=Paper=
{{Paper
|id=Vol-2481/paper51
|storemode=property
|title=Defining Action Types: on the roles of Thematic Structures and Argument Alternations
|pdfUrl=https://ceur-ws.org/Vol-2481/paper51.pdf
|volume=Vol-2481
|authors=Massimo Moneglia,Alessandro Panunzi,Rossella Varvara
|dblpUrl=https://dblp.org/rec/conf/clic-it/MonegliaPV19
}}
==Defining Action Types: on the roles of Thematic Structures and Argument Alternations==
https://ceur-ws.org/Vol-2481/paper51.pdf
Defining Action Types: on the roles of Thematic Structures and Argument
Alternations
Massimo Moneglia Alessandro Panunzi Rossella Varvara
Università di Firenze Università di Firenze Università di Firenze
massimo.moneglia@unifi.it alessandro.panunzi@unifi.it rossella.varvara@unifi.it
Abstract of action conceptualization, the linguistic and the
cognitive levels of categorization are not equiv-
English. The paper focuses on the relation alent and should not be confused (Croft, 2012;
between Action Types (ontological objects Moneglia, 2014). As a matter of fact, the lexical
that identifies the referential potential of a category instantiated by an action verb can refer to
verb) and the Thematic Structure and Al- more than one cognitive entity.
ternations of verbs. The empirical anal- For instance, the verb to push can refer to actions
ysis shows that these linguistic features in which the force causes the movement of the ob-
are not properties of the verb itself, but ject in space (e.g. in a sentence like John pushes
vary in relation to its referential variation. the basket under the table), as well as to actions in
Given this evidence, we argue that The- which the object does not move (e.g. John pushes
matic Structure and Argument Alternation the fabric into a ball). This differential property is
can help in the identification of the differ- more than enough to cognitively distinguish these
ent types of action to which a verb refers, events in different action concepts. As a conse-
providing evidences to define the granular- quence, the need for a cognitive level of action cat-
ity of action concepts in the development egorization which is independent from the lexical
of an ontology of actions. one becomes clear.
In this paper, we investigate the role of one type
1 Introduction of linguistic information, specifically Thematic
Structure and Argument Alternations, in the defi-
Action verbs are among the most frequent words
nition of action types, i.e. types of action concepts
in ordinary communication, and their correct pro-
that gather together single datapoint in the IMA-
cessing constitutes an underpinning element for a
GACT ontology of actions. We point out that The-
wide series of human-machine interaction tasks.
matic Structure is not a property of the verb itself,
The formalization of action verb meanings has of-
since different structures may be present for the
ten been linked to propositional representations
same verb. Our aim is to show how these features
within decompositional approaches (Dowty, 1979;
are linked to action types and how this correlation
Rappaport Hovav and Levin, 2012), in which the
can be useful for the definition and the induction
semantic core of the verb remains a non-analyzed
of Action Types1 .
lexical root. Other traditional components used in
the representation and annotation of the meaning In section 2, we show the innovative methodol-
of action verbs are: the temporal and aspectual ogy assumed by the IMAGACT Ontology of Ac-
properties of verbs (Vendler, 1957; Pustejovsky, tion for representing the meaning of action verbs,
1991); the thematic roles of participants (Fill- focusing on their referential properties rather than
more, 1967; Gruber, 1965); the force dynamics on their intensional definition. In sections 3 and
and causal relations implied (Talmy, 1988; Croft, 4, we will see through a case study that the induc-
2012; Gärdenfors, 2014). tion of the referential variation of verbs can take
Nevertheless, even if these semantic components advantage of linguistic features. Thematic Struc-
are usually assumed to reflect the general structure
1
Similarly, previous work in Word Sense Disambiguation
Copyright c 2019 for this paper by its authors. Use per- (Dang and Palmer, 2005; Roberts and Kordoni, 2012) have
mitted under Creative Commons License Attribution 4.0 In- shown that thematic information can improve verb sense dis-
ternational (CC BY 4.0). ambiguation.
tures and their Alternations can have an impact in that could co-extend to the same scene (Mon-
the definition and characterization of the objects in eglia et al., 2018a). An additional validation,
an ontology of actions. in which raters were asked to assign scenes to
ATs, was conducted with an overall agreement of
2 The IMAGACT ontology 0.8 (Gagliardi, 2014). Lastly, during the ontol-
ogy’s development, Thematic Structure, Alterna-
In the IMAGACT multilingual Ontology of Ac- tions and Aktionsart were manually annotated for
tions2 (Moneglia et al., 2012b; Panunzi et al., the linguistic captions of each scene. These lat-
2014) action concepts are not defined through ter annotations will be the starting point of the
a propositional and truth conditional perspective, present study, in which we analyze the correlation
but they are rather identified and visually repre- between ontological entities (ATs) and linguistic
sented through scenes. Each scene is conceived as features, specifically Thematic Structure and Al-
a prototypical instance (Rosch, 1983) of an action ternations.
concept and constitutes the basic entity of refer-
ence of the action ontology. Scenes have been de- 3 Thematic Structure and Action Types
rived from a complex annotation procedure (Mon-
eglia et al., 2012a) of the occurrences of action Thematic structure and syntactic frame informa-
verbs3 in two large spoken resources of English tion of verbs are usually provided by most lex-
and Italian. After this bootstrapping phase, the ical resources, such as VerbNet (Kipper-Schuler,
ontology has been extended to many other lan- 2005), FrameNet (Fillmore et al., 2004) and Prop-
guages exploiting competence judgments by na- Bank (Palmer et al., 2005). In these resources,
tive speakers (Brown et al., 2014; Pan et al., 2018; the different entries of a verb are associated to
Moneglia et al., 2018b). The whole IMAGACT their possible thematic structures. They include
database is currently comprised of 1,010 scenes manually annotated data and have been useful for
linked to more than 8,700 verbs in 13 languages4 . the development of statistical approaches for Se-
As a result, action concepts have been represented mantic Role Labeling (Gildea and Jurafsky, 2002)
by language independent scenes, each one linked and for various NLP applications (e.g. informa-
to the series of verbs that can be used to refer to tion extraction (Surdeanu et al., 2003), summa-
it. The scenes are described by linguistic captions rization (Melli et al., 2006), and machine trans-
(i.e. short sentences) that have as predicates each lation (Boas, 2002)).
of those verbs. Simultaneously, each verb is con- In this section, we show that Thematic Struc-
nected to a set of scenes in the ontology, represent- ture (TS) is not a property of the verb and we will
ing in this way its referential variation. verify: 1) to what extent it can be considered a
The scenes linked to a verb have been then property of the action types in the variation of a
grouped in broader categories called Action verb; 2) to what extent it can provide a differential
Types5 (hereafter also ATs or Types). ATs are de- feature for the identification of ATs. We consider
fined as higher level concepts which fall in the se- as TS the minimal themathic structure7 which is
mantic variation of a verb, useful to represent its necessary to interpret a verb as an instance of a
referential potential in a more compact way, re- specific AT.
ducing an excessive granularity in the representa- There are cases in which the TS is the same
tion of meaning6 . ATs have been created exploit- all through the verb variation. Frequently, one
ing similarity judgments among scenes and con- specific thematic structure is associated to activity
sidering Local Equivalent verbs, i.e. all the verbs verbs that show almost no variation in their mean-
2
ing. This is the case of the verb to drink, who has
Freely accessible at http://www.imagact.it/
3 only one AT. The verb to close, on the contrary,
Only in their basic, physical meaning, so excluding all
metaphorical and phraseological uses. shows a significant variation in the IMAGACT on-
4
Besides English and Italian, the list of fully mapped lan- tology (7 ATs, four of them represented in table
guage comprenends: Arab, Chinese, Danish, German, Hindi, 1), but all types present the same TS (AG-V-TH).
Japanese, Polish, Portuguese, Serbian, Spanish, Greek.
5
See, for instance, Table 1 which represents the main vari- In these cases, thematic structure cannot play any
ation of the action verb to close. role in the definition of different types, which are
6
As a matter of fact, some verbs in IMAGACT can be
7
linked to several dozen scenes, and the most general ones, The set of roles used in IMAGACT is based mainly on
like to take and to put, are linked to about 100 scenes. the set used in VerbNet.
Mary presses the
fabric into a ball
AG-V-TH-DEST
to push, to squeeze,
to compress
Mary closes her hand Mary closes the umbrella
Table 3: To press, type b
The doctor presses
the shoulder
AG-V-TH
to push, to poke
Mary closes the door Mary closes the lock
Table 1: Variation of to close Table 4: To press, type c
John presses the
identified on the basis of ontological features of button
the theme (e.g. a body part vs an artifact) or by the AG-V-TH
kind of result produced (spatial consequences vs to push
functional consequences), and even by the set of
equivalent verbs which provide a differential prop-
Table 5: To press, type d
erty of each ATs (to shut vs. to lock vs. to close
up vs. to clench). John presses the
Verbs like to close shows that TS is not a nec- pedal
essary differential of ATs, but, as the next exam- AG-V-TH
ples will point out, it can help to select among to push
the interpretations of a general verb. This is the
case with verbs like to press and to push which Table 6: To press, type e
record different TSs possibilities across their vari-
ation. Let’s consider the verb to press. In the IMA-
GACT ontology it shows 10 different ATs. We can and the concept implies a compression as an inter-
observe groups of Types that share the same TS. subjective activity, whereas in type d the Theme
Types a (table 2) and b (table 3) present Agent- is an object and the compression implies a func-
Verb-Theme-Destination structure. In both cases, tional correlation. In type e the Theme undergoes
the destination is necessary to represent the event a continuous scalar pressure, not limited to a sin-
type, which cannot be identified otherwise. In type gle impulse.
a, the Agent compacts the Theme into a block, Although these TS commonalities among types
changing its shape but not its volume. In type b show that TS is not necessarily predictive of a
the Agent squeezes the Theme, reducing its vol- single type, TS helps in the distinction of action
ume. types. For example, TS restricts the range of pos-
sible interpretation of a general verb like to press
John presses the in the case of type a and type b (table 2 and 3)
scraps into a block
versus type c, d and e (tables 4, 5 and 6). The
AG-V-TH-DEST
distinction between these groups of types (which
to push, to compact
is independently motivated) is mirrored by the dif-
ferent TSs.
Table 2: To press, type a TS may constitute an important feature for the
definition of granularity of action types in the verb
Types c, d and e (tab.4, 5, 6) differ from types a variation. Type c (the doctor presses the shoul-
and b since Destination is not necessary and AG- der, tab.4) and type f (the thief presses the gun
V-TH is sufficient to identify the action. into Marys back, tab.7) are distinguished in reason
Despite the common Thematic Structure, they of their TS: they are similar actions from a cog-
clearly identify different actions for cognitive rea- nitive point of view and they can be paraphrased
sons. In type c the Theme is a humans body part, both with to push, but the TS of the verb in the
two events is different. has been previously stated that it shows the cona-
tive alternation, i.e. “a transitivity alternation in
The thief presses the gun
into Mary’s back which the objects of the verb in the transitive vari-
AG-V-TH-DEST ant turns up in the intransitive conative variant
to push as the object of the preposition in a prepositional
phrase headed by the preposition at (sometimes on
with certain verb of ingesting and the push/pull
Table 7: To press, type f
verbs)” (Levin, 1993, p.42). However, only four
types of press allow for the conative alternation ,
Two cases in the variation of to press are char-
as illustrated in the examples below:
acterized by a specific TS: type g (AG-V-TH-
INSTR) and type h (AG-V-TH-RESULT). Type g • Type c: The doctor presses the shoulder →
(tab. 8) necessarily requires the instrument in its The doctor presses on the shoulder
minimal structure, contrary to all other types; type
h (tab. 9) requires the expression of the result of • Type d: John presses the button → John
the action. TS is predictive of the Action Type in presses on the button
those cases.
• Type e: John presses the pedal → John
The tailor presses the
cloth with the iron
presses on the pedal
AG-V-TH-INST
to push
Other types result in agrammatical sentences when
the conative alternation is applied:
Table 8: To press, type g • Type a: *John presses at/on the scraps into a
block
John presses the can flat • Type g: *The tailor presses at/on the cloth
AG-V-TH-RESULT with the iron
to push
Considering now to push, a verb that shares
many interpretations with to press, only some
Table 9: To press, type h types of to push (types a, b, c, d but not e, f and g)
allow this alternation:
Considering the variation of a verb like to
press8 , we can conclude that TS is not peculiar • Type a: John pushes the button → John
of the verb but is related to its types. Given the pushes on the button
cases in which one TS is shared by multiple types,
it is clear that types distinction is not a function • Type b: John pushes the shoulder → John
of the thematic variation (which is actually related pushes on the shoulder
to the intersection of multiple features). However,
TS has a role in type prediction, since it helps iden- • Type c: John pushes the lever → John pushes
tifying the features of a type. on the lever
4 The role of Argument Alternation • Type d: John pushes the pedal → John pushes
on the pedal
Argument Alternations (AAs) are one of those
properties of the verb that have received great at- • Type e: Mary pushes the chair → *Mary
tention in a large body of literature after Levin pushes on the chair
(1993). As we will see, also AAs are not proper-
ties of the whole verb, but rather characterizes the • Type f : Mary pushes the toothpaste → *Mary
verb in its types. Considering the verb to press, it pushes on the toothpaste
8
Further similar examples have been extracted from the
IMAGACT ontology; however, for space limitations, we re- • Type g: Mary pushes the fabric → *Mary
fer only to the cases already discussed. pushes on the fabric
In addition to the conative alternations, other • Type f : the thief presses the gun into Marys
two alternations can be seen in the variation of the back → the thief presses Marys back with the
verbs considered: the resultative construction and gun9
the theme-instrument alternation. The resultative
For the verb to push, only types i and d allow it:
construction presents a phrase that describes the
state achieved by the referent of a noun phrase as • Type i: The thief pushes the gun into Marys
a result of the action. As noted already by Levin back → The thief pushes Marys back with the
(1993, p. 100), it cannot be predicated in case of gun
oblique:
• Type d: John pushes the pedal → John pushes
(1) a. The silversmith pounded the metal the foot on the pedal
flat.
As a whole, if considered together, TS and AA
b. *The silversmith pounded on the metal
can reduce the underdetermination of types. In
flat.
other words, when two types share the same TS,
This alternation is found only in type h for to they can be predicted from a difference in their
press: argument alternations. This is the case, for ex-
ample, for types a (table 2) and f (table 7) of to
• John presses the can → John presses the can press, which share the TS AG-V-TH-DEST, but
flat differ with respect to the theme-instrument alter-
ation: only type f allows it, not type a.
Lastly, we find an alternation between the In the variation of to push, types e and a share
Theme and the Instrument, not listed in Levin the same TS (AG-V-TH) but type e does not al-
(1993). In this case, the Instrument from sentence low the conative alternation (6=Mary pushes on the
2b (which can be seen as the result of a conative chair), contrary to types a (John pushes on the but-
alternation) becomes the Theme in sentence 2c. ton).
(2) a. The doctor pushes the shoulder with 5 Conclusion
his hand
b. The doctor pushes on the shoulder In this paper we have investigated the relation be-
with his hand tween Thematic Structure and Action Types. The
c. The doctor pushes his hand on the empirical analysis reveals that Thematic Structure
shoulder and Argument Alternations are not properties of
the whole verb, but rather of the verb in its Types.
This alternation can be considered as a particu- We have provided evidence about the saliency of
lar case of locative alternation. In terms of Levin both Thematic Structure and Argument Alterna-
(1993), the noun shoulder would represent the lo- tions in the identification of Action Types. Al-
cation argument, whereas hand would be consid- though TS and AA do not determine the variation
ered the locatum. Also in this case, the theme- of a verb across different ATs, these linguistic fea-
instrument alternation does not apply to all types tures can, indeed, reveal characterizing features of
of the variation of to press, but rather characterizes a Type, helping us in the disambiguation of con-
specific types. cepts and in the recognition of the necessary level
of granularity in building our ontologies.
• Type g: the tailor presses the clothes with
the iron → The tailor presses the iron on the
clothes References
Hans Christian Boas. 2002. Bilingual framenet dic-
• Type c: the doctor presses the shoulder → the tionaries for machine translation. In Proceedings of
doctor presses the shoulder with the hand → LREC.
the doctor presses the hand on the shoulder 9
Other types do not allow the theme-instrument alterna-
tion: *John presses the hand on the scraps (type a); *Mary
• Type d: John presses the button → John presses the hand on the clothes (type i). For completeness,
we report some examples of to push for which this alterna-
presses the button with the hand → John tion does not hold: *Mary pushes on the chair with her hand
presses the hand on the button (type e); *Mary pushes the hand on the box (type h).
Susan Windisch Brown, Gloria Gagliardi, and Mas- Massimo Moneglia, Monica Monachini, Omar Cal-
simo Moneglia. 2014. Imagact4all. mapping span- abrese, Alessandro Panunzi, Francesca Frontini,
ish varieties onto a corpus-based ontology of action. Gloria Gagliardi, and Irene Russo. 2012b. The
CHIMERA: Journal of Romance Corpora and Lin- imagact cross-linguistic ontology of action. a new
guistic Studies, (1):91–135. infrastructure for natural language disambiguation.
In Nicoletta Calzolari, editor, Proceedings of the
William Croft. 2012. Verbs: Aspect and causal struc- Eight International Conference on Language Re-
ture. OUP Oxford. sources and Evaluation, pages 948–955. European
Language Resources Association (ELRA).
Hoa Trang Dang and Martha Palmer. 2005. The role
of semantic roles in disambiguating verb senses. In Massimo Moneglia, Alessandro Panunzi, and Lorenzo
Proceedings of the 43rd Annual Meeting on Asso- Gregori. 2018a. Action identification and local
ciation for Computational Linguistics, pages 42–49. equivalence of action verbs: the annotation frame-
Association for Computational Linguistics. work of the imagact ontology. In James Pustejovsky
and Ielka van der Sluis, editors, Proceedings of the
David Dowty. 1979. Word Meaning and Montague
LREC 2018 Workshop AREA Annotation, Recogni-
Grammar. Reidel Publishing Co, Dordrecht.
tion and Evaluation of Actions, pages 23–30. Euro-
Charles J Fillmore, Josef Ruppenhofer, and Collin F pean Language Resources Association (ELRA).
Baker. 2004. Framenet and representing the link
Massimo Moneglia, Alessandro Panunzi, and Lorenzo
between semantic and syntactic relations. Frontiers
Gregori. 2018b. Taking events in hindi. a case study
in linguistics, 1:19–59.
from the annotation of indian languages in imagact.
Charles J. Fillmore. 1967. The case for case. In In Proceedings of the LREC 2018 Workshop WIL-
E. Bach and R. Harms, editors, Universals in Lin- DRE4 4th Workshop on Indian Language Data: Re-
guistic Theory, pages 1–89. Holt, Rinehart and Win- sources and Evaluation, pages 46–51. LREC.
ston, New York.
Massimo Moneglia. 2014. Natural language ontology
Gloria Gagliardi. 2014. Validazione dellontologia del- of action: A gap with huge consequences for natu-
lazione IMAGACT per lo studio e la diagnosi del ral language understanding and machine translation.
Mild Cognitive Impairment. Ph.D. thesis, Univer- In Z. Vetulani and J. Mariani, editors, Human Lan-
sity of Florence. guage Technology. Challenges for Computer Sci-
ence and Linguistics., pages 370–395. Springer,
Peter Gärdenfors. 2014. The geometry of meaning: Berlin/Heidelberg.
Semantics based on conceptual spaces. MIT Press,
Cambridge (MA). Martha Palmer, Daniel Gildea, and Paul Kingsbury.
2005. The proposition bank: An annotated cor-
Daniel Gildea and Daniel Jurafsky. 2002. Automatic pus of semantic roles. Computational linguistics,
labeling of semantic roles. Computational linguis- 31(1):71–106.
tics, 28(3):245–288.
Yi Pan, Massimo Moneglia, Alessandro Panunzi, and
Jeffrey Gruber. 1965. Studies in Lexical Relations. Lorenzo Gregori. 2018. Imagact4all. una ontolo-
Ph.D. thesis, M.I.T. gia per immagini dell’azione per l’apprendimento
del lessico verbale di base delle lingue seconde. In
Karin Kipper-Schuler. 2005. VerbNet: A broad- Anna De Meo and Margaret Rasulo, editors, Usare
coverage, comprehensive verb lexicon. Ph.D. thesis, le lingue seconde, pages 120–150. Officinaventuno.
University of Pennsylvania.
Alessandro Panunzi, Irene De Felice, Lorenzo Gre-
Beth Levin. 1993. English verb classes and alter- gori, Stefano Jacoviello, Monica Monachini, Mas-
nations: A preliminary investigation. University of simo Moneglia, and Valeria Quochi. 2014. Trans-
Chicago press. lating action verbs using a dictionary of images: the
imagact ontology. In Proceedings of the XVI EU-
Gabor Melli, Yang Wang, Yudong Liu, Mehdi M
RALEX International Congress: The User in Focus.
Kashani, Zhongmin Shi, Baohua Gu, Anoop Sarkar,
Bolzano: EURAC research, pages 1163–1170.
and Fred Popowich. 2006. Description of squash,
the sfu question answering summary handler for the James Pustejovsky. 1991. The syntax of event struc-
duc-2005 summarization task. Proceedings of the ture. Cognition, 41:47–81.
HLT/EMNLP Document Understanding Workshop
(DUC). Malka Rappaport Hovav and Beth Levin. 2012. Build-
ing verb meanings. In Miriam Butt and Wilhelm
Massimo Moneglia, Gloria Gagliardi, Alessandro Pa- Geuder, editors, The projection of arguments: Lexi-
nunzi, Francesca Frontini, Irene Russo, and Mon- cal and compositional factors, pages 97–134. CSLI
ica Monachini. 2012a. Imagact: Deriving an action Publications, Stanford, CA.
ontology from spoken corpora. In Proceedings of
the Eight Joint ACL - ISO Workshop on Interopera- Will Roberts and Valia Kordoni. 2012. Using verb
ble Semantic Annotation (ISA-8). Pisa, October 3-5, subcategorization for word sense disambiguation. In
2012, pages 42–47. LREC, pages 829–832.
Eleanor Rosch. 1983. Prototype classification and log-
ical classification: The two systems. New trends in
conceptual representation: Challenges to Piaget’s
theory, pages 73–86.
Mihai Surdeanu, Sanda Harabagiu, John Williams, and
Paul Aarseth. 2003. Using predicate-argument
structures for information extraction. In Proceed-
ings of the 41st Annual Meeting of the Association
for Computational Linguistics.
Leonard Talmy. 1988. Force dynamics in language
and cognition. Cognitive science, 12(1):49–100.
Zeno Vendler. 1957. Verbs and times. The philosophi-
cal review, 56:97–121.