (2005).
20. Koeva, S.: Derivational and morphosemantic relations in Bulgarian wordnet. Intelligent
Information Systems 16, 359-369 (2008).
21. Koeva S., Krstev C., Vitas D.: Morpho-semantic Relations in Wordnet – a Case Study for
two Slavic Languages. In: Tanács, A., Csendes, D., Vincze, V., Fellbaum, Ch., Vossen, P.
(eds.) Global Wordnet Conference (GWC'08). Global Wordnet Conference, 2008, pp. 239-
253, Szeged, (2008).
22. Kunze, C.: Semantics of verbs within GermaNet and EuroWordNet. In: Proceedings of
11th European Summer School in Logic, Language and Information, pp. 189-200, Utrecht
(1999).
23. Kunze, C., Lemnitzer, L.: GermaNet – representation, visualization, application. In:
Proceedings of LREC 2002, main conference, volume V, pp. 1485–1491, Las Palmas
(2002).
24. Laskowski, R.:. Kategorie morfologiczne języka polskiego–charakterystyka funkcjonalna.
In: Grzegorczykowa, R., Wróbel H., Laskowski R. (eds.), Gramatyka współczesnego
języka polskiego. Morfologia 1, Warszawa (1998).
25. Lis, M.: Annotation scheme for multimodal communication: Employing plWordNet 1.5.
In: Proceedings of the Formal and Computational Approaches to Multimodal
Communication Workshop. 24th European Summer School in Logic, Language and
Information (ESSLLI 2012), (2012).
26. Lyons, J.:Semantics. CUP, Cambridge (1977).
27. Marcińczuk, M., Oleksy, M., Bernaś, T., Kocoń, J., Wolski, M.: Towards an event
annotated corpus of Polish. Cognitive Studies| Études cognitives 15, 253-267 (2015).
28. Maziarz, M., Piasecki, M., Szpakowicz, S., Rabiega-Wiśniewska, J., Hojka, B.: Semantic
Relations between Verbs in Polish Wordnet 2.0. Études Cognitives/Studia Kognitywne 11,
183-200, (2011).
29. Maziarz, M., Piasecki, M., Szpakowicz, S.: The chicken-and-egg problem in wordnet
design: synonymy, synsets and constitutive relations. Language Resources and Evaluation
47, 769-796, (2013).
30. Maziarz, M., Piasecki, M., Rudnicka, E., Szpakowicz, S., Kędzia, P.: (2016) plWordNet
3.0 - a Comprehensive Lexical-Semantic Resource. In: Calzolari, N., Matsumoto, Y.,
Prasad, R. (eds.), COLING 2016, 26th International Conference on Computational
Linguistics, Proceedings of the Conference: Technical Papers, December 11-16, pp. 2259-
2268, Osaka (2016).
31. Mędak, S.: Słownik form koniugacyjnych czasowników polskich. Universitas, Kraków
(1997).
32. Młynarczyk, A.: Aspectual Pairing in Polish. LOT, Utrecht (2004).
33. Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D., Miller, K. J.: Introduction to
WordNet: An on-line lexical database. International journal of lexicography 3(4), 235-244
(1990).
34. Miller, G., Fellbaum, Ch. (eds.): WordNet – An Electronic Lexical Database. The MIT
Press, Cambridge (1998).
35. Paducheva, E. V.: Taxonomic categories and semantics of aspectual opposition. Temporal
reference, aspect and actionality 1, 71-89 (1995).
36. Pala, K., Smrž, P.: Building czech wordnet. Romanian Journal of Information Science and
Technology 7(2-3), 79-88 (2004).
37. Pala K., Hlaváčková D.: Derivational relations in Czech WordNet. In: Proceedings of the
Workshop on Balto-Slavonic Natural Language Processing: Information Extraction and
Enabling Technologies (ACL '07), pp. 75-81, Association for Computational Linguistics,
Stroudsburg, PA (2007).
38. Piasecki, M., Broda, B., Szpakowicz, S.: A wordnet from the ground up. Oficyna
Wydawnicza Politechniki Wrocławskiej, Wrocław (2009).
39. Piasecki, M., Marcińczuk, M., Ramocki, R., Maziarz, M.: WordNetLoom: a WordNet
development system integrating form-based and graph-based perspectives. International
Journal of Data Mining, Modelling and Management 5(3), 210-232 (2013).
40. Piasecki, M.; Ramocki, R. & Maziarz, M.: Recognition of Polish Derivational Relations
Based on Supervised Learning Scheme In Calzolari, N.; Choukri, K.; Declerck, T.; Dogan,
M. U.; Maegaard, B.; Mariani, J.; Odijk, J. & Piperidis, S. (Eds.) Proceedings of the Eight
International Conference on Language Resources and Evaluation (LREC'12), European
Language Resources Association (ELRA), pp. 916-922 (2012).
41. Perlin, J.: Ile jest we współczesnej polszczyźnie czasowników dwuaspektowych?.
Linguistica Copernicana, 3(1), 165-172 (2010).
42. Rozwadowska, B.: Initial Boundary and Telicity in the Semantics of Perfectivity. In:
Kosta, P., Błaszczak, J., Frasek, J., Geist, L., Żygis, M. (eds.) Investigations into Formal
Slavic Linguistics. Berlin: Peter Lang, pp. 859-872 (2003).
43. Rudyk, A.: O subiekcie semantycznym w rosyjskich zdaniach z predykatywem
wartościującym pod względem emocjonalnym w porównaniu z językiem polskim. Acta
Universitatis Lodziensis. Folia Linguistica Rossica. (2011).
44. Šojat, K., and Srebačic, M.: Morphosemantic relations between verbs in Croatian Word-
Net. In Proceedings of the Seventh Global WordNet Conference, pp. 262–267, (2014).
45. Vendler, Z.: Verbs and times. In: Vendler, Z. (eds.) Linguistics in philosophy, Ithaca,
Cornell University Press, pp. 97-121, New York (1957).
46. Verkuyl, H.J. (2005): Aspectual Composition: Surveying the Ingredients. In: Verkuyl,
H.J., de Swart, H., van Hout, A. (eds.), Perspectives on Aspect [Studies in Theoretical
Psycholinguistics, v. 35], Springer, pp 19-39, (2005).
47. Vossen, P., Bloksma, L., Alonge, A., Marinai, E., Peters, C., Castellon, I., Rigau, G.:
Compatibility in interpretation of relations in EuroWordNet. In: Vossen, P. (eds.),
EuroWordNet: A multilingual database with lexical semantic networks, pp. 81-112,
Springer Netherlands, Dordrecht (1998).
48. Vossen, P.: EuroWordNet. General document. University of Amsterdam, Amsterdam
(2002).
49. Wróbel, H.: Czasownik. In: Grzegorczykowa, R., Wróbel, H., Laskowski, R. (eds.),
Gramatyka współczesnego języka polskiego. Morfologia 2, Warszawa (1998).