=Paper=
{{Paper
|id=Vol-2518/paper-CREOL3
|storemode=property
|title=Sweetening Ontologies Cont’d: Aligning
bottom-up with top-down ontologies
|pdfUrl=https://ceur-ws.org/Vol-2518/paper-CREOL3.pdf
|volume=Vol-2518
|authors=Elisabetta Jezek
|dblpUrl=https://dblp.org/rec/conf/jowo/Jezek19
}}
==Sweetening Ontologies Cont’d: Aligning
bottom-up with top-down ontologies==
https://ceur-ws.org/Vol-2518/paper-CREOL3.pdf
Sweetening Ontologies Cont’d: Aligning
bottom-up with top-down ontologies
Elisabetta JEZEK a
a
University of Pavia, Department of Humanities
Section of Theoretical and Applied Linguistics
jezek@unipv.it
Abstract. This paper addresses an issue at the interface between language and
ontology. Specifically, we report the results of the alignment we performed between
the T-PAS ([1]) and DOLCE categories ([2]), and discuss the distinctions and
similarities we observed from a cognitive and application-based perspective. The
motivation for our work lies in the different nature of the two resources; while T-
PAS is a bottom-up system, in which semantic types are identified by manual
clustering the fillers of argument positions of verbs gathered from large corpora,
DOLCE is top-down ontology, in which categories are not based on extensive
linguistic evidence and are stipulated on formal grounds. The preliminary results of
the alignment task reveal that the most general types in T-PAS can be mapped fairly
well into DOLCE’s upper level. Two substantial issues remain open, the mapping
of the Abstract category and the treatment of systematic polysemy. The experiment
also assesses the anthropic character of the bottom-up system compared to the top-
down system, and the fine-grained granularity of the first compared to the second.
On the other hand, the taxonomy of DOLCE is ontologically more solid than the T-
PAS hierarchy. The resulting alignment benefits both sides.
Keywords. semantic types, ontological category, endurant, corpus analysis,
anthropic concepts
1. Introduction
By applying the methodology of Corpus Pattern Analysis ([3]) to the analysis of corpus
evidence for about 1200 Italian verbs, with the goal of acquiring their recurrent semantic
structures (e.g. HUMAN partecipa in ACTIVITY), we have compiled a list of 180 semantic
types to characterize the semantic preferences of verbs for each argument position in
each verb sense.1 These semantic types (EVENT, LOCATION, FOOD, VEHICLE, PART OF
BODY, etc.) are obtained from manual clustering of lexical items found in the argument
positions of verbal structures in the corpus:2 they can thus be seen as human judgments
abut the selectional preference of verbs as shown in the corpus. These type look very
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
1
The analysis is carried out within the T-PAS (Typed Predicate-Argument Structures) project, which
aims at building an inventory of semantic structures for Italian verbs for linguistic analysis and text processing
([1]).
2
For example, for ACTIVITY in the structure above we have: gara, riunione, selezione, manifestazione,
seduta, cerimonia, conferenza, votazione, elezione, celebrazione, esequia, competizione, maratona,
discussione, messa, festa, and so forth.
much like ontological categories; however, instead of being stipulated, they are induced
by the analysis of selectional properties of verbs, and reflect the way in which humans
talk about events and states of entities through language. Despite the obvious correlations,
the methodology underlying the identification of types in T-PAS differs from the way
categories are defined in resources such as the DOLCE ontology (Descriptive Ontology
for Linguistic and Cognitive Engineering), which, despite “aiming at capturing the
ontological categories underlying natural language and human common sense’ ([4]) does
not derive the categories from systematic observation and clustering of linguistic data.
In this paper, we report the results of an experiment of alignment of the semantic
type system developed in T-PAS to the DOLCE category system. The goals of this
exercise are to verify the possibility of the alignment, and to identify the similarities and
differences between a corpus-based and a speculative system of categories, highlighting
the distinctions and similarities between the two from a cognitive and application-based
perspective. It is expected that both resources can benefit from this analysis, and
consequently be improved. We regards such issues of significant importance for applied
ontology.
The structure of the contribution is as follows. Section 2 provides a description of
the T-PAS semantic type system focusing on the aspects that are relevant for the present
discussion; section 3 presents an overview of the content and structure of DOLCE;
section 4 reports the results of the alignment between the two resources and the analysis
of the results. Finally, section 5 reports the concluding observations together with
directions for further research.
2. System of Semantic Types in T-PAS
The system of semantic types in T-PAS currently contains 180 semantic types (as of
August 2019). The list is organized in a hierarchy in order to identify the appropriate
level of specificity of the selectional properties of individual verbs. The taxonomic
structure is based on human judgments, and on the manual comparison of the members
of the set of lexical items that instantiate each type in the corpus. The only relation
represented in the system is subsumption. The part_of relation is currently not
represented. This has consequence as to how types such as PART OF BODY are currently
positioned in the structure. Fig. 1 reports the top level of the system.
Figure 1. Top-level of the T-PAS system with a selection of leaf types
As shown in Fig. 1, the starting point of the hierarchy is the type ANYTHING. The top
level has ENTITY, EVENTUALITY (in Emmond Bach's [5] terminology) and PROPERTY as
branches. The main distinction in the domain of the ENTITY is between PHYSICAL and
ABSTRACT ENTITY. PHYSICAL ENTITY is further distinguished in INANIMATE, ANIMATE,
BODY, PART OF BODY, PLANT and LOCATION. BODY, PART OF BODY and PLANT are
considered ambiguous with respect to animacy, and therefore classified as subtypes of
PHYSICAL ENTITY. ARTIFACT forms a large and articulated branch of INANIMATE (34
nodes in total, not visible in the Figure), together with the sister note STUFF (17 nodes).
The system contains no type for NATURAL KIND (as opposed to ARTIFACT) nor a type for
INDIVIDUATED ENTITY (as opposed to STUFF). The prevailing distinction in the domain
of PHYSICAL ENTITY is between ANIMATE and INANIMATE; this finds motivation in the
role that this distinction plays in language, in particular in defining the semantic
preferences that verbs impose on their arguments. The domain of EVENTUALITY has
EVENT and STATE as main braches, whereas PROPERTY has, inter alia, COLOR, ROLE, and
WEIGHT as subtypes. The system includes multiple inheritance, as with DOCUMENT,
which is typed as an ARTIFACT and as an INFORMATIONAL SOURCE. For our current
purposes, we do not discuss the domains of EVENTUALITY and PROPERTY, and focus our
attention on PHYSICAL ENTITY.
3. DOLCE
Fig. 2 below depicts the structure of DOLCE, with focus on the Endurant branch.
Figure 2. Taxonomy of DOLCE’s categories (fragment)
The upper level of DOLCE distinguishes between Endurant, Perdurant, Quality and
Abstract. While Endurants are entities that exist over time (objects), Perdurants are
entities that happen over time (events and states). The main relation between the two is
that of "PARTICIPATION": an Endurant exists at a given time by participating in a
Perdurant. Qualities are inherent in the entities we can perceive or measure: shapes,
colors, sizes, sounds, smells, and so on. They manifest themselves with entities and exist
as long as the entity exists. Abstracts are entities without spatial or temporal qualities.
Within Endurants, DOLCE distinguishes between Physical and NonPhysical, depending
on whether or not they have direct spatial qualities. Child nodes of Physical are
AmountOfMatter, PhysicalObject and Feature, distinguished on the basis of Unity and
the relation of dependency (cf. [6, 7]). In particular, PhysicalObjects are Endurants with
Unity, whereas AmountsOfMatters are Endurants without Unity. PhysicalObject and
AmountsOfMatter do not depend on other objects, while Feature depends on another
object, its “host”. Examples of Features are RelevantPart, such as a bump in an object,
and Place, such as a hole in a piece of cheese. PhysicalObjects are analyzed in Agentive
and NonAgentive depending on whether or not they may have intentions. In DOLCE the
AgentivePhysical are made of Non-agentivePhysical: for example, a person is made up
of an organism ("CONSTITUTION" relation). NonPhysicalObjects (“abstract” in
common parlance) are distinguished in MentalObjects and SocialObjects depending on
whether or not they are generically dependent on a community of agents. SocialObjects
are further divided into Agentive and NonAgentive. AgentiveSocialObjects are for
example companies, like Sony. NonAgentiveSocialObjects are laws, norms, peace
treaties, etc.; they are generically dependent on Societies.
4. Mapping T-PAS into DOLCE
After presenting the content and structure of T-PAS and DOLCE, in this section we
report the results of the tentative alignment of the system of semantic types in T-PAS (in
red in Fig. 3) to the categories and structure of DOLCE (in blue, accompanied by the
definitions). As previously clarified, the attention is focused on the Endurant branch, in
particular PhysicalEndurant. The alignment is depicted in Fig. 3, which is followed by
comments on the individual operations.
■ Endurant!live%in%time%(and%can%therefore%exhibit% ■ Alcoholic!Drink!
changes)%by!participating%in%a%Perdurant!">!ENTITY!! ■ Wine!
■ Physical/Endurant!have%direct%spatial%qualities! ■ Water![Beverage,!Liquid]!
■ Amount/of/Matter!Endurants%with/no% ■ Food!
unity,%none%of%them%is%an%essential%whole,% ■ Building![Artifact,!Location]!
change%identity%when%they%change%parts% ■ Garment!
(mereologically%invariant)!">!STUFF! ■ Artwork!!
■ SOLID! ■ Movie![Artwork,!
■ MATERIAL!! Performance]!includes%video%
■ CLOTH! ■!Musical!Composition![Concept,!
■ THREAD! Artwork]!
■ METAL! ■ Picture!
■ DUST! ■ Document![Artifact,!Information!source]!
■ SOIL! ■ Agreement![Document,!Speech!Act]!
■ FLUID! ■ Machine!
■ LIQUID! ■ Vehicle!
■ BEVERAGE![ARTIFACT,!LIQUID]! ■ Road!Vehicle!
■ ALCOHOLIC!DRINK! ■ Water!Vehicle!
■ WINE! ■ Flying!Vehicle!
■ WATER![BEVERAGE,!LIQUID]! ■ Computer!
■ WATER! ■ Device!
■ VAPOUR!! ■ Software!
■ GAS! ■ Container!
■ AIR! ■ Engine!
■ SMELL! ■ Flag!
■ Physical/Object!Endurants%with%unity,% ■ Furniture!
mereologically%variant,%non%dependent%on% ■ Image!
■ Medium![Artifact,!Abstract],!e.g.!
other%objects!
radio,!TV,!the!Press!
■ Agentive!Endurants%with%intentions,!
■ Sound!Maker!e.g.!alarm!clock,!bell
constituted%by%non?Agentive%Physical%
■ Musical!Instrument!
Objects%(spatially%co?localized%with%them)%">!
■ String!
Animate!
■ Ball!
■ Human!!
■ Drug!
■ Human!Group!
■ Body!
■ Institution![Human!Group,!Abstract!
■ Parts!of!the!Body!
Entity]!
■ Plant!
■ Business!Enterprise!
■ Location!!
■ Animal!
■ Natural!Landscape!Feature!
■ Cat!
■ Watercourse!includes%lakes,%the%
■ Cow!
sea,%rivers%and%streams!
■ Horse!
■ Waterway![Watercourse,!Route]!
■ Dog!
e.g%canals,%navigable%rivers!
■ Sheep!
■ Hill!
■ Goat!
■ Route!includes!roads,%railways!
■ Snake!
■ Waterway![Route,!Watercourse]!
■ Spider!!
e.g%canals,%navigable%rivers!
■ Bird!
■ Area!includes%geographical%area,%e.g.%
■ Insect!
states!
■ Fish!
■ Building![Location,!Artifact] !
■ Animal!Group! ■ Light!Source![Location,!Inanimate]!
■ Non Inanimate! dependent%on%physical%objects%8!their%hosts%
■ Artifact! (not%spatially%co?localized%with%them)!
■ Weapon!
■ Relevant!Part!e.g.!bump,%damage!
■ Bomb!
■ Place!e.g.%crack,%hole,%opening,%
■ Firearm!
window,%doorway!
■ Beverage![Artifact,!Liquid]!
■ Aperture!
Figure 3. Mapping T-PAS into DOLCE. DOLCE categories are marked in black and italic,
together with their definitions, whereas TPAS semantic types are in grey and small caps.
Endurant vs. Entity
The first high-level observation is that DOLCE's Endurant category is a node that aligns
very well with the T-PAS organization. DOLCE Endurant corresponds to ENTITY in T-
PAS. On the other hand, Entity is the label used in DOLCE for the top node, which
corresponds to ANYTHING in T-PAS. We regard ANYTHING as a better term for the top
node of the system, as the word ENTITY is often used in the linguistic literature in a way
which excludes events. Note that ANYTHING is T-PAS has a specific interpretation as it
stands for all semantic types that play the role of participant in the event described by the
verb selecting them (“PARTICIPATION” relation between endurants and perdurants in
DOLCE).
Endurants and the Object / Stuff distinction
DOLCE PhysicalEndurant corresponds to PHYSICAL ENTITY in T-PAS; the internal
organization of the two nodes, however, differs. Amount of Matter is a sister node of
PhysicalObject in DOLCE (together with FEATURE), while in T-PAS, its closest
equivalent (STUFF) is a type of PHYSICAL ENTITY (INANIMATE PHYSICAL ENTITY, see Fig.
3). The solution in DOLCE appears more adequate, as the animate/inanimate distinction
apparently applies only to objects with Unity.
In T-PAS, BODY and PART OF BODY are child nodes of PHYSICAL ENTITY, and sister
nodes of ANIMATE and INANIMATE. The “CONSTITUTION” relation, used in DOLCE
for co-located entities, as in the case of a person (agent) and its organism (not agent) (see
[8]), and the “PARTHOOD” relation, which defines the relation between a body and its
parts, are not represented in T-PAS, in which the only relation between the semantic
types is the “IS_A” relation. In the future it would be convenient to expand the relations
in T-PAS to include “CONSTITUTION” and “PARTHOOD”.
Abstracts and the tangible / intangible distinction
ABSTRACT ENTITY in T-PAS defines all intangible entities. DOLCE distinguishes among
Abstract (entity without temporal qualities, such as mathematical objects) and
NonPhysicalEndurant (entity with temporal properties such as Mental and Social Object,
see Fig. 2); these two categories appear in different nesting levels. There is no possible
one-to-one alignment in this case. For application purposes, the two DOLCE’s category
can be conflated into T-PAS ABSTRACT ENTITY, inasmuch as the latter does not draw a
distinction between intangible entities with or without temporal qualities. Such a
modeling decision, however, is far from being without consequences.
Agents and the animate / inanimate distinction
The Agent label is used in DOLCE to express a potential agent, that is, a living being
endowed with intentions, beliefs, and desires. Physical objects that have intentionality
(i.e. the ability to direct / handle objects or states of the world) are called Agentive, those
that have no intentionality are called NonAgentive.
In T-PAS, agent is not present, as it is considered a role assumed by a HUMAN in an
eventuality rather than a type - a thematic role in linguistic terms, which, according to
Guarino [9], corresponds to the processual role theorized by Loebe. Therefore, the
DOLCE Agentive / NonAgentive PhysicalObject distinction does not have a direct
equivalent in T-PAS. The closest type to which DOLCE's AgentivePhysicalObject can
be associated in T-PAS is ANIMATE. In T-PAS ANIMATE subsumes, among others,
HUMAN and HUMAN GROUP; it does not include PLANT but it includes the taxonomy of
the animal kingdom (ANIMAL and ANIMAL GROUP). The animal kindgdom differs from
the scientific taxonomy of Linnaeus. T-PAS includes semantic types for animals for
whom there exists a verb that selects the class or species as argument (see Fig. 1);
typically these are verbs of sound emission such as barking (DOG), or verbs of motion
such as galloping (HORSE). Linnaeus categories such as MAMMAL are not present, as no
verb has been identified yet that selects for it.
Feature and parasitic entities
DOLCE assumes the category Feature for “parasitic entities” that are constantly
dependend on physical objects (their so-called Hosts). Feature subsumes Place (holes in
a cheese) and Relevant part (bumps or edges). T-PAS does not have a type that matches
Relevant Part but has APERTURE as a type of LOCATION, which can be aligned to
DOLCE’s Place category.
Locations
In T-PAS we find the semantic type LOCATION, which is used for both natural places and
artifactual ones (an island, a parking lot). DOLCE has the category Place, which,
however, does not correspond to T-PAS LOCATION. In DOLCE, the spatial dimension is
considered a Quality of an entity (specifically Spatial Location > Spatial Region). There
is therefore no direct mapping between the two systems as regards the type LOCATION.
From a linguistic point of view, the solution in T-PAS appears more apt to account for
the geographical entities denoted by words that qualify as independent entities:
mountains, lakes, islands, and so forth.
Natural kinds / Artifactual Types distinction
Neither DOLCE nor T-PAS draw a distinction between manufactured objects and natural,
mind-independent entities. T-PAS has ARTIFACT as a type of INANIMATE but does not
have its counterpart NATURAL KIND. DOLCE has neither one nor the other. In the ENTITY
branch of T-PAS and the Endurant branch of DOLCE the prevailing distinction is that
between concrete and abstract, and between individuated (i.e. with Unity) and mass
(without Unity). The distinction between natural kind and artifactual type is orthogonal
to the other categories: for example, STUFF in T-PAS subsumes both natural entities
(METAL) and artifacts (CLOTH), LOCATION subsumes both natural entities (HILL) and
artifactual ones (ROUTE), and so forth. Note that from a linguistic perspective, the
distinction between individuated and mass appears to be the most represented formally
in the world’s languages, that is, the grammatical behavior of nouns appears to be
primarily determined by their encoding individuals or masses (cf. [10, 135]).
Types vs. roles
T-PAS has FOOD and BEVERAGE as types of ARTIFACT. In [7, 218] it is observed that
“nothing is necessarily food, and just about anything is possibly food”. In other words,
food is considered a role that an entity can play in a food event, not an ontological
category (roles being anti-rigid properties that characterize the way something
participates to a contingent event). While sharing the same theoretical position, in T-PAS
it is believed that there being a large numbers of verbs selecting for the two types
(currently 78 for FOOD and 11 for BEVERAGE) it is pragmatically useful to keep the two
labels in the repertoire of semantic types, also in the light of the presence of artificial
food, that is, man-made entities to be consumed as food.3
Systematic polysemy
Systematic polysemy is the phenomenon whereby a word or expression exhibits an
alternation of meanings that is also exhibited by other words in the lexicon, so that this
alternation can be considered “regular” (in Apresjan’s terms: cf. [10] for an overview),
for example “content” | “container” in the case of glass, dish, bottle (“break a glass” vs.
“drink a glass”), and “physical object” | “information” in the case of book, letter, novel
(“The book is heavy to carry around” vs. “The book examines the life of Dante”). These
cases are currently treated in T-PAS through multiple inheritance, that is, a semantic type
inherits from more than one type, and each subsumption relation implicitly represents
one of the types that are conflated in the ambiguous class. For example, the DOCUMENT
type (instanciated by nouns such as libro ‘book’ and lettera ‘letter’) inherits from the
ARTIFACT type and the INFORMATION SOURCE type: this is a case of cross domain multiple
inheritance, as the two types are situated in different branches of the type system
(PHYSICAL ENTITY and ABSTRACT ENTITY respectively). Systematic polysemy is
currently not represented in DOLCE: see, however, the proposal in [11] to formally
represent complex categories that ‘overlap’ with disjoint domains of entities using the
ontological relationship of “CONTITUTION” and the notion of coincidence.4
Although it is used in other lexical ontologies such as WordNet, multiple
inheritance is normally avoided in formal ontologies because it introduces incoherence
and inconsistency from an ontological perspective and it creates problems for calculating
inferences. It makes the hierarchy more like a graph than a tree and the inference is more
complex. However, for the purposes of natural language processing task such as sense
disambiguation we believe that multiple inheritance constitutes a valid ad interim
solution to the problem of systematic polysemy, until a formal characterization is
standardized.
Leaf Categories
DOLCE is an upper level ontology, and for this reason it does not comprise fine-grained
child categories. T-PAS comprises just as many child categories as they are requested by
verb selectional behaviour: the set of categories is not finite and may increase as long as
the analysis of new verbs requires new semantic types.
At present the deepest type in the system is the type WINE (stuff > fluid > liquid >
beverage > alcoholic drink > wine), motivated by a specific sense of the verbs
invecchiare (‘to grow old’) and maturare (‘to ripen’). The largest node is Artifact, with
35 types. Child categories in T-PAS disclose the anthropocentric character of the type
3
Note that T-PAS also contains a semantic type called ROLE in the PROPERTY branch, used for words and
expressions that denote functions, for example Prime Minister in the expression: “Mattarella has appointed
Giovanni Conte Prime Minister" (see Fig. 1).
4
Specifically, [11] proposes to characterize such cases as ontological constructs based on coincidence
relations between entities of disjoint categories.
system, that does not reflect the state of the art in scientific knowledge but how everyday
speakers talk about the world: for example, the Artifact node includes WEAPON (and its
subtypes FIREARM and BOMB), three subtypes of VEHICLES (ROAD, FLYING and WATER),
FOOD, BEVERAGE, BUILDING, GARMENT, FURNITURE, and peculiar types such as FLAG and
STRING.
5. Concluding Observations
We illustrated the main features of the system of semantic types developed in T-PAS
starting from the analysis of linguistic evidence from corpora, and we compared it with
DOLCE foundational ontology, whose categories are not the result of a data-driven
approach.
In general, the study shows that the analysis based on linguistic evidence induces
semantic types that can be linked to the upper level of a top-down ontology like DOLCE
quite successfully, at least as far as the endurant domain is concerned. Two substantial
issue emerge: first, the category ABSTRACT in T-PAS maps to two disjoint classes in
DOLCE (Abstract and NonPhysicalEndurant) and there is no straightforward one-to-one
alignment in this case; secondly, the phenomenon of systematic polysemy is currently
not represented in DOLCE, while it is systematically encoded as multiple inheritance in
T-PAS.
Furthermore, the study shows that the data-driven inventory of types in T-PAS is
populated by semantic types that point to cognitive categories that are relevant to human
communication, which do not necessarily match scientific classifications: hence the
“anthropic” character of the T-PAS system.
In the alignment described in the previous section, DOLCE, whose structure satisfies
the OntoClean methodology, provides several hints for improvement of the general
architecture of the type system in T-PAS, where the only relation is subsumption: for
example, the relation of “PARTHOOD” and the relation of “CONSTITUTION” will be
helpful to improve the current treatment of BODY, PART OF BODY, HUMAN GROUP, and
ANIMAL GROUP. Conversely, from an applied perspective, T-PAS provides the “leaf
categories” that DOLCE does not possess - being DOLCE an upper level ontology. The
benefit of the experiment can be confirmed as reciprocal.
As regards future work, a first step concerns the completion of the alignment with
DOLCE's NonPhysicalEndurant, Perdurant and Quality branches, not included here. A
second step concerns systematic polysemy, and the implementation of a distinct
treatment of complex categories [11] and metonymic shifts [12]. Finally, a third step
regards the evaluation of corpus-derived semantic types agains automatically obtained
clusters of argument fillers, in line with works such as [13] and more recently - in a
distributional semantic framework - [14].
Acknowlegments
We would like to thank three reviewers for their helpful comments. All remaining errors
and infelicities are our own.
References
[1] Jezek, Elisabetta & Magnini, Bernardo & Feltracco, Anna and Bianchini, Alessia & Popescu, Octavian.
T-PAS: A resource of Typed Predicate-Argument Structures for Linguistic Analysis and Semantic
Processing. In Calzolari, Nicoletta et al. (ed) Proceedings of the Ninth International Conference on
Language Resources and Evaluation (LREC’14), 890-895. Paris, ELRA, 2014.
[2] Masolo, Claudio & Borgo, Stefano & Gangemi, Aldo & Guarino, Nicola & Oltramari, Alessandro.
WonderWeb Deliverable D18: Ontology Library (final). IST Project 2001-33052 WonderWeb: Ontology
Infrastructure for the Semantic Web, 2003.
[3] Hanks, Patrick. Lexical Analysis: Norms and Exploitations. Cambridge MA: The MIT Press, 2013.
[4] Gangemi, Aldo & Guarino, Nicola & Masolo, Claudio & Oltramari, Alessandro & Schneider, Luc.
Sweetening Ontologies with DOLCE. In Proceedings of the International Conference on Knowledge
Engineering and Knowledge Management: 166-181. Berlin – Heidelberg: Springer, 2002.
[5] Bach, Emmon. The Algebra of Events. Linguistics and Philosophy, 9 (1986), 5-16.
[6] Guarino, Nicola & Welty, Christopher A. Evaluating Ontological Decisions with OntoClean.
Communications of the ACM 45(2) (2002), 61-65, New York, ACM Press.
[7] Guarino, Nicola & Welty, Christopher A. An overview of OntoClean In Handbook on Ontologies: 151-
171. Berlin – Heidelberg, Springer, 2009.
[8] Gaio, Silvia & Borgo Stefano & Masolo Claudio & Oltramari Alessandro & Guarino Nicola.
Un’introduzione all’ontologia DOLCE. AIDA informazioni 28(1-2) (2010), 107-125.
[9] Guarino Nicola. BFO and DOLCE: So Far, So Close. Cosmos + Taxis 4 (4) (2017), 10-18.
[10] Jezek, Elisabetta. The Lexicon: An Introduction. Oxford, Oxford University Press, 2016.
[11] Arapinis, Alexandra & Vieu, Laure. A plea for complex categories in ontologies. Applied Ontology 10(3-
4) (2015), 285-296.
[12] Jezek, Elisabetta & Vieu, Laure. 2014. Distributional analysis of copredication: Towards distinguishing
systematic polysemy from coercion. In Basili Roberto et al (eds.) First Italian Conference on
Computational Linguistics CLiC-it 2014, 219-223. Torino: Accademia University Press, 2014.
[13] Snow, Rion & Jurafsky, Daniel & Ng, Andrew Y. Semantic Taxonomy Induction from Heterogenous
Evidence. In Proceedings of the 21st International Conference on Computational Linguistics and 44th
Annual Meeting of the Association for Computational Linguistics, 801–808. Sydney, Australia:
Association for Computational Linguistics, 2006.
[14] Jezek, Elisabetta & Ponti, Edoardo Maria & Magnini Bernardo. Evaluating distributional representations
of verb semantic selection. In Harry Bunt (ed) Proceedings of the 15th Joint ACL - ISO Workshop on
Interoperable Semantic Annotation (ISA-15), IWCS 2019, Gothenburg University, 2019.