Representing Verbal Semantics with Diagrams
An Adaptation of the UML for Lexical Semantics
Andrea C. SCHALLEY
School of Languages, Cultures and Linguistics
University of New England
Armidale, NSW 2351, Australia
andrea.schalley@une.edu.au
Abstract
The paper presents a new way of accounting for
the meaning of verbs in natural languages, using
a diagrammatic notation based on the Uni ed
Modeling Language (UML). We will introduce
the new framework by outlining some model-
ing elements and indicating major differences
to the UML. An extended example will be dis-
cussed in more detail. We will then focus on
the cognitive background of the framework, and
in particular address the question why the usage
of graphical elements within a linguistic model-
ing language proves to be very fruitful. Finally,
we will brie y indicate the potential of the new
framework and its applicability.
1 Introduction
Today, the Uni ed Modeling Language (UML) is
accepted as lingua franca for the design of object-
oriented systems, being widely used for software
development processes. Although the UML has also
been employed in other  elds such as business mod-
eling (cf. the example pro le in the UML speci -
cation, Object Management Group 2003), research
in theoretical and computational linguistics has not
yet tried to apply a graphical language as rich as the
UML. Such an approach will be advocated in this
paper, focussing on the question how verbal mean-
ing is to be represented adequately. Our answer is a
new framework adapted from the UML to model the
meaning of verbs, as developed in extenso in Schal-
ley (2004). This framework for linguistic semantics
is called Uni ed Eventity Representation (UER),
because it is a true extension of the UML and not
just a pro le. Living up to its name, the UER tries
to unify both intuitivity and formality. It employs
intuitive semantic primes as building blocks and in-
cludes these within a framework of speci ed mod-
eling elements. Semantics and syntax of the mod-
eling elements are explicated in the speci cation of
the UER, which was done in the style of the UML
speci cation. Being an adaption from the UML, the
UER introduces a third formal paradigm of com-
puter science into linguistic semantics, one that is
neither functional nor logical but object-oriented in
nature. This is one of the factors contributing to the
cognitive adequacy of the UER. Since the UER is
based on the UML, it can be easily put to use in
computational linguistics.
In Section 2, the UER as adapted from the UML
will be sketched. The cognitive relevance of the
UER is outlined in Section 3  graphical model-
ing elements in general represent prominent kinds
of concepts, or, respectively, structural or meta-
concepts. In particular, we will discuss the impor-
tance of these graphical modeling elements for the
cognitive adequacy of representational frameworks
such as the UER. Section 4  nally comprises an out-
look, listing some potential areas of application of
the UER.
2 A Diagrammatic Modeling Language
for Linguistic Semantics
The UER exhibits a novel use of a diagrammatic no-
tation. But even though it adapts a well-known and
well-elaborated framework used in computer sci-
ence, it constitutes not only a new use of the UML’s
diagrammatic notation but is a redesign in order to
achieve better cognitive and in particular linguistic
adequacy. In order to get a grasp of the character
of the UER we will indicate some major differences
between the UER and the UML and then discuss
an extended example, modeling two particular con-
crete readings of a verb in more detail.
The UER’s focus does not rest on computational
adequacy to the same extent as the UML’s does. It
is not designed to develop software systems, but to
represent meaning. This objective entails, for exam-
ple, that pure software speci c modeling elements
are not part of the UER. Instead, it aims to be close
to conceptualization as revealed in natural language.
Nevertheless, the UML has been a perfect starting
point for the endeavour of developing an adequate
modeling language for verbal semantics, because
on a coarse level of granularity the UML itself sup-
ports cognitive modeling in the sense that it allows
to model software requirements, without going too
deep into implementational issues in the beginning.
But why has a new framework been developed  
why has it not been sensible or feasible to establish
a UML pro le? We believe that for the purpose of
a linguistic modeling language which is designed to
represent verbal semantics, new modeling elements
are required and therefore new metamodel elements
are inevitable.
Roughly, verbs encode events or similar entities,
entities that are called eventities within the UER.1
That is, the semantics of a verb or, as in most cases,
one of its readings, corresponds to the eventity that
is encoded by the verb.2 Then, to represent the
meaning is to model the eventity.
We believe that eventities are conceptual units in
human cognition and comprise particular compo-
nents that are combined in a particular way. In order
to model this appropriately, the UER has a graphical
container for eventity concepts: octagons represent
eventities as such and contain their components in a
structured way. As modeling elements of the UER,
they are called eventity frames and model eventi-
ties, the speci cities of which are graphically con-
tained within these diagrammatic elements. Simi-
larly, the components are again diagrammatic ele-
ments as long as they represent structural or meta-
concepts that hold content. Here the notions of state
and transition (which are part of the UML), or the
new notion of participant class, which is a modeling
element reminiscent of the UML’s classi er role,
could be mentioned. New modeling elements not
being part of the UML, such as the eventity frame
or participant class, clearly establish an extension of
the UML. Therefore, the UER is not a pro le of the
UML, but a close relative.
In addition to de ning new modeling elements
for the UER (and adapting UML ones), the UML’s
division into different modeling views resulting in
separate diagrams (such as class, statechart, activ-
ity, or collaboration diagrams, cf. Object Manage-
ment Group 2003: I 2)3 is given up in the UER.
Since both dynamic as well as static aspects are part
1 The term ‘eventity’ has been adopted from Zaefferer
(2002).
2 We use the term eventity as a term for a kind or type, not an
instantiation or token. Hence, ‘to wake up’ is an eventity,
termed WAKE UP 1 in the following (eventities are usually
notated with capital letters), whereas ‘John wakes up’ is an
instantiation of this eventity.
3 Not all parts of the UML are relevant to our endeavour. Ac-
cordingly, non-relevant parts such as use case and imple-
mentation diagrams are not included into the speci ciation
of the UER at all in order not to overload it and to adjust it
to our purposes.
of an eventity concept (in modeling an eventity one
has to answer the questions what is happening and
to whom it is happening), both aspects are modeled
within one eventity frame and not in several sepa-
rate diagrams. Nevertheless, we take care to distin-
guish the aspects within the eventity frame, with the
dynamic aspects being graphically contained by the
so-called dynamic core, thereby forcing users of the
UER to go for clear distinctions. The integration of
dynamic and static aspects seems to be feasible in
the UER, because  differently from software sys-
tems  we expect eventity concepts not to exceed
a particular level of complexity due to an assumed
maximum of complexity applying to any cognitive
unit. Accordingly, we expect eventity frames to
only come up with straightforward models which
are in general easily manageable.
To get a better idea what a model of an eventity,
or, respectively, a model of a verb’s reading looks
like, we will discuss two readings of wake up in the
following. Consider the eventity frame representing
the semantics of its non-causative reading in Fig. 1
(as in He woke up or in Suddenly Eleni woke up).
«intrinsic» ani : Animacy = animate
[[y]] / Patient : Ineventity
wake_up_1
«do»
«undergo»
y
Awake
«spontaneous»
Figure 1: The non-causative reading of wake up
The octagon depicts the eventity that is encoded
by the verb and hence represents the conceptual unit
corresponding to the verb’s reading. Each even-
tity frame can have a name shown in the upper left
corner, in this case the name wake up 1 was se-
lected according to the  rst ‘wake up’ eventity, the
WAKE UP 1 eventity. The eventity’s components
are nested within the octagon. First of all, partic-
ipants of an eventity are rendered as solid-outline
rectangles similar to UML class symbols and are
attached to the dynamic core by a dashed line (as
in UML collaboration diagrams) that indicates their
status as participants, the participate association. In
the case of WAKE UP 1, there is only one partici-
pant, the undergoer that wakes up and thus endures
the course of events.4 The participant’s name ex-
pression, [[y]], means that y is a representative
name for entities that could potentially be partici-
pants in the eventity (therefore, a notation reminis-
cent of mathematical equivalence class notation has
been chosen).
The dynamic components of the eventity concept
are contained in the dynamic core itself which is dis-
played as a dashed-outlined rectangle with rounded
corners and generally comprises the state machines
of prominent participants.5 For the sake of clarity
(and because it is necessary when there is more than
just one participant), the representative’s name, in
this case y, is cited in the upper left corner of its
state machine. y experiences a transition from an
unknown source state to the known passive simple
state of being Awake. In order to undergo a tran-
sition into the state of being awake and thus to un-
dergo a change of state at all, the undergoer must
have been in a source state differing from the tar-
get state. This is the only information we do have,
we do not know whether the undergoer was asleep,
dozing or unconcious while in the source state. But
y must have been in a state which was not the one
of being awake, such that a transition could result
in the state of being awake. Accordingly, the target
state is speci ed, whereas the source state is dis-
played as an unspeci ed source state. Moreover,
there is no reason and thus trigger for the transition
conceptualized, therefore the transition is marked as
 spontaneous and distinguished from completion
transitions or transitions triggered by signals.
yin the dynamic core is a reference to the partici-
pant class displayed outside of the dynamic core, in
the static periphery. The static periphery in general
depicts the participants, the roles they hold within
the eventity, and the relationships that hold between
them. In the example modeling in Fig. 1, the un-
dergoer y is in fact a patient, i.e., has the role of
undergoing some change of state or condition (and
not only a change of location as a theme, for in-
4 The term undergoer  as well as its counterpart actor  
are taken from Van Valin and LaPolla (1997: 141 147).
5 There are at most two prominent participants in each
eventity, the most active one (the actor, with the stereo-
type  do attached to its participate association) and the
most passive one (the undergoer, with the stereotype  un-
dergo attached to its participate association). Only promi-
nent participants are assigned their own state machine
within the dynamic core.
stance, would). This is indicated in the role speci-
 cation Patientwithin y’s participant class. Ad-
ditionally, in order to be ‘wakeable’ (that is to say,
to be a potential undergoer of a WAKE UP 1 even-
tity), y has to be an animate Ineventity, a
non-eventity entity in the UER’s participant ontol-
ogy which has the intrinsic property of being ani-
mate. The speci cation of both an ontological cate-
gory as well as a property (technically captured in
the UER in form of a type-expression and an at-
tribute) rules out that SLEEP, as an eventity, or a
stone, as an inanimate ineventity, could be potential
undergoers of the WAKE UP 1 eventity. In other
words, the participant class speci cation works like
a  lter on the set of all entities, ruling out those en-
tities which cannot be participants of the eventity in
question because they do not ful l the required char-
acteristics. This is important in describing verbal
semantics because it is an essential part of a verb’s
semantics which selectional restrictions apply.
The  rst, non-causative reading of wake up
is simple in the sense that only one participant
is involved. Turning to the causative reading,
WAKE UP 2 (as in He woke me up or The storm
woke him up), the modeling in Fig. 2 becomes more
complex: there are two interacting participants,
both of which are prominent participants  they
are both assigned their own state machine, each of
which is depicted in a swimlane in the dynamic core
(the two swimlanes are divided by a solid vertical
swimlane border).
«do» «undergo»
yx
[[x]] / Instigator : Entity «intrinsic» ani : Animacy = animate
[[y]] / Patient : Ineventity
wake_up_2
/ Agent / Effector
cause
cause Awake
Figure 2: The causative reading of wake up
The undergoer y essentially endures the same
course of events, although this time the transition
has a trigger in that it is caused by the active insti-
gator, the actor. That is, the transition is triggered
by a cause signal, the receipt of which is rendered
in the concave pentagon in y’s swimlane. The sig-
nal is sent from the actor x, with the signal sending
being represented in x’s swimlane as convex pen-
tagon. The signal sending is the result of an un-
speci ed action state. In other words, the actor per-
forms some action (where action is broadly under-
stood and does not necessarily entail intention), the
nature of which is not conceptualized and irrelevant,
thus leaving the speci cs of the action state irrele-
vant (which is indicated by the underscore). All that
is important is that there is some ‘action’ by x in-
volved so that x wakes y up.6
In the case of the actor’s speci cation in the static
periphery, there are not many restrictions. The actor
is primarily an entity that instigates the transition.
The italicized role description pays tribute to the
fact that natural languages distinguish between vo-
litional and involitional instigators. Instigator is an
abstract role description, meaning that it cannot be
directly instantiated but only by its children (we em-
ploy the object-orientational concept of inheritance
at this point), namely Agent (volitional instigator)
or Effector (involitional instigator). Although
in English users are not forced to decide whether
they are conceptualizing an agent or an effector, in
Hare, for example, an Athapascan language spo-
ken in Canada, users have to mark agents differ-
ently than effectors (cf. Frawley 1992: 207). Thus,
Hare forces its users to de nitely make a decision
whether it is an agent or an effector they are talking
about. If the model was a modeling of Hare (and
not English as it is), we would add the constraint
fdisjointgto the inheritance relations in order
to indicate that the actor can be either an agent or
an effector, but not both at the same time. This way,
natural language speci cities come into the models.
3 Cognitive Relevance of Diagrammatic
Modeling Elements
We trust that the above illustrations suf ce to give
an impression what UER diagrams entail, although
there are details of the diagrams that have not been
explained. But in exactly what way does such a
graphical representation as in Fig. 2 differ from a
Wunderlich-style decomposition as in (1) (cf. also
Wunderlich 1996, 1997)?
(1) CAUSE (x, BECOME (AWAKE(y)))(s)
6 Action states are rendered with convex arcs on both
sides, whereas passive states are shown as rectangles with
rounded corners.
Essentially, the same information concerning dy-
namic structuring seems to be included in (1): there
is an xwhich causes that ybecomes awake. But ob-
viously necessary information about the participants
is not included. Of course, one could add this infor-
mation as in (2) (in this case, the decomposition is
not within Wunderlich’s framework any more).
(2) (AGENT(x)_EFFECTOR(x))^
PATIENT(y)^ANIMATE(y)^
INEVENTITY(x)^INEVENTITY(y)^
CAUSE (x, BECOME (AWAKE(y)))(s)
Comparing (2) to Fig. 2, the diagrammatic represen-
tation is to be preferred for several reasons, one of
these being the intuitivity that is brought forward
in the graphical structure: those modeling elements
that are cognitively connected are graphically con-
nected via connectors, containment resp. nesting, or
visual attachments. That is, the cognitive structur-
ing is re ected in the diagrammatic representation
in a straightforward way, which is not the case in the
linearized representation in (2). Moreover, the ex-
plicit partition of static and dynamic aspects within
one eventity frame as well as the speci ed syntax of
the modeling elements facilitates not only the un-
derstanding of the representation, but at the same
time forces users of the UER to make sure they
produce sound diagrams. That entails re ecting on
what exactly causes, for instance, the undergoer’s
transition in WAKE UP 2. In (1) and (2) it is x as a
participant that directly causes the transition, while
in Fig. 2 it is some action of x that causes the tran-
sition. The latter is more appropriate and also sup-
ported by speakers’ intuition and conceptualization
 something has, in a very broad sense, to ‘happen’
(even if it was pure presence) in order to cause a
transition.7 To represent that ‘something’, a feature
has been included into the UER that is not part of
the UML, namely unspeci ed elements (generally
rendered with underscores in the name slot). These
are elements where only the graphical  and thus
cognitive  structure is present (such as the action
state in the actor’s swimlane in Fig. 2), but no con-
tent of the structure is given. That is, the exact con-
cept is irrelevant and underspeci ed, the only thing
that matters is structure: in Fig. 2 it is merely con-
ceptualized that some kind of action takes place, but
7 This is also supported by the fact that (a) is  ne, but (b) is
not:
(a) The ball broke the window.
(b) *The window broke as a result of the ball.
In other words, explication (b) would need some action of
the ball to be speci ed in order to be sound, such as in (c):
(c) The window broke as a result of the ball’s rolling into it.
the speci cs of the action are not speci ed. Since
the UER aims at representing cognitive structures,
this is a sensible feature  which it would not be
within the UML, of course, as the UML has to head
towards determinism, being a computational mod-
eling language.
The general layout of both the UML and the UER
as graphical languages supports cognitive adequacy.
In graphical languages such as the UML and the
UER, prominent structural concepts can be distin-
guished by non-textual symbols, namely their cor-
responding graphical modeling elements. Exam-
ples in the UER we have seen are the octagon for
eventities, the rectangle for participants, and the
rectangle with rounded corners for passive simple
states. In other words, the structure is passed on
into the graphical-visual domain, whereas the con-
tents are kept in linearized form. Representational
languages in linguistic semantics to date only rarely
distinguish in their formalizations between struc-
tural meta-concepts, such as the ones mentioned
above, e.g., or state and transition, and the con-
cepts themselves, such as the state of being Awake.
Meta-concepts exist at the most implicitly in the
arity of predicates, in particular predicate names
(where the reader has to know that BECOME, for
example, represents a transitional concept), or in
 xed combinations of predicates. Although meta-
concepts are extensively discussed in the literature
(cf. the discussions about Aktionsarten), it is not
taken care to explicate them in a distinct way and to
distinguish them from speci ed concepts, i.e. their
contents.
The UER is the  rst linguistic representa-
tional framework that explicitly accounts for meta-
concepts, rendering them graphically and thus fun-
damentally different from their contents, and dis-
playing different structural concepts with visually
different modeling elements. That way, an intuitive
line is drawn that divides these two levels of rep-
resentation, that divides the speci ed concepts em-
bedded in a structure from the structure itself. The
distinction between structural concepts and ‘ lled’
concepts is in our eyes a very vital one not only
in terms of modeling precision  users have to
clearly distinguish these levels of representation  
but also in terms of cognitive adequacy and univer-
sality. We believe that structural meta-concepts are
universal due to them being based on human expe-
riences, and that they are in principle stable. This
does most certainly not apply in this generality to
contents, although one might assume a very funda-
mental shared knowledge across cultures and thus
languages.
Hence, in  xing the meta-concepts but not their
contents, the UER is a modeling language that can
readily accommodate different linguistic facts and
allows for adequate recording of language speci-
 city due to its  exibility concerning the ‘ lling’ of
the meta-concept. Since it moreover includes the
UML’s extension mechanisms such as constraints
and stereotypes, it can even be more easily adapted
to modeling needs  modeling elements and thus
cognitive components can, if necessary, be adjusted.
The degree of granularity of a representation and
thus the understanding of what is primitive within a
modeling can be aligned to the modeling purposes.
It is our hope that the UER is a modeling language
that can be universally applied to model verbal se-
mantics because of its general  exibility, no matter
what natural language’s verbal system is described
and what granularity is needed for the semantic de-
scriptions.
4 Applicability in Linguistic Research
This last section is devoted to an outlook on what
the UER should be able to do and where poten-
tial applications of this new modeling language can
possibly be found. Within linguistic semantics, the
UER is expected to facilitate and enhance research.
First of all, as has already been indicated, it forces
semanticists to re ect on their representations, in
particular on the structure they model and on what
they consider to be primitive components within
their endeavour. The syntax of the UER has been
speci ed with the aim to allow sensible linguistic
modeling; ensuring that diagrams are syntactically
correct (that is, conform to the speci cation) will be
of invaluable help in semantic work, and the design
of the UER allows, as has been argued, to achieve
cognitive adequacy more easily than other rigorous
linguistic frameworks do. Secondly, the UER sup-
plies mechanisms to capture any potential compo-
nent of verbal semantics, thus allowing for compre-
hensive modeling. Thirdly, case studies applying
the UER framework have shown a strong potential
of the UER in capturing structural relations between
different readings of one lexical item and also be-
tween different lexical items. In other words, the
UER is a practical tool for the study of semantic re-
lationships.
In the study of polysemy, that is, the study of lex-
ical items (elements of our lexicon) and their differ-
ent readings shared on systematic grounds, expli-
cating the different readings in the UER allows to
pin down the systematic changes from one reading
to another (cf., for instance, Chapter 9 in Schalley
2004). A comparison of UER diagrams and intu-
itions of native speakers in cases of polysemy has
shown a thrilling interconnection. The closer the
modeling of the different readings were structurally
to each other, the surer were native speakers intu-
itively that the readings in questions were instances
of polysemy. On the other hand, it seems as if at
least a major change affecting one of the graphi-
cal modeling elements has to occur  such as the
gain of a participant (cf. the two readings of wake
up as modeled in Fig. 1 and 2)  in order for na-
tive speakers to readily identify different readings
of a lexical item and not to consider instances of
both readings to be instantiations of only one read-
ing used in different contexts.
Taking the distinction between structure and con-
tent into account, the UER offers a new perspective
on decompositional semantics. In eliminating all
content within modeled readings and just keeping
the graphical structure, one can ask what changes
the remaining structures can undergo and whether
one  nds instances of such changes in the seman-
tics of natural languages. This enables scholars to
systematically study not only polysemy, but also se-
mantic change and verb classi cation, and to deter-
mine where irregularities in meaning structures of
natural languages are to be found. Moreover, al-
ternations such as causativation or resultatives can
be systematically captured and studied within the
UER.
In addition, it might be interesting to ask what
degree of complexity eventity structures can maxi-
mally obtain while still being lexicalized in a single
verb. In other words: what are possible verbs? This
can be investigated within the UER framework, be-
cause the UER in principle allows for the modeling
of eventities which are too complex to be coded in a
single verb. For instance, one would expect that, if
there are two prominent participants involved (i.e.,
actor and undergoer), there has to be some inter-
action between the participants, some causation to
take place in order for the modeled eventity to be
lexicalized. A verb expressing that something be-
comes liquid and another something becomes full
at the same time is not likely to exist (also cf. Kauf-
mann 1995: 198f.). Systematic analysis applying
the UER can show which structures are most likely
ruled out because they do not constitute a cognitive
unit, with the missing unity showing up in uncon-
nected swimlanes within the dynamic core, for in-
stance.
There are, due to the proximity of the UER to
the UML, not only theoretical linguistic, but natu-
rally computational linguistic applications that sug-
gest themselves. For example, the usefulness of
UER structures for machine translation could be ex-
plored. Pattern matching could be applied to UER
diagrams. Having the semantics of a verb in the
source language captured in a diagram, one could
ask which verb and thus which representation dia-
gram in the target language would be most appro-
priate for the translation.8 This should be the verb
the representation of which comes ‘closest’ to the
modeled one in the source language. Evidently, cri-
teria for what is considered to be ‘closest’ would
have to be identi ed.
Apart from progress in scienti c discovery, the
UER as a UML-based modeling language is a new
modeling language that allows for envisaging ap-
plications also in speech processing systems, for
example. The UER is a rigorous, but cognitively
oriented, non-iconic, but intuitive decompositional
modeling language both suitable for human and
computational usage. We hope that the UER will be
tested extensively and applied to different research
areas within theoretical and computational linguis-
tics. Since it is a rather new modeling language, its
testing is just about to begin, but we are con dent
that the UER will prove very fruitful for many re-
search enterprises.
8 This is of course a simpli cation of the translation prob-
lem, because idiomatic constructions and syntactic environ-
ments are, for instance, not taken into account  or one
would have to model them within the UER as well.

References

William Frawley 1992. Linguistic Semantics. Hills-
dale: Lawrence Erlbaum.

Ingrid Kaufmann 1995. Konzeptuelle Grundla-
gen semantischer Dekompositionsstrukturen. Die
Kombinatorik lokaler Verben und prcurrency1adikativer
Komplemente. (Linguistische Arbeiten 335.)
Tcurrency1ubingen: Niemeyer.

Object Management Group 2003. OMG Uni ed
Modeling Language Speci cation. Version 1.5.

Andrea C. Schalley 2004. Cognitive Modeling and
Verbal Semantics. A Representational Frame-
work Based on UML. (Trends in Linguistics.
Studies and Monographs 154.) Berlin/New York:
Mouton de Gruyter.

Robert D. Van Valin and Randy J. LaPolla 1997.
Syntax: Structure, Meaning and Function. (Cam-
bridge Textbooks in Linguistics.) Cambridge:
Cambridge University Press.

Dieter Wunderlich 1996. Models of lexical de-
composition. Lexical Structures and Language
Use. Proceedings of the International Confer-
ence on Lexicology and Lexical Semantics, ed.

Edda Weigand and Franz Hundsnurscher. Tcurrency1ubin-
gen: Niemeyer, 169 183.

Dieter Wunderlich 1997. Cause and the structure of
verbs. Linguistic Inquiry 28(1):27 68.

Dietmar Zaefferer 2002. Polysemy, polyvalency,
and linking mismatches. The concept of rain and
its codings in English, German, Spanish, and Ital-
ian. DELTA  Documentac‚  ao de Estudos em
Lingcurrency1u· stica T·eorica e Aplicada 18(spe.):27 56.
