Fine-Grained Lexical Semantic Representations and
Compositionally-Derived Events in Mandarin Chinese
Jimmy Lin
MIT Computer Science and Artificial Intelligence Laboratory
Cambridge, MA 02139
jimmylin@csail.mit.edu
Abstract
Current lexical semantic representations for
natural language applications view verbs as
simple predicates over their arguments. These
structures are too coarse-grained to capture
many important generalizations about verbal
argument structure. In this paper, I specifi-
cally defend the following two claims: verbs
have rich internal structure expressible in terms
of finer-grained primitives of meaning, and at
least for some languages, verbal meaning is
compositionally derived from these primitive
elements. I primarily present evidence from
Mandarin Chinese, whose verbal system is very
different from that of English. Many empiri-
cal facts about the typology of verbs in Man-
darin cannot be captured by a “flat” lexical se-
mantic representation. These theoretical results
hold important practical consequences for nat-
ural language processing applications.
1 Introduction
Lexical semantics is becoming increasingly important in
a variety of natural language applications from machine
translation to text summarization to question answering.
Since it is generally agreed that the verb is the locus of
“meaning” in a natural language sentence, theories of ver-
bal argument structure are extremely important for our
understanding of lexical semantics.
An appropriate lexical semantic representation can il-
luminate difficult problems in language processing, ex-
pose facets of meaning relevant to the surface realization
of sentential elements, and reveal insights about the or-
ganization of the human language faculty. In machine
translation, a “good” representation of verbs can straight-
forwardly capture cross-linguistic divergences in the ex-
pression of arguments. In question answering, lexical se-
mantics can be leveraged to bridge the gap between the
way a question is asked and the way an answer is stated.
This paper explores fine-grained lexical seman-
tic representations—approaches that view a verb as
more than a simple predicate of its arguments (e.g.,
Dang et al., 2000). This contrasts with recent semantic
annotation projects such as PropBank (Kingsbury and
Palmer, 2002) and FrameNet (Baker et al., 1998). For
example, while it is undeniable that throw(John, the ball,
Mary), is a valid representation for the sentence “John
threw the ball to Mary”, it is widely believed (at least by
theoretical linguists) that decomposing verbs in terms of
more basic primitives can better capture generalizations
about verb meaning and argument realization. I will ar-
gue that finer-grained semantics is not only theoretically
motivated, but necessary for building applications.
I first provide a brief overview of theories of verbal ar-
gument structure, and then contrast the typology of Man-
darin verbs with that of English verbs. I will present evi-
dence from Chinese that verb meaning is compositionally
“built up” from primitive notions of stativity and activity.
The consequence, therefore, is that “flat” representations
lacking internal structure are unable to capture the verbal
semantics of a language like Mandarin. Productive phe-
nomena such as verbal compounding render enumeration
of all permissible verbs impossible. Verb meaning, there-
fore, must be represented decompositionally in terms of
underlying primitives. This paper does not propose a con-
crete lexical semantic representation, but rather focuses
on the requirements, for natural language applications, of
such a representation.
2 Event Types
The earliest theory of verbal argument structure involves
generalized collections of semantic roles, known as a
case frame (Fillmore, 1968) or a theta-grid (Stowell,
1981) under the framework of Government and Binding
Theory. The idea of semantic roles was first explicated
in Fillmore’s seminal paper, “The Case for Case” (1968),
which argues that the propositional component of a sen-
tence can be represented as an array consisting of the verb
and a number of noun phrases specifically marked with
roles such as agent, patient, instrument, and goal. These
labels identify the grammatically relevant aspects of the
roles that pertain to argument realization in the syntax. A
verb is defined by the semantic roles that it “takes”, i.e.,
its case frame. For example, love takes an agent and a pa-
tient, while frighten takes an experiencer and a stimulus.
A theory of argument structure is not complete with-
out an associated linking theory that explicitly maps ar-
guments in the lexical semantic representation (semantic
roles) to syntactic arguments. Approaches based on se-
mantic roles often formulate a linking theory in terms of a
thematic hierarchy (Jackendoff, 1972): semantic roles are
arranged in an abstract “prominence” hierarchy, and the
realization of syntactic arguments is based on the position
of roles in this hierarchy. The highest role in the thematic
hierarchy is assigned the highest argument position in the
syntactic structure (the subject), the next highest role is
assigned the next highest argument, and so forth. The-
matic hierarchies are believed to be an independent and
irreducible module of grammar.
There has been considerable debate over the ordering
of roles on thematic hierarchies. In fact, the actual in-
ventory of semantic roles, along with precise definitions
and diagnostics, remains an unsolved problem. These are
not the only drawbacks associated with theories of argu-
ment structure that rely on semantic roles:1 Some anal-
yses show that semantic roles are too coarse-grained to
account for certain semantic distinctions. The only re-
course, to expand the collection of roles, comes with the
price of increased complexity, e.g., in the linking rules.
Fillmore’s original assumption that each noun phrase in
an utterance occupies a unique thematic role is often
called into question. For some verbs, e.g., resemble, mul-
tiple noun phrases appear to have the same semantic role.
Finally, because case frames are “flat”, i.e., lacking any
internal structure, a theory based purely on semantic roles
lacks real explanatory power. Why is it, for example, that
love takes an obligatory agent and an obligatory patient?
Why is the instrument role in open optional? These theo-
ries cannot offer satisfactory answers because they do not
directly refer to the meaning of predicates.
Recognizing the drawbacks of theories based purely on
semantic roles, there is now a general consensus among
linguists that argument structure is (to a large extent)
predictable from event semantics—hence, patterns of ar-
gument realization should be inferable from lexical se-
mantic representations grounded in a theory of events.
These event representations typically decompose seman-
1see (Dowty, 1991) and (Levin and Rappaport Hovav, 1996)
tic roles in terms of primitive predicates representing
concepts such as causality, agentivity, inchoativity, and
stativity (Dowty, 1979; Jackendoff, 1983; Pustejovsky,
1991b; Rappaport Hovav and Levin, 1998).
3 From Event Types to Event Structure
Although Aristotle (Metaphysics 1048b) observed that
the meanings of some verbs involve an “end” or a “re-
sult”, and other do not, it wasn’t until the twentieth cen-
tury that philosophers and linguists developed a clas-
sification of event types which captures logical entail-
ments and the co-occurrence restrictions between verbs
and other syntactic elements such as tenses and adver-
bials. Vendler’s (1957) classification of events into states,
activities, accomplishments, and achievements is ground-
breaking in this respect. In his event ontology, activi-
ties and states both depict situations that are inherently
temporally unbounded (atelic); states denote static situa-
tions, whereas activities denote on-going dynamic situa-
tions. Accomplishments and achievements both express a
change of state, and hence are temporally bounded (telic);
achievements are punctual, whereas accomplishments ex-
tend over a period of time. Examples of the four event
types are given below:
(1)
States Activities
know run
believe walk
desire push a cart
Accomplishments Achievements
paint a picture recognize
make a chair find
deliver a sermon lose
Although activities group naturally with states
and accomplishments with achievements in terms
of telicity, it has also been observed that states
can be grouped with achievements and activities
with accomplishments in that that first pair lacks
the progressive tense, while the second pair allows
them (cf. Lakoff, 1966; Shi, 1988). To capture these
properties, Vendler’s classes can be further decomposed
in terms of independent features (cf. Andersen, 1990;
Van Valin and LaPolla, 1997:91-102):
(2) a. State: [¡telic, ¡durative, ¡dynamic]
b. Activity: [¡telic, +durative, +dynamic]
c. Achievement: [+telic, ¡durative +dynamic]
d. Accomplishment: [+telic, +durative
+dynamic]
Vendler’s work on ontological types of events serves
as a foundation upon which others have grounded lexi-
cal semantic representations and theories of verbal argu-
ment structure. Dowty’s seminal work (1979) attempts
to decompose states, activities, accomplishments, and
achievements in terms of the primitives DO, CAUSE, and
BECOME:
(3) a. state: …n(fi1;:::;fin)
b. activity: DO(fi1;[…n(fi1;:::;fin)])
c. achievement: BECOME[…n(fi1;:::;fin)]
d. accomplishment:
[[ DO(fi1;[…n(fi1;:::;fin)])] CAUSE
[ BECOME […n(fi1;:::;fin)]]]
(Dowty, 1979:123-124)
Examples of Dowty’s theory applied to English sen-
tences are shown below:
(4) a. He sweeps the floor clean.
[ [ DO(he, sweeps(the floor)) ] CAUSE
[ BECOME [ clean(the floor) ] ] ]
b. John walks.
[ DO(John, walk) ]
In what later becomes a standard analysis adopted by
subsequent linguists, Dowty breaks causative sentences
down into two subevents: a causing subevent and a result
subevent. The representation of the resultative sentence
(4a) is comprised of the causing subevent “he sweeps
the floor” and the result subevent “the floor is clean”.
Unergative verbs, on the other hand, are represented by
a single subevent with the primitive DO.
Rappaport Hovav and Levin’s more recent theory of
event templates (1998) also defines a basic inventory of
event building blocks in terms of Vendler’s event types:
(5) a. [ x ACT<MANNER> ] (activity)
b. [ x <STATE> ] (state)
c. [ BECOME [ x <STATE> ] ] (achievement)
d. [ x CAUSE [ BECOME [ y <STATE> ] ] ]
(accomplishment)
e. [ [ x ACT<MANNER> ] CAUSE [ BECOME
[ y <STATE> ] ] ] (accomplishment)
(Rappaport Hovav and Levin, 1998:108)
A verb’s meaning consists of a constant paired with a
particular event template drawn from the basic inventory
above. Constants are open-class items drawn from a fixed
ontology (e.g., manner, instrument, state, etc.) and are
represented within the angle brackets of the event tem-
plate. An important claim of this theory is that verbs di-
rectly encode, or lexicalize, complex event structures.
To account for complex events and secondary predi-
cation, Rappaport Hovav and Levin propose a process
called Template Augmentation that allows basic event
templates to be freely “augmented” to any other event
template. This process, for example, explains the resulta-
tive form of surface contact verbs like sweep:
(6) a. Phil swept the floor.
[ Phil ACT<SWEEP> floor ]
b. Phil swept the floor clean.
[ [ Phil ACT<SWEEP> floor ] CAUSE
[ BECOME [ floor <CLEAN> ] ] ]
In this case, an activity has been augmented into an ac-
complishment through the addition of another subevent,
i.e., the floor becoming clean (note similarities with
Dowty’s representation). In order to bring the lexical
semantic representation “into alignment” with syntactic
structure for the purpose of argument realization, Levin
and Rappaport Hovav (1995) propose well-formedness
constraints and linking rules such as the following:
(7) a. Immediate Cause Linking Rule. The argument
of a verb that denotes the immediate cause of
the eventuality described by that verb is its
external argument.
b. Directed change Linking Rule. The argument
of the verb that corresponds to the entity
undergoing the directed change described by
that verb is its internal argument.
Vendler’s ontology of verbal types has paved the way
for many important developments in lexical semantics.
Although the role of lexical aspect in argument realiza-
tion has been called into question (Levin, 2000), this
generally-accepted classification of events figures promi-
nently in most theories of verbal argument structure. Of
great interest both theoretically and for the purposes of
building language applications, therefore, is the typologi-
cal organization of verbal systems in different languages.
Can Vendler’s event type ontology, which was originally
developed for English, be directly applied to other lan-
guages as well? The answer, I will demonstrate, at least
for Mandarin Chinese, is no.
4 The Mandarin Verbal System
I will argue that the typology of Mandarin Chinese verbs
is very different from that of English verbs. Specifically,
I make the following claims:
(8) a. Activity and state are the only two primitive
verbal types in Mandarin Chinese.
Accomplishments and achievements are
derived compositionally.
b. With a small number of exceptions, there are
no monomorphemic verbs in Mandarin that
are telic—no monomorphemic verb
necessarily encodes a result, an end state, or
the attainment of a goal.
c. The particle le, among other uses, signals
inchoativity.
The somewhat controversial claim that Mandarin lacks
monomorphemic accomplishments and achievements has
been previously made by a number of linguists, most no-
tably Tai (1984); see also (Shi, 1988). These works serve
as a starting point for my inquiry into the typological or-
ganization of Mandarin verbs.
One important bit of evidence is the existence of ac-
tivity/achievement verb pairs in English, which are not
present in Mandarin:
(9)
English
activity achievement
look (at) see
listen (to) hear
study learn
look for find
(10)
Mandarin
activity achievement
kan4 ‘look’ kan4 jian4 ‘look-perceive’
= see
ting1 ‘listen’ ting1 jian4 ‘listen-perceive’
= hear
xue2 ‘study’ xue2 hui4 ‘study-able’
= learn
zhao3 ‘look for’ zhao3 dao4 ‘look.for-arrive’
= find
In English, for example, the verb look expresses
an atelic activity, while the verb see expresses a telic
achievement that lexicalizes the attainment of a goal (i.e.,
the successful act of perception). Mandarin Chinese,
however, does not have monomorphemic counterparts
for English achievements. To encode an end state, Chi-
nese speakers must resort to resultative verb compounds,
where the first verb denotes the activity, and the second
verb denotes the result. For verbs of perception, two dif-
ferent result morphemes are typically used: jian4, best
glossed as ‘perceive’, and dao4, literally ‘arrive’.
The claim that resultative verb compounds are required
to explicitly encode the result state is supported by the
grammaticality of sentences that explicitly deny the at-
tainment of the goal:
(11) ta1
he
kan4
look
le5
LE
ban4
half
tian1,
day
ke3shi4
but
mei2
not-have
kan4
look
jian4
perceive
‘He looked for a long time, but couldn’t see it.’
In contrast, using a resultative verb compound in the
first clause triggers a contradiction:
(12) *ta1
he
kan4
look
jian4
perceive
le5
LE
ban4
half
tian1,
day
ke3shi4
but
mei2
not-have
kan4
look
jian4
perceive
intended: ‘He saw for a long time, but couldn’t see
it.’
Another important bit of evidence comes from the in-
terpretations of accomplishments. In English, accom-
plishments are compatible with both in and for adver-
bials, the standard diagnostic for telicity:
(13) a. John wrote a letter for an hour.
b. John wrote a letter in a hour.
As demonstrated in the above example, writing a letter
can be interpreted as either atelic (13a) or telic (13b). The
atelic interpretation is to be understood as “John engaged
in the activity of letter writing for an hour”, whereas the
telic interpretation implies the completion of the letter.
Both readings are generally available, but in the past
tense, the telic accomplishment is much more salient.
Thus, to deny the completion of the goal renders the sen-
tence decidedly odd:
(14) #John wrote a letter yesterday, but he didn’t finish
it.
It is, however, not very difficult to construct a context
that renders the above sentence felicitous:
(15) John is always writing letters, but he never finishes
any of them. In fact, John wrote a letter yesterday,
but as usual, he didn’t finish it.
The situation in Mandarin, however, is very different.
It appears that the Chinese counterpart of write, xie3, has
no reading that necessarily implies completion of the di-
rect object (incremental theme):
(16) wo3
I
zou2tian1
yesterday
xie3
write
le5
LE
yi1
one
feng1
CL
xin4,
letter
ke3shi4
but
mei2
not-have
xie3
write
wan2
finish
‘I wrote a letter yesterday, but I didn’t finish it.’
In fact, the only way to encode completion of the letter
writing is, once again, through a resultative verb com-
pound such as xie3 wan2 ‘write-finish’.
I have thus far demonstrated that the Mandarin
equivalent of many English verbs cannot be expressed
monomorphemically, but rather must involve a verbal
compound. In order to defend my claims, however, the
following (apparent) counterexamples must be explained:
(17) a. shu4
tree
dao3
fall
le5
LE
‘The tree fell.’
b. bo1li2
glass
sui4
shatter
le5
LE
‘The glass shattered.’
It appears that dao3 and sui4 are monomorphemic
verbs that express change of state. In order for my claims
to be correct, I would have to demonstrate that such verbs
are actually derived from more basic forms. Indeed, this
is the case: the examples above are derived from underly-
ing stative predicates—the particle le signals inchoativity.
The following stative/inchoative minimal pair presents
evidence for my theory:
(18) a. shu4
tree
gao1
tall
shi2
ten
gung1fen1
centimeter
‘The tree is ten centimeters tall.’
b. shu4
tree
gao1
tall
le5
LE
shi2
ten
gung1fen1
centimeter
‘The tree grew ten centimeters.’
The only difference in the two above sentences is the
presence/absence of le. The particle, therefore, must con-
tribute the semantic component of inchoativity. Similar
minimal pairs related to prenominal modifiers show this
same contrast:
(19) a. sui4
shattered
(de5)
DE
bo1li2
glass
‘shattered glass’ (stative/adjective)
b. sui4
shattered
le5
LE
de5
DE
bo1li2
glass
‘glass that was shattered’ (resultative
participle)
The above pair represents a subtle but detectable dif-
ference in meaning; whereas (19a) describes a pure state,
(19b) describes the result of an event. This distinction ex-
actly parallels the difference between an open door and
an opened door in English. Once again, since the sen-
tences differ only by le, the particle must be contributing
that semantic component. As further evidence, consider
the following minimal pair:
(20) a. Zhang1san1
Zhangsan
you3
has
yi1
one
da4
big
bi3
amount
qian2
money
‘Zhangsan has a lot of money.’
b. Zhang1san1
Zhangsan
you4
has
le5
LE
yi1
one
da4
big
bi3
amount
qian2
money
‘Zhangsan has acquired a lot of money.’
Once again, the addition of le creates a change of state
acquire out of a simple stative predicate have. The se-
mantic contribution of the particle le is also seen in a sub-
ordinate clause:
(21) a. wo3
I
kan4
see
jian4
perceive
shu4
tree
dao3
fall
zhai4
at
lu4
road
bian1
side
‘I see the fallen tree at the side of the road.’
(tree may have fallen a long time ago)
b. wo3
I
kan4
see
jian4
perceive
shu4
tree
dao3
fall
le5
LE
zhai4
at
lu4
road
bian1
side
‘I see the tree falling at the side of the road.’
(eye witness account)
Once again, the stative reading is contrasted with the
change of state reading. The interpretation of the above
two sentences is consistent with the analysis of le as a
signal of inchoativity.
It is clear from the above minimal pairs that the particle
le combines with stative predicates to gives rise to change
of state interpretations. Are these derived events achieve-
ments or accomplishments? Dowty (1979) provides the
following diagnostics:
(22)
compatible with complement
progressive? of stop
state no ok
activity yes ok
accomplishment yes ok
achievement maybe bad
Accomplishments are generally compatible with the
progressive; some achievements appear felicitous (e.g.,
okis winning), while others do not (e.g., *is noticing).
Accomplishments, since they are durative, are generally
acceptable as the complement of stop, whereas the punc-
tual nature of achievements renders them ungrammatical.
These diagnostics clearly demonstrate that the addition of
le shifts stative predicates into achievements:
(23) a. *bo1li2
glass
zheng4zai4
in.process.of
sui4
shatter
le5
LE
man3
whole
di4
floor
intended: ‘The glass is in the process of
shattering all over the floor.’
b. *bo1li2
glass
ting2zhi3
stop
sui4
shatter
le5
LE
man3
whole
di4
floor
intended: ‘The glass stopped shattering all
over the floor.’
It is interesting to note that many achievements in Man-
darin cannot directly causativize into the transitive form:
(24) a. *Zhang1san1
Zhangsan
dao3
fall
le5
LE
shu4
tree
intended: ‘Zhangsan fell the tree.’
b. ??Zhang1san1
Zhangsan
sui4
shatter
le5
LE
bo1li2
glass
intended: ‘Zhangsan shattered the glass.’
Instead, a resultative verb compound is necessary to
express an accomplishment. Typically, the second verb
denotes the result (end state) of the event, while the first
verb denotes the activity that brings about the end state:
(25) a. Zhang1san1
Zhangsan
kan3
chop
dao3
fall
le5
LE
shu4
tree
‘Zhangsan chopped the tree down.’
b. Zhang1san1
Zhangsan
da3
hit
sui4
shatter
le5
LE
bo1li2
glass
‘Zhangsan shattered the glass.’
Putting all the pieces together, the organization of the
Mandarin verbal system can be summarized as follows:
(26) primitive event types: activity, state
state + le ! achievement
activity + achievement ! accomplishment
Activity and state are the two primitive verbal cate-
gories in Mandarin. Non-causative change of state predi-
cates (achievements) are derived from states with the ad-
dition of the particle le. Accomplishments are further de-
rived from achievements through the formation of resul-
tative verb compounds in which the first verb denotes an
activity, and the second verb the end state.
Traditionally, the particle le that appears post-verbally
has been analyzed as an aspectual marker denoting per-
fectivity (Li and Thompson, 1981). This contrasts with
my analysis of it as a signal of inchoativity. How are
these two approaches to be reconciled? In (Lin, 2004b),
I argue that le is a reflex, rather than an overt realiza-
tion of the underlying inchoative marker. As generally
defined, perfective aspect is not compatible with stative
predicates. However, the addition of a covert inchoative
functional head, in effect, licenses the perfective aspect.
5 Computational Significance?
Why is this peculiar organization of the Mandarin verbal
system important for lexical semantic representations de-
signed for language applications? It demonstrates that, at
least for languages such as Mandarin Chinese, the verb
phrase must be rich in internal structure; a verb cannot be
simply viewed as a predicate of its arguments. Evidence
from Mandarin resultative verb compounds demonstrate
that verbal predicates themselves must be composition-
ally built from underlying primitives.
It is important to note that the formation of verbal com-
pounds in Chinese is a fully productive process—the only
constraint on verb combinations appears to stem from
plausible real-world associations between cause and ef-
fect. The following shows but a small range of possible
resultative verb compounds with the dao3 ‘fall’ result:
(27) kan3 dao3 chop-fall to chop down
zhuang4 dao3 crash-fall to knock over
tui1 dao3 push-fall to push over
la1 dao3 pull-fall to pull down
In principle, verbal compound formation in Mandarin
could be a lexical process, but I present elsewhere in-
dependent evidence for a non-lexicalist approach that
captures these constraints in the theoretical framework
of Distributed Morphology, an extension of Chomsky’s
Minimalist Program (Lin, 2004a; Lin, 2004b). How-
ever, the actual machinery for formalizing these insights
is not important for the present discussion. The important
lessons are the theoretical constraints imposed by verbal
typology on lexical semantic representations designed for
language applications. More specifically:
(28) a. verbs have rich internal structure expressible
in terms of finer-grained primitives of
meaning, and
b. at least for some languages, verbal meaning is
compositionally derived from these primitive
elements.
These claims imply that a PropBank or FrameNet ap-
proach to lexical semantics will not be sufficient for many
language applications, at least for languages such as
Mandarin Chinese. While I may disagree with the tech-
nical details, I believe that the approach taken by (Dang
et al., 2000) is on the right path. Due to the produc-
tivity of verbal phenomena in Mandarin, it is impossi-
ble to exhaustively enumerate all felicitous predicates—
verbal meaning, therefore, must be compositionally de-
rived from primitive elements. This however, does not
mean that PropBank or FrameNet are not useful; quite
the contrary! Existing semantic resources serve as the
foundation from which we can bootstrap finer-grained se-
mantic representations.
While the approach Palmer and Wu (1995) take to lex-
ical semantics captures many selectional restrictions and
finer-grained facets of meaning, it still does not model
the arbitrary productivity of verbal compounds. For the
purposes of translating English change of state verbs into
Mandarin, they developed a conceptual lattice that uni-
fies verbs from both languages. Distances between nodes
in this lattice correspond to “semantic distance”, and is
used to find the closest translation if a specific meaning is
unavailable. Although this approach results in better lex-
ical selection, the semantic lattice still assumes that all
verbal forms can be exhaustively enumerated. Although
this certainly may be true within the context of a specific
corpus, the productivity of Mandarin verbal phenomena
is limitless in the real world.
I believe that, for all languages in general, verbal
meanings are compositionally built up from states and
activities. Furthermore, this process is syntactic in na-
ture (Lin, 2004b), governed by well-known syntactic pro-
cesses such as MERGE (Chomsky, 1995) and subjected
to well-studied constraints such as selectional restric-
tions and the Head Movement Constraint (Travis, 1984).
This contrasts with Rappaport Hovav and Levin’s (1998)
“event template” approach, which is lexicalist in that
large chunks of event structure are directly associated
with verbs. Under their analysis, the lexical entry associ-
ated with sui4 ‘shatter’ would be something like:
(29) sui4 ‘shatter’ =
[ [ x ACT<UNDEF> ] CAUSE [ BECOME
[ x <SHATTERED> ] ] ]
Rappaport Hovav and Levin’s theory argues that a
verb’s meaning is composed of an event template that
captures the structural component of meaning and open-
class constants that capture the idiosyncratic component
of meaning (represented by items in angle brackets). This
separation is a major breakthrough in lexical semantic
theories because it allows grammatically relevant facets
of meaning to be untangled from facets of meaning not
directly relevant to the encoding of arguments. Descrip-
tively, the structural component of meaning is what a verb
shares with other verbs in the same verb class, whereas
the idiosyncratic component of meaning is what separates
verbs within the same verb class.
In Rappaport Hovav and Levin’s account of verbal ar-
gument structure, complex event representations are di-
rectly introduced in the syntax; that is, the verb lexi-
calizes a complete causative accomplishment—to shatter
implicates an agent participating in an unspecified activ-
ity that brings about a change of state where an entity
becomes shattered. In English, they propose that intran-
sitive verbs are derived by a process of “decausativiza-
tion” through which the external argument is “absorbed”,
and therefore remains unexpressed (Levin and Rappaport
Hovav, 1995). Such a theory is unable to account for the
derivation of Mandarin resultatives such as da3 sui4 ‘hit-
shatter’. If (29) is indeed the representation of sui4 ‘shat-
ter’, then what is the lexical semantic representation of
da3 ‘hit’? There are, in principle, two alternatives:
(30) Option 1: da3 ‘hit’ = [ x ACT<HIT> ]
Option 2: da3 ‘hit’ = < HIT >
One might suggest that da3 ‘hit’ is associated with
its own event template that somehow gets merged with
the lexical entry of sui4 ‘shatter’. In order for this ap-
proach to be tenable, one has to explicate the process by
which verbs are “fused” (and in many cases, how argu-
ments of both verbs are sometimes merged or remain un-
expressed); Li (1990) provides exactly such a lexical ac-
count, although it has been found to be problematic for
many cases (Cheng and Huang, 1994). The other op-
tion is to suggest that da3 ‘hit’ merely encodes the id-
iosyncratic component of meaning, without an associated
event template. This, however, cannot be true because
da3 ‘hit’ itself can be used as a main verb:
(31) Zhang1san1
Zhangsan
da3
hit
le5
LE
bo1li2
glass
‘Zhangsan hit the glass.’
The only plausible solution is that verbs encode small
fragments of event structure, which are compositionally
built up by regular syntactic processes. This approach
also provides a natural solution for handling verbs that
are derivationally related to other lexical categories, e.g.,
deadjectival verbs such as flatten, widen, modernize, and
legalize. These derivational affixes obviously contribute
the inchoative component of meaning that turns states
(adjectives) into change of states:
(32) flat: [state flat]
-en: ‚s‚x:BECOME(x; BE(s))
flat-en: ‚x:BECOME(x; BE([state flat]))
In such a treatment, for example, the complete se-
mantics of a word can be compositionally derived from
its component morphemes. This framework, where the
“semantic load” is spread more evenly throughout the
lexicon to lexical categories not typically thought to
bear semantic content, is essentially the model advo-
cated by Pustejovsky (1991a), among others. Such an
analysis of verbal phenomena marks a departure from
the standard architectural view of morphological analysis
as a preprocessor—instead, morphological and syntactic
derivation can be integrated under a common framework.
6 Conclusion
The key claim of this paper is that results from the
theoretical study of verbal argument structure are rele-
vant to computational lexical semantic representations for
language applications. Although the simplest possible
argument representation treats verbs as predicates over
their arguments, I have demonstrated that this approach
is woefully inadequate for handling a language such as
Mandarin Chinese. I have presented evidence that verb
meaning in Mandarin is compositionally built up from
underlying state and activity primitives—this organiza-
tion of the verbal system must be mirrored by any lex-
ical semantic representation aspiring to capture general-
izations about argument realization patterns. This paper
takes an important step in laying out some of the con-
straints for such a representation.

References
Roger W. Andersen. 1990. Papiamentu tense-aspect,
with special attention to discourse. In J. V. Singler,
editor, Pidgin and Creole Tense-Mood-Aspect Sys-
tems, pages 59–96. John Benjamins, Amsterdam, The
Netherlands.
Collin F. Baker, Charles J. Fillmore, and John B. Lowe.
1998. The Berkeley FrameNet project. In Proceedings
of the 36th Annual Meeting of the Association for Com-
putational Linguistics and 17th International Con-
ference on Computational Linguistics (COLING/ACL
1998).
Lisa Lai-Shen Cheng and C.-T. James Huang. 1994. On
the argument structure of resultative compounds. In
Matthew Chen and Ovid Tzeng, editors, In honor of
William S.-Y. Wang Interdisciplinary Studies on Lan-
guage and Language Change, pages 187–221. Pyra-
mid Press, Taipei, Taiwan.
Noam Chomsky. 1995. The Minimalist Program. MIT
Press, Cambridge, Massachusetts.
Hoa Trang Dang, Karin Kipper, and Martha Palmer.
2000. Integrating compositional semantics into a verb
lexicon. In Proceedings of the 18th International
Conference on Computational Linguistics (COLING
2000).
David Dowty. 1979. Word Meaning and Montague
Grammar. D. Reidel Publishing Company, Dordrecht,
The Netherlands.
David Dowty. 1991. Thematic proto-roles and argument
selection. Language, 67(3):547–619.
Charles J. Fillmore. 1968. The case for case. In E. Bach
and R. Harms, editors, Universals in Linguistic The-
ory, pages 1–88. Holt, Rinehart, and Winston, New
York.
Ray Jackendoff. 1972. Semantic Interpretation in
Generative Grammar. MIT Press, Cambridge, Mas-
sachusetts.
Ray Jackendoff. 1983. Semantics and Cognition. MIT
Press, Cambridge, Massachusetts.
Paul Kingsbury and Martha Palmer. 2002. From Tree-
Bank to PropBank. In Proceedings of the Third In-
ternational Conference on Language Resources and
Evaluation (LREC-2002).
George Lakoff. 1966. Stative adjectives and verbs in En-
glish. NSF-Report 17, Harvard Computational Labo-
ratory.
Beth Levin and Malka Rappaport Hovav. 1995. Unac-
cusativity: At the Syntax-Lexical Semantics Interface,
volume 26 of Linguistic Inquiry Monograph. MIT
Press, Cambridge, Massachusetts.
Beth Levin and Malka Rappaport Hovav. 1996. From
lexical semantics to argument realization. Unpub-
lished manuscript, Northwestern University and Bar
Ilan University.
Beth Levin. 2000. Aspect, lexical semantic representa-
tion, and argument expression. In Proceedings of the
26th Annual Meeting of the Berkeley Linguistics Soci-
ety.
Charles N. Li and Sandra A. Thompson. 1981. Man-
darin Chinese: A Functional Reference Grammar.
University of California Press, Berkeley, California.
Yafei Li. 1990. On V-V compounds in Chinese. Natural
Language and Linguistic Theory, 9:177–207.
Jimmy Lin. 2004a. A computational framework for
non-lexicalist semantics. In Proceedings of the 2004
HLT/NAACL Student Research Workshop.
Jimmy Lin. 2004b. Event Structure and the Encoding of
Arguments: The Syntax of the English and Mandarin
Verb Phrase. Ph.D. thesis, Department of Electrical
Engineering and Computer Science, Massachusetts In-
stitute of Technology.
Martha Palmer and Zhibiao Wu. 1995. Verb semantics
for English-Chinese translation. IRCS Report 95-22,
Institute for Research in Cognitive Science, University
of Pennsylvania.
James Pustejovsky. 1991a. The generative lexicon.
Computational Linguistics, 17(4):409–441.
James Pustejovsky. 1991b. The syntax of event structure.
Cognition, 41:47–81.
Malka Rappaport Hovav and Beth Levin. 1998. Building
verb meanings. In Miriam Butt and Wilhelm Geuder,
editors, The Projection of Arguments: Lexical and
Compositional Factors. CSLI Publications, Stanford,
California.
Ziqiang Shi. 1988. The Present of Past of the Particle Le
in Mandarin Chinese. Ph.D. dissertation, University
of Pennsylvania.
Tim Stowell. 1981. Elements of Phrase Structure. Ph.D.
thesis, Massachusetts Institute of Technology.
James H-Y. Tai. 1984. Verbs and times in chinese:
Vendler’s four categories. In David Testen, Veena
Mishra, and Joseph Drogo, editors, Papers from the
Parasession on Lexical Semantics. Chicago Linguistic
Society, Chicago, Illinois.
Lisa Travis. 1984. Parameters and Effects of Word Order
Variation. Ph.D. dissertation, Massachusetts Institute
of Technology, Department of Linguistics.
Robert D. Van Valin and Randy J. LaPolla. 1997. Syntax.
Structure, Meaning and Function. Cambridge Univer-
sity Press, Cambridge, England.
Zeno Vendler. 1957. Verbs and times. Philosophical
Review, 56:143–160.
