From: AAAI Technical Report SS-93-02. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.
Three-Level Knowledge Representation of Predicate-Argument Mapping
for MultilinguM Lexicons
Chinatsu
Aone and Doug McKee
Systems Research and Applications
2000 15th Street North
Arlington,
VA 22201
aonec@sra.com,
mckeed@sra.com
lexical
1
Introduction
levels,
automatic acquisition
arise from predicate-argument mapping mismatches.
Namely, verbs in different languages denoting the
Constructing a semantic concept hierarchy covers same semantic concept do not always represent their
only part of the problem in building any knowledge- thematic roles with the same syntactic structures (cf.
based natural language processing (NLP) system:
McCord[17], Oorr [10]). For example, "I like Mary"
Language understanding requires mapping from syn- is translated into "Maria me gusta a m{", where the
tactic structures into conceptual (i.e. interlingua)
subject and the object in the English sentence are
representation (henceforth verb-argument mapping), the object and the subject in the Spanish sentence.
while language generation requires the inverse map- Furthermore, a verb with two different predicateping. Jacobs [13] proposed a solution in which argument mappings may correspond to two different
he augmented his concept hierarchy in the knowl- verbs in another language. For example, the Enedge base with verb-argument mapping information.
glish word "break," when AGENTis expressed as
While his solution works for monolingual understand- in "John broke the window," is mapped to Japanese
ing/generation, associating each semantic concept of word "kowasu", while without AGENT,as in "The
a verb with English specific mapping information in window broke," "break" should map to "kowareru".
the knowledge base is not ideal for an interlinguabased NLP system, since verb-argument mapping
varies greatly amonglanguages.
2 Three-Level Representation
In this paper, we discuss how the lexicon of
our interlingua-based
NLPsystem 1 abstracts the Each lexical sense of a verb in our lexicon encodes
language-dependent portion of predicate-argument
its semantic meaning (i.e. semantic concept), its demapping information from the core meaning of verb fault predicate-argument mappingtype (i.e. situation
senses (i.e. semantic concepts as defined in the knowl- type), and any word-specific mappingexceptions (i.e.
edge base). We also show how we represent this
idiosyncrasies) in addition to its morphological and
mapping information in terms of cross-linguistically
syntactic information. In the following, these three
generalized mappingtypes called situation types and levels are discussed in detail.
word sense-specific idiosyncrasies. This three-level
representation
enables us to keep the languageSituation
Types
2.1
independent knowledgebase intact and make efficient
use of lexical data by inheritance. Wetake advantage Eachof a verb’s lexical senses is classified into one of
of large multilingual corpora to automatically acquire the four default predicate-argument mapping types
language- and domain-specific lexical information.
called situation types. As shown in Table 1, situaIn addition, this representation framework eas- tion types of verbs are defined by two kinds of inily deals with some of the MTdivergences which formation: 1) the number of subcategorized NPor
1 Ourinterlingua-basedNLPsystemconsists of Solomon,a arguments and 2) the types of thematic roles which
broad-coverage
text understandingsystemfor English, Spanish these arguments should or should not map to. Since
and Japanese, and an experimental generation system which this kind of information is applicable to verbs of any
constructssentencesfrominterlingua representations.
language, situation types are language-independent
predicate-argument mapping types. Thus, in any lan- whole sentence rather than a verb itself (cf. Krifka
guage, a verb of type CAUSED-PROCESS
has two [15], Dorr [11], Moensand Steedman [20]). For inarguments which map to AGENTand THEMEin the stance, as shown below, the same verb can be classidefault case (e.g. "kill"). A verb of type PROCESS-fied into two different aspectual classes (i.e. activity
OR-STATEhas one argument whose thematic role
and accomplishment) depending on the types of obis THEME,and it does not allow AGENTas one ject NP’s or existence of certain PP’s.
of its thematic roles (e.g. "die"). An AGENTIVEACTIONverb also has one argument but the argu- (1) a. Sue drank wine for/*in an hour.
b. Sue drank a bottle of wine *for/in an hour.
ment maps to AGENT(e.g. "look"). Finally,
INVERSE-STATEverb has two arguments which (2) a. Harry climbed for/*in an hour.
b. Harry climbed to the top *for/in an hour.
map to THEMEand GOAL; it does not allow
AGENT
for its thematic role (e.g. "see").
Situation types are intended to address the issue of
Althoughverbs in different languages are classified
into the same four situation types using the same cross-linguistic predicate-argument mappinggeneraldefinition,
mapping rules which map grammatical ization, rather than the semantics of aspect.
functions (i.e. subject, object, etc.) in the syntactic
structures 2 to thematic roles in the semantic strucSemantic
Concepts
2.2
tures maybe different from one language to another.
This is because languages do not necessarily express Each lexical meaning of a verb is represented by
the same thematic roles with the same grammati- a semantic concept (or frame) in our languagecal functions. This mappinginformation is language- independent knowledge base, which is similar to the
one described in Onyshkevych and Nirenburg [22].
specific (cf. Nirenburg and Levin [21]).
Each verb frame has thematic role slots, which have
The default mapping rules for the four situation
TYPE and MAPPING. A TYPE facet
types are shown in Table 2. They are nearly iden- two facets,
tical for the three languages (English, Spanish, and value of a given slot provides a constraint on the type
of objects which can be the value of the slot. This
Japanese) we have analyzed so far. The only difference is that in Japanese the THEME
of an INVERSE- information is used during the semantic disambiguaSTATEverb is expressed by marking the object NP tion process and during Debris Semantics, a semantically based syntax-to-semantics mapping strategy
with a particle "-ga", which is usually a subject
3
employed
whena parse fails (cf. [1]).
marker (cf. Kuno[16]). 4 So we add such information
In
the
MAPPINGfacets,
we have encoded some
to the INVERSE-STATE
mapping rule for Japanese.
cross-linguistically
general
predicate-argument mapGeneralization expressed in situation types has saved
ping
information.
For
example,
we have defined
us from defining semantic mappingrules for each verb
that
all
the
subclasses
of
#COMMUNICATIONsense in each language, and also made it possible to
EVENT¢/: (e.g.
#REPORT#, #CONFIRM#, etc.)
acquire them from large corpora automatically.
map
their
sentential
complements (SENT-COMB)
This classification system has been partially deTHEME,
as
shown
below.
rived from Vendler and Dowty’s aspectual classifications [24, 12] and Talmy’s lexicalization patterns (#COMMUNICATION-EVENT#
[23]. For example, all AGENTIVE-ACTION
verbs
(AKO #DYNAMIC-SITUATION#)
(AGENT (TYPE
#PERSON#
#ORGANIZATION#))
are so-called activity verbs, and so-called stalive verbs
(THEME (TYPE #SITUATION#
#ENTITY#)
fall under either INVERSE-STATE
(if transitive)
(MAPPING (SENT-COMB W)))
(GOAL
(TYPE
#PERSON#
#ORGANIZATION#)
5 However,
PROCESS-OR-STATE
(if intransitive).
(MAPPING (P-AP~G GOAL))))
the situation types are not for specifying the semantics of aspect, which is actually a property of the
Also, by grouping prepositions (for languages like
English
and Spanish) and postpositions (for those
2 We use structures
sinfilar
to LFG’s f-structures.
However,
like Japanese) into semantically related concepts, we
unlike those defined in Bresnan [4] or used in KBMT-89 [5],
we normalize passive constructions.
express other default mapping rules in the knowl3There is a debate
over whether
the NP with "ga" is a
edge base language-independently. For example, the
subject
or object.
However, our approach
can accolmnodate
MAPPINGfacet of the INSTRUMENT
slot in Figeither analysis.
4The GOAL of some INVERSE-STATE verbs
in Japanese
ure 1 requires that the thematic role INSTRUMENT
call be expressed
by a "hi" postpositional
phrase.
However,
of the event @BREAK@
be mapped from a P-ARG
as Kuno [16] points out, since this is an idiosyncratic
phe(i.e. a pre/postpositional argument in the syntactic
nomenon, such infomnation
does not go to the default
mapping
structure) whose pre/postposition
denotes INSTRUrule but to the lexicon.
5h~ both cases, the reverse does not hold.
MENT.In the following three sentences in English,
CAUSED-PROCESS
PROCESS-OR-STATE
AGENTIVE-ACTION
INVERSE-STATE
# of required NP or S arguments
2
1
1
2
default thematic roles
Agent Theme
Theme
Agent
Goal Theme
prohibited thematic roles
Agent
Agent
Table 1: Definitions of Situation Types
CAUSED-PROCESS
AGENT
THEME
PROCESS-OR-STATE THEME
AGENTIVE-ACTION AGENT
INVERSE-STATE
GOAL
THEME
English/Spanish Mapping
Japanese Mapping
(SURFACE SUBJECT)
(SURFACE SUBJECT)
(SURFACE OBJECT)
(SURFACE OBJECT)
SURFACESUBJECT)
(SURFACE SUBJECT)
SURFACESUBJECT)
(SURFACE SUBJECT)
SURFACESUBJECT)
(SURFACE SUBJECT)
(SURFACE OBJECT)
(SURFACEOBJECT) (PARTICLE "GA")
Table 2: Default MappingRules for Three Languages
Note that our representation
can also handle cases
where two internal arguments "flip-flop".
For example, in (5), "infected"
maps its object NP to GOAL
and the PP to THEMEidiosyncratically,
while the
Japanese counterpart
"utsushita"
maps its object
NP ("infuruenza
uirusu-o")
to THEMEand the
("saru-ni")
to GOAL.6 Such idiosyncratic
mapping
for "infect" is stated in the lexicon as in Figure 2.
(#BREAK#
(AKO #ANTI-CREATION-EVENT#)
(AGENT (TYPE #CREATURE# #ORGANIZATION#))
(THEME (TYPE #ENTITY#))
(INSTRUMENT(TYPE #PHYSICAL-OBJECT#)
(MAPPING(P-ARG INSTRUMENT))))
Figure 1: KB entry #BREAK#
Spanish and Japanese, the words which indicate the
instrument used for the #BREAK#event are different (i.e. "with", "con", "-de"). However, by assigning
the same semantic concept INSTRUMENTto each of
these words in the lexicons, the same default mapping
rule is shared across languages.
(3) John broke the vase with a hammer.
Juan rompi5 el vaso con un martillo.
John-wa kanazuchi-de kabin-o kowashita.
2.3
Idiosyncrasies
Idiosyncrasiesslots in the lexicon specify wordsensespecific idiosyncratic phenomenawhich cannot be
(5) a. The researchers
infected
the monkeys with flu
virus.
b. kenkyusha-tachi-wasaru-ni infuruenza uirusu-o
utsushita.
Word-specific (hence language-specific) TYPEidiosyncrasies can be also stated in IDIOSYNCRASIES
slots in the lexicon to deal with someof the semantic mismatches (cf. Barnett et al. [2], Kameyama
el al. [14]). For example,Japanese verbs like "shibousuru (die)" and "shiyousuru (use)" take only
ple as their subjects while "shinu (die)" and "tsukau
(use)" take any animate subjects. Suchword-specific
semantic constraint
lows and overrides
knowledge base.
is defined in the lexicon as folthe constraint
inherited from the
captured by semantic concepts or situation types. In (SHIBOUSURU
(CATEGORY.V)
particular,
subcategorized pre/postpositions
of verbs
(SENSE-NAME.SHIBOUSURU-1)
are specified here. For example, the fact that "look"
(INFLECTION-CLASS.IRS)
(SEMANTIC-CONCEPT#DIE#)
denotes its THEMEargument by the preposition
(IDIOSYNCRASIES (THEME(TYPE #PERSON#)))
"at" is captured by specifying idiosyncrasies as shown
(SITUATION-TYPEPROCESS))
in Figure 2. As discussed in the next section, we derive this kind of word-specific information automati2.4
Sharing
Information
through
Incally from corpora.
heritance
In addition, word-specific mapping rules are stated
in idiosyncrasies slots to deal with thematic divergenTheinformation stored in the three levels discussed
cies (e.g. (4)). In Figure 2, a Spanish entry "gustar"
aboveis shared throughinheritance (cf. Jaeobs [13]).
is shown.
6The Japanese mappingis done by using the situation type
of the verb (i.e. CAUSED-PROCESS)
and the semantic con(4) a. li ke Ma
ry.
cept assigned to the postposition "ni" (i.e. GOAL)
(cf. Section
b. Maria me gusta a m~.
2.2).
3
(LOOK (CATEGORY .
(SENSE-NAME.
LOOK-l)
(SEMANTIC-CONCEPT
#LOOK#)
(IDIOSYNCRASIES
(THEME (MAPPING (P-ARG
(SITUATION-TYPE
AGENTIVE-ACTION))
"AT"))))
(GUSTAR (CATEGORY .
(SENSE-NAME.
GUSTAR-1)
(SEMANTIC-CONCEPT
#LIKE#)
(IDIOSYNCRASIES
(THEME (MAPPING (SURFACE SUBJECT)))
(GOAL (MAPPING (SURFACE OBJECT))))
(SITUATION-TYPE
INVERSE-STATE))
(INFECT
(CATEGORY.
V)
(SENSE-NAME.
INFECT-I)
(SEMANTIC-CONCEPT
#INFECT#)
(IDIOSYNCRASIES
(THEME (MAPPING (P-ARG "WITH")))
(GOAL (MAPPING (SURFACE OBJECT))))
(SITUATION-TYPE
CAUSED-PROCESS))
Figure 2: Lexical entries for "look", "gustar", and "infect"
(BREAK.1
(ISA #BREAK#)
(THEME WINDOW.2)
(TENSE PAST))
(BREAK.5
Figure 4: "The window broke."
Figure 5: "John broke the window."
3
The idiosyncrasies take precedence over the situation types and the semantic concepts, and the situation types over the semantic concepts. For example,
two mappingrules derived by inheritance for "break"
are shown in Figure 3. Notice that the semantic
TYPErestriction
and INSTRUMENT
role stored in
the knowledgebase are also inherited at this time.
(ISA
#BREAK#)
(AGENT JOHN.6)
(THEME WINDOW.7)
(TENSE PAST))
Automatic
Corpora
Acquisition
from
In order to expand our lexicon to the size needed
for broad coverage and to be able to tune the system to specific domains quickly, we have implemented
algorithms to automatically build multi-lingual lexicons from corpora. The three-level representation described here lends itself to automatic acquisition. In
this section, we discuss howthe situation types and
By
separating
language-specific
lexical idiosyncrasies are determined for verbs.
predicate-argument mapping information (i.e. situOur overall approach is to use simple robust
ation types and idiosyncrasies) from semantic concepts, we can capture predicate-argument mapping parsing techniques that depend on a few languagedivergences without proliferating semantic concepts. dependent syntactic heuristics (e.g. in English and
For instance, we generate Japanese sentences from Spanish, a verb’s object usually directly follows the
English within this framework in the following way. verb), and a dictionary for part of speech information. Wehave used these techniques to acquire inforGiven a semantic representation
for the sentence
"The window broke" (cf. Figure 4, detail omit- mation from English, Spanish, and Japanese corpora
ted), a verb is chosen from the Japanese lexicon varying in length from about 25000 words to 2.7 milwhose semantic concept is #BREAK#and which lion words.
expresses its THEMEbut not its AGENT,namely
"kowareru", which has situation type PROCESS-OR-3.1
Acquiring
Situation
Type InforSTATE.On the other hand, given a semantic repremation
sentation for the sentence "John broke the window"
(cf. Figure 5), a verb whose semantic concept maps Weuse two surface features to restrict the possible
to #BREAK#and which expresses both its AGENT situation types of a verb: the verb’s transitivity raling
and THEME,namely "kowasu", which has situationand its subjec~ animacy.
type CAUSED-PROCESS,
is chosen.
The transitivity rating of a verb is defined to be
YPE #ENTITY#))
NT (TYPE #PHYSICAL-OBJECT#)
J
(MAPPING
(P-ARG INSTR~
!
CAUSED-PROCESS
PROCESS-OR-STATE
Situation
Types
~m-l~
V
¯ break"
(MAPPING (SURFACE SUBJECT)
(THEME (TYPE #ENTITY#)
(MAPPENG (SURFACEOBJECT)
(INSTRUMENT(TYPE #PHYSICAL-OBJECT#)
(MAPPING (P-ARG INSTRUMENT)
Figure 3: Information from the KB, the situation
argument mappings for the verb "break."
~xioon
(MAPPING (SURFACESUBJECT)
INSTRUMENT TYPE #PHYSICAL-OBJECT#
MAPPING P-ARG INSTRUMENT
l
type, and the lexicon all combine to form two predicate-
Transitivity:
CP/IS
,
,
..
~ .... .,.--.¯
().0 O1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Ambig : ¯ : .. :...:
- :. --:-- -:
.....
t
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
AA/PS
..........
o o 02 03 04
66 67 d8 d9 i o
Subject Animacy:
CP/AA ().0
Ambig.
IS/PS
d.1 (~.2 6.3
,
, ee t ¯ ¯ ~,e
eee
,ee
6.5 0.6 0.7 0.8 0.9 1.0
¯
,- -.. ,.
¯
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
,
,
¯ ..
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
¯
,
0.8 0.9 1.0
0.8 0.9 1.0
Figure 6: This graph shows the accuracy of the Transitivity
and Subject Animacymetrics.
verb
SUFFICE
TIME
TRAIN
WRAP
SORT
UNITE
TRANSPORT
SUSTAIN
SUBSTITUTE
TA RG ET
STORE
STEAL
SHUT
STRETCH
STRIP
THREATEN
WEAR
TREAT
TERMINATE
WEIGH
TEACH
SURROUND
TOTAL
VARY
WAIT
SPEAK
SURVIVE
UNDERSTAND
SURGE
SUPPLY
SIT
TEND
BREAK
WRITE
WATCH
SUCCEED
STAY
STAND
TELL
SPEND
SUPPORT
SUGGEST
TURN
START
LOOK
THINK
TRY
WANT
USE
TAKE
0CC8
8
15
2O
22
25
27
28
32
33
36
36
36
36
53
57
58
61
77
79
81
82
85
97
112
130
139
146
180
188
188
199
200
219
243
268
277
300
310
368
445
454
57O
852
890
1084
1227
1272
1659
2211
2525
TR
0.6250
0,8333
1.0000
0,7222
0.4211
0.5833
0,8571
0.9062
0.7500
0.7778
0.9091
0.9167
0.2400
0.5278
0.7609
0.8793
0.8033
0.8052
0,9726
0,2069
0.7794
0.8000
0.0515
0.1354
0.1923
0.1667
0.4754
0,6946
0.0182
0.7176
0,0625
0.8594
0.4771
0.4637
0.7069
0.5379
0.2156
0.2841
0.8054
0.3823
0.8486
0.7782
0.3418
0.3474
0.1718
0.7602
0.7904
0.8559
0.8416
0.7447
SA
0.0000
1.0000
1.0000
0.6667
1.0000
1.0000
0.6667
0.6842
0.5000
0.8000
1.0000
0.6667
0.5000
0.5000
0.8571
0.4419
0.6667
0.8000
1.0000
0.5294
0.6875
0.6667
0.2759
0.0294
1.0000
0.7500
0.3846
0.8684
0.3125
0.8571
0.7027
0.4340
0.5000
0.9123
0.8462
0.8899
0.6604
0.7237
0.8101
0,8125
0.5370
0.5918
0.5891
0.6221
0.6520
0.9237
0.8743
0.8787
0.7725
0.5933
Pred.
(IS)
ST
Correct ST
IS)
CP)
CP PS)
(CP)
(CP AA)
(CP AA)
(CP)
(CP)
(CP PS)
(CP)
(CP)
(CP)
(cP is)
CP IS AA PS)
CP IS AA PS)
(CP
/i%~ IIsS/
(CP IS)
(CP IS)
(CP IS)
(is ps)
(IS PS)
(CP IS)
(CP)
(CP IS)
(IS)
CP)
CP PS)
(CP PS)
(is)
(CP IS)
(IS PS)
CP
/ IS
CP IS
PS)
(IS PS)
IS AA
(cp)
(CP)
(CP PS)
(CP PS)
AA)
AA cp)
IS PS)
IS)
PS)
cPisAA;
t
ICP
IS PS)
CP IS)
(PS)
(CP IS)
(AA PS)
(IS)
(IS PS)
(CP IS AA PS)
CP IS)
CP IS AA PS)
(CP IS AA PS)
(CP IS AA PS)
CP IS)
CP IS AA PS)
(IS)
(IS)
IS PS)
CP IS AA PS)
(CP IS AA PS)
(CP IS
cp)
AA PS)
CP IS)
(CP PS)
(CP AA)
OP)
CP PS)
(PS)
(PS CP AA)
(CP)
(CP)
(CP IS)
(CP IS)
PS
l
I
I
(AA PS)
is)
Prepositional
Idio
at
up over in with
out
for
on
from
up for
over into out from
from into of
over
over
as
on with into
at
at
from over
for up
out at up
with
on with at out in up
up into out
off
out over
out up on with at
up at as out on
on over
out into up over
with off out
at into for up
(Is)
(cP)
(CP IS)
(IS)
(CP IS)
over off out into up
Table 3: Automatically Derived Situation Type and Idiosyncrasy Data
6
the number of transitive occurrences in the corpus
Verbs that have a low subject animacy candivided by the total occurrences of the verb. In En- not be either
CAUSED-PROCESSor AGENTIVEglish, a verb appears in the transitive wheneither:
ACTION,since the syntactic subject must map to
the AGENTthematic role. A high subject animacy
¯ the verb is directly followed by a noun, de- does not correlate with any particular situation type,
terminer, personal pronoun, adjective, or wh- since several stative verbs take only animate subjects
pronoun (e.g. "John owns a cow.")
(e.g. perception verbs).
The predicted situation types shown in Figure 6
¯ the verb is directly followed by a "THAT"as a
were
calculated with the following algorithm:
subordinating conjunction (e.g. "John said that
he liked cheese.")
1. Assumethat the verb can occur with every situation type.
¯ the verb is directly followed by an infinitive (e.g.
"John promised to walk the dog.")
¯ the verb past participle is preceded by "BE," as
would occur in a passive construction (e.g. "The
apple was eaten by the pig.")
.
If the transitivity rating is greater than 0.6,
then discard
the AGENTIVE-ACTION and
PROCESS-OR-STATE
possibilities.
If the transitivity rating is below 0.1, then dis.
For Spanish, we use a very similar algorithm, and
card the CAUSED-PROCESS and INVERSEfor Japanese, we look for noun phrases with an obSTATEpossibilities.
ject marker near and to the left of the verb. A high
transitivity
is correlated
with CAUSED-PROCESS . If the subject animacy is below 0.6, then discard the CAUSED-PROCESSand AGENTIVEand INVERSE-STATE
while a low transitivity
corACTION
possibilities.
relates
with AGENTIVE-ACTIONand PROCESSOR-STATE.Table 3 shows 50 verbs and their calcuWeare planning several improvements to our situlated transitivity rating. Figure 6 shows that for all
ation
type determination algorithms. First, Because
but one of the verbs that are unambiguously transisome
stative
verbs can take animate subjects (e.g.
tive the transitivity
rating is above 0.6. The verb
perception
verbs
like "see", "know", etc.), we some"spend" has a transitivity
rating of 0.38 because
times
cannot
distinguish
between INVERSE-STATE
most of its direct objects are numeric dollar amounts.
or
PROCESS-OR-STATE
and
CAUSED-PROCESS
Phrases which begin with a number are not recogor
AGENTIVE-ACTION
verbs.
This
problem, hownized as direct objects, since most numeric amounts
ever,
can
be
solved
by
using
algorithms
by Brent [3]
following verbs are adjuncts as in "John ran 3 miles."
or
Dorr
[11]
for
identifying
stative
verbs.
Wedefine the a verb’s subject animacy to be the
Second, verbs ambiguous between CAUSEDnumber of times the verb appears with an animate
PROCESS
and PROCESS-OR-STATE(e.g. "break",
subject over the total occurrences of the verb where
"vary")
often
get inconclusive results because they
we identified the subject. Any noun or pronoun diappear
transitively
about 50% of the time. When
rectly preceding a verb is considered to be its subject.
these
verbs
are
transitive,
the subjects are almost
This heuristic fails in cases where the subject NPis
always
animate
and
when
they
are intransitive,
the
modified by a PP or relative clause as in "The man
subjects
are
nearly
always
inanimate.
We
plan
to
recunder the car wore a red shirt." Wehave only implemented this metric for English. The verb’s subject ognize these situations by calculating animacy sepais considered to be animate if it is any one of the rately for transitive and intransitive cases.
following:
Acquiring
Idiosyncratic
Informa3.2
¯ A personal pronoun ("it" and "they" were extion
cluded, since they may refer back to inanimate
objects.)
Weautomatically identify likely pre/postpositional
argument structures for a given verb by looking for
¯ A proper name
pre/postpositions in places where they are likely to
¯ A word under "agent" or "people" in WordNet attach to the verb (i.e. within a few words to the
right for Spanish and English, and to the left for
(cf. [19])
Japanese). Whena particular pre/postposition
ap¯ A word that appears in a MUC-4template slot
pears here much more often than chance (based on eithat can be filled only with humans(eft [9])
ther Mutual Information or a chi-squared test [7, 6]),
word
know
vow
eat
w~,nt
resume
possible clausal complements
THATCOMP
THATCOMP,
TOCOMP
TOCOMP
INGCOMP
Table 4: English Verbs which Take Complementizers
we assume that it is a likely argument. A very similar strategy works well at identifying verbs that take
sentential complements by looking for complementizers (e.g. "that", "to") in positions of likely attachment. Some examples are shown in Tables 3 and
4. The details of the exact algorithm used for English are contained in McKeeand Maloney [18]. Areas for improvement include distinguishing between
cases where a verb takes a prepositional arguments,
a prepositional particle, or a commonadjunct.
4
Conclusion
[3] Michael Brent. Automatic Semantic Classification of Verbs from Their Syntactic Contexts: An
ImplementedClassifier for Stativity. In Proceedings of the 5th European ACLConference, 1991.
[4] Joan Bresnan, editor. The Mental Representation of GrammaticalRelations. MITPress, 1982.
[5] Center for Machine Translation, Carnegie Mellon University. KBMT-89:Project Reprot, 1989.
[6] Kenneth Church and William Gale. Concordances for Parallel Text. In Proceedings of the
Seventh Annual Conference of the University of
Waterloo Centre for the New OEDand Text Research: Using Corpora, 1991.
[7] Kenneth Church and Patrick Hanks. Word Association Norms, Mutual Information, and Lexicography. Computational Linguistics,
16(1),
1990.
[8] Ido Dagan and Alon Itai. Automatic Acquisition
of Constraints for the Resolution of Anaphora
References and Syntactic Ambiguities. In Proceedings of the 13th International Conference on
Computational Linguistics, 1990.
Wehave used this three-level knowledge representation for our English, Spanish and Japanese lexicons, and have built lexicons with data derived from
corpora for each language. The lexicons have been
[9] Defense Advanced Research Projects Agency.
Proceedings of Fourth Message Understanding
used for several multilingual data extraction applicaConference (MUC-4). Morgan Kaufmann Pubtions (cf. Aone et al. [1]) and a prototype MTsyslishers, 1992.
tem (Japanese - English). The fact that this representation is language-independent, bidirectional and
amenable to automatic acquisition has minimized our [10] Bonnie Dorr. Solving Thematic Divergences in
Machine Translation. In Proceedings of 28th Andata acquisition effort considerably.
nual Meeting of the ACL, 1990.
Currently we are investigating ways in which thematic role slots of verb frames and semantic type [11] Bonnie Dorr. A Parameterized Approach to Inrestrictions on these slots are derived automatically
tegrating Aspect with Lexical-Semantics for Mafrom corpora (cf. Dagan and Itai [8], Zernik and Jachine Translation. In Proceedings of 3Oth Annual
cobs [25]) so that knowledgeacquisition at all three
Meeting of the ACL, 1992.
levels can be automated.
[12] David Dowty. Word Meaning and Monlague
Grammar. D. Reidel, 1979.
References
[1] Chinatsu Aone, Doug McKee, Sandy Shinn,
and Hatte Blejer. SRA: Description of the
SOLOMON
System as Used for MUC-4. In Proceedings of Fourth Message Understanding Conference (MUC-]t), 1992.
[2] Jim Barnett, Inderjeet Mani, Elaine Rich, Chinatsu Aone, Kevin Knight, and Juan C. Martinez. Capturing Language-Specific Semantic
Distinctions in Interlingua-based MT. In Proceedings of MTSummit III, 1991.
[13] Paul Jacobs. Integrating Language and Meaning in Structured Inheritance Networks. In John
Sowa, editor, Principles of Semantic Networks.
Morgan Kaufmann, 1991.
[14] Megumi Kameyama, Ryo Ochitani, and Stanley
Peters. l~esolving Translation Mismatches with
Information Flow. In Proceedings of 29lh Annual
Meeting of the ACL, 1991.
[15] Manfred Krifka. Nominal Reference, Temporal
Construction, and Quantification in Event Semantics. In R. Bartsch et al., editors, Semantics
and Contextual Expressions. Forts, Dordrecht,
1989.
[16] Susumu Kuno. The Structure
Language. MIT Press, 1973.
of the Yapanese
[17] Michael McCord. Design of LMT: A prologbased Machine Translation System. Computational Linguistics, 15(1), 1989.
[18] Doug McKee and John Maloney. Using Statistics Gained from Corpora in a KnowledgeBased NLP System. In Proceedings
of The
AAAI Workshop on Statistically-Based
NLP
Techniques, 1992.
[19] George Miller, Richard Beckwith, Christiane
Fellbaum, Derek Gross, and Katherine Miller.
Five papers on WordNet. Technical Report
CSL Report 43, Cognitive Science Laboratory,
Princeton University, 1990.
[20] Marc Moens and Mark Steedman. Temporal ontology and temporal reference. Computational
Linguistics, 14(2), 1988.
[21] Sergei Nirenburg and Lori Levin. Syntax-Driven
and Ontology-Driven Lexical Semantics. In Proceedings of A CL Lexical Semantics and Knowledge Representation Workshop, 1991.
[22] Boyan Onyshkevych and Sergei Nirenburg. Lexicon, Ontology and Text Meaning. In Proceedings
of A CL Lexical Semantics and Knowledge Representation Workshop, 1991.
Patterns: Se[23] Leonard Talmy. Lexicalization
mantic Structure in Lexical Forms. In Timothy Shopen, editor, Language Typology and Syntactic Descriptions. CambridgeUniversity Press,
1985.
[24] Zeno Vendler. Linguistics in Philosophy. Cornell
University Press, 1967.
[25] Uri Zernik and Paul Jacobs. Tagging for Learning: Collecting Thematic Relations from Corpus. In Proceedings of the 13th International
Conference on Computational Linguistics, 1990.