American Journal of Computational Linguistics Microfiche 38 
V~nn PARADIGMS FOR 
SENTENCE RECOGNITION 
MARIANNE CELCE-MURCIA 
DEPARTMENT OF ENGLISH 
UNIVERSITY OF CALIFORNIA 
LOS ANGELES 90024 
Co~yr: yht 1976 by the ,->ssocja tion for Computational ~inguistics 
Summary of "Verb Paradigms for Sentence Rccognit.lonff 
by 
Piqrinnnc Cclcc-Murcio 
Department of English, UCL4 
This paper dascribcs a linguisticnlly-based recognition 
grammar modcl, which was dcvdlopcd as part of a Co~up~~tl.tc~-r\iJcd 
Ins tructinn Project, to the tasks cf recognizing and anolyzin~ 
n variety of basic sentence types in English. Ways of estending 
the model to the analysis of comples sentences are also suggested. 
The procedures and the model described herein are original; 
however, tjhey owe much tb insights found in the work of tws 
linguis ts , Grub,er and Fillmore. 
The general problem of grammar recognition is that of 
going from a surface string of words rn a deep rcprcscntation 
that permits semantic interprcta8ion. PIorc specifically , our 
grammar recognition procedure depends on thc identlfica tion of 
the ~recise function or semantic role that each noun phrase 
actant occurring in a given sentence exhibits with respect tcp 
the verb of that sentence. 
By assigning verbs--or, to be more precise, verb, senses-- 
to one or more paradigms ( , perceptually and functionally 
defined surf ace con£ igurations) , it becomes possible to determine 
algorithmically for every sentence the functional relation (e. g . , 
theme, causal actant, goal, source, locus) that each noun phrase 
in the sentence bears to the verb, thereby assisting greatly Ln 
arriving at a representation of the mearring of each sentence. 
A nuhber of verb paradigms such ae intrans Ltive, transitive 
and ergative Rre defined. Verbs belonging to the intransitive 
paradigm such a8 die, fall, go, etc. always have subjects that 
function ae t%emes. Verbs belonging to the transitive paradigm 
such ae kill, - - read, -S eat etc. have subjects that function as 
causal actants -. - and objects that function as themes. The ergative 
paradigm, which is more complex, cons Fs ts of change-of -state 
verbs such as open, melt, increase, etc. If an ergative para- 
digm verb has both a subject and- an object, the subject is a 
causal actant and the object is a theme; however; if such a 
verb takes only a subject, then the subject functions as a theme. 
The paradigm membership of each verb sense in the data base is 
determined and is recorded as a lexical feature of that verb. 
The number o+ verb paradigms would ,proliferate almost in- 
definitely were it not for several devices, built into the grammar. 
One of these devices is the reversal of transformations such as 
passive and interrogative so that subject and object functional 
relations remain the same as in active, declarative sentences 
Another such device is the recovery of noun phrase dssations 
such as the one occurring in a sentence likel"John ate. " 
In 
this case, an indefinite object is reco~structed and the 
traneitive paradigm feature of the verb eat remains intact. A 
- 
third device of thie nature is use of a notion called incorpor; 
ation. For exardple, a sentence srtch as "It is raining." will 
be analyzed as having incorporated the theme subject rain' into 
4 
the verb with the result that a abstract structure resembling 
"Rain is falling" gets rec80nstructcd and processed as an in- 
transitive verb paradigm item. 
To overview the entire procedure, we start by parping the 
sprface structure of any given sentence. The major constituents 
a,nd parts ofi speech are identified. Nokt, we determine the type 
nf sentence involved ( , declarative, interrogative, imperative) 
and transform the wrd order where necessary. Following that 
the form QE the verb (ie voice and tense) is identified and 
again the word order is transformed if need be. A surface role 
(i*. e. , subject, object) is then assigned to each noun phrase not 
preceded by a preposition. At this time the lexical entry of 
the verb is consulted for features of movement, deletion, in- 
corgoratlon, etc. so that any necessary final adjustments can 
be made. The paradigm me~ibership of the verb is then identified 
thereby permitting a predetermined heutis tic to assign a relation 
(Le., theme, causal actant, locative, etc.) Eo every noun phrase 
in the sentence. Finally, by applying the above information as 
well as other lexical. and grammatical information the surface 
structure is transformed ihto an appropriate abstract structure 
that permi ts setuantic interpretation 
TABLE OF CONTENTS 
Preface ..............mm....m..mm.............m...... Paw 6 
Surface Subjects. Surface Objects 
and Marked vs . Unmarked Noun Phrases ............ ..Page 10 
......... 
Functional Relations: Elementary and Derived Page 1.3 
.............. Classification of Verbs into Paradigms Page 20 
Paradigms and Transformations ....m...mm.m...mm...... Page 30 
Paradigms and Aspect ............................ .... Pa 41 
Paradigms and Incorporation ......................... Page 48 
Frm Surf ace S truc ture to 
Deep Functional Structure .................a.mm *Page 53 
........ Extension 8f the System to Complex Sentences Page 57 
Concluding Statement ....m..........m......mm....m... a 63 
References ........................m.............mm. Page 67 
6 
Verb Paradigms for Sentence Recognition* 
Preface 
For a number of years in the early 1970's the U.S. Air 
Force sponsored research at System Developmefit Corporation in 
the applicati~n of natural language processing techniques to 
computer-aided ins truc ti~p (CAI) . The purpose of this research 
was to increme the overall effectiveness of CAI as an instruc- 
tional method, with particular emphasis on extending the s tudent 
computer interface to permit s tudent-generated f tee f om responses 
and queries. The research included experiments with a deduc- 
tive question-answering system designed for use in CAI, the 
modeling of the behavior of a hypothetical tutzor, and lastly 
the computer evaluation of constructed student responses, and a 
question-answering system driven by a dynamic model of the CAI 
lesson content. The subject matter used in the project was 
intreductory meteorology--specifically concepts relating t.0 the 
nature of psecipitation. 
*A slightly shorter version of this paper originally appeared 
in April, 1972, ag one of several reports generated under Contract 
F33615-70-C-1726, which the Hwn Resources Laboratory of the 
U.S. Air Force entered into with System Development Corporatioh, 
Santa Monica, California. The views expressed herein are those 
of the author and do not necessarily reflect the views of the 
U.S Air Force or the Department of Defense. 
7 
One of the goals oP this C~mputer-Aided Instruction Project 
was the development of an efficient ~ecognitian grammar that is 
as a imple and rcalis tic as possible without s'tlcrif icing in'for 
mation critical to semantic interpretation The primqry purpose 
of this grammar was the analysis of student responges phrased 
in natural English- a necessary step in the answer-evaluation 
process. 
This paper describes in a general way the grammar re- 
cognition model deve1ope.d as part of the CAI Project. 
The model developed was based on the assumption that recog- 
nition grammars should be constructed quite differently 0-om 
production grammars (e., they are something other than mere 
inverse algorithms of pr.oduction grawars and involve certain 
questions of heuristics that are not a£ concern in the constact- 
ion of production grammars); therefore an attempt was made to 
develop a unique model specif ical ly oriented to recognition tasks. 
1 
~. . . . .... . . 
l~h~ author is not a computer programmer but an applied 
linguist, who collaborated on a grammar recogni't ion model for 
the c~mputer-aided instruction project wiilh a team a£ oomputer 
programers and speciali-s ts in artificial. intelligence that in- 
oiMed William J. Schoene, John S. Brown and Robert F. Simmons, 
and the author is greatly indebted to these colleagues for their 
assistance. Ques tzoris conc arning the details of actual programs 
and rourines Lmplemented should therefore be addressed to one 
of these specialists rather than the author, who served the pro- 
ject in a nari-technical capacity as a linguist. 
8 
At the time this CAI project was undertaken two of, the moat 
successful recognition grammars that had been cons truc ted worn 
the ~BEJ "Automated Recognition Grammar for English," of Culi- 
CdVQr, et al. (1969), and the work dona at NIT on Project bfAlbC, 
which is desctibed in Winograd (1971) . The Culicover racognitio 
grammar is based largely on the gener ative-transfometionaL work 
of Chomskg (1965) and others, end is a highly complex mechanism 
that attempts to capture the insights of Chamsky ' s generative- 
transformattonal model from the recognlt.ion gramar point Q£ 
view. Winograd's highly pragmatic recognition procedures, on 
the other hand, make use of a systemic-type of grammar a la 
Halliday (1961, 1966, 1967) ; however, the power of WLnograd ' s 
system is much more in his limited world-model and semantic 
networks than in his gramar. 
Rather than produce an imitation oC either the IBM Recog- 
nition Gramar or the MIT Project--it is, in fact, possible that 
neither of these approaches will ultiiately be the most useful 
one for recognii tion purpbses- -we tr Led to reexamine recent 
insights , especially Fillmore 's Dteep Case Hypo thesis (1968b) , 
Chafe's semantically-based grammar (1970), and Gruber's work on 
the Lexicon (1965, 1967) in order ta see how the functions and 
co-occusrence relations. they describe .might best be incorporated 
int~ a recognition grammar without sacrificing the positional 
insights that much of Chomsky's work has made explicit. In 
setraspect, eur greatest debt was to Grbber (1965) This re 
search was, of necessity, carrred our: somewhat independent of the 
9 
lingu Ls t LC descriptions presented by these authors s ince they 
typically endeavor to remain neutral with respect to questions 
of productton and recognitian in their work. 
INTRODUCTION 
.II 
The problem we faced is, in general terms, the following: 
Given (more or ' less) acoura te pareings of s tudents ' responses, 
how do we, as simply as possible, automate the reconstruction of 
deeper, more abstract message representations that permit us to 
make reasonable semantic interpretations? This paper outlines 
the essentials of our solution to chis prohlern. The solution 
is described in terms of the surfaae structures and deep rep- 
resentations of simple sentences beaause it is eas ies t to under- 
stand and evaluate the solution at this level. The final section 
will di9suss some complex sentence s-tructures and will demon& 
s tra te that the paradigma tic technique discussed in this paper 
can be extended quite naturally to bandle complex sentences as 
well a's simple sentences. 
Given an utterance that is a simple EngLish sentence, or part 
of a simple English sentence, the grammar driving the parser must, 
first of all, be capable of recognizing constituents such as 
noun phrases, prepositianal phrases., finite verbs or verb phrases, 
2 
By 'verb phtase' we mean not the full 'predicate' but only 
the root verb with its tense marker and optional modal and aspect- 
ual auxiliaries, etc. 
10 
and adverbs or adverb phreaee in ehte surface string of mxds. 
The noun phrase preceding the verb phrase is referred to as the 
surface sublect; and if there is one noun phrase following the 
verb, it is identified as the surface oblect. If two noun 
phrases follow the verb phrase, there ,are objects, identi- 
fied as surface object1 end surface ablecr2 respectively. The 
vezb phraae, plus whatever follows it in a simple sentence, has 
traditionally been referred to as thg psemdicate; we shall be 
making. use of the term in this sense too; hotrwer, the notions 
swface subjeCp, surface verb phrase,' and 'surface objece(s) 
are the notions that are most critical in the development of 
our approach. 
Surface Subjects, Surface Objects, and Marked Vs. unmarked Noun 
Phrases 
L 
All noun phrases occurring in English sentences will be 
viewed as actants that bear a specific functional relation to 
the verb or element of predicatio in the sentmenee. 
These func- 
tional relations (e.g., theme, causal actant, locus, etc.) will 
be dfscussed in the nexksection of the paper. We view the 
identification of the precise functions that all noun phrase 
actants in a given sentence exhibit to their verb as the basic 
problem of sentence recognition. English, more so than languages 
such as German or Russian, tends to give relatively little direct 
indication in surhce structure as to what the function of a 
given noup phrase in a sentence ls. This is because noun pbrases 
occurring ae eurface subj ec te and eurface ollj ec ts--with the rx- 
caption of oome pronominal and interrogat~se or relative dorms-- 
are completely unmarked in English ( . they bear no inflection 
that would exclude or eugges t a particular function) . 
Thus 
we may have a eentence with en unmarked noun pnrase as the sur 
face subject: 
( 1) The object deecended. 
Surface Subj . 
The same unmarked noun phrase may occur as a single unmarked 
eurface object (2) or aa one of two unmarked surface objects 
(3) : 
(2) Jqhn saw the object 
Surface Surface 
Sub j . Obj . 
( 3) 
- John ga,ve Mary the gbj ec t. 
Surface Surface Surf ace 
Subj . Obj. 1 Obj. 2 
Thus in sentences such as (1) through (3), the only inform- 
ation we can use if we want to identify the function of the noun 
phrases is: 
( 4) a) 
The order of the NP's with respect to the verb 
(e , subject and object (if present)). 
b) When two objects are present, the order of objects 
with respect to each other (e., obj 1 or obj 
c) The semantic class and syntactic voice (e , active/ 
passive) of the verb. 
In cases such as these, where the only infomation we hevs 
about the noun phrases in a sentence has to do. with their serial 
order, we say thet the noun phrases are unmarked. 
The surface subject of any English sentence is, by definition 
always unmarked ( it is not preceded bg a preposition) ; 
however, it is possible for the other noun phrase(s) in e sen- 
tence all to be marked e, preceded bv a  reposition) (5) , or 
fox the non-subject noun phrase(s) to be a combination of marked 
and unmarked noun phrases (6) : 
(5) 
The object descended to the --  round. 
Surface 
Sub j , 
(by definition: 
unmarked) 
(6) John gave the object to Mary 
Surfkce Surf ace Marked WP 
Sub j Object 
(by definitbn:. (by definition* 
unmarked) unmarked) 
From this point of view, most prepositional phrases are 
analyzed as marked noun phrases m our grammar, and the lexical 
shape of a preposition preceding 
the noun phrase usually gives us 
some information as to what function the following noun phrase is 
fulfilling in the septence. This $nformation is of a different 
order from the information stated in (4) ; however, we believe 
13 
that it must be fully tntegtated with such infamation if sen- 
tenc6es are to be meaningfully recognized. 
At the surface LeQel, then, we are interested in recognizing 
umerked nopn phrase@ and marked noun phrases (i., noun 
phrase$ preceded by prepositi~ns); ufi.thin the class of bmarked 
noun &rases we also distinguish surface subjects from surfficc 
objects in terms of their po8ttion with respect to the verb* 
and where two surface objects occur, we number them object1 
an& object2 on a left-to-right basis. When marked noun phrases 
occur, special attention is given to the preposdtion that pre- 
cedes the noun phrase. The voice of the verb phrase of the 
sentence is also recogntzed at the surface level. None of this 
inform&tion, however, will directly cons titute the deep represen- 
tation of a sentence. Exactly how this surface-structure in- 
Earnation will be used to arrive at accurate deep representations 
will becme clear in subsequent sections of the paper. 
The elementary functianal relations in our grammar are the following: 
theme 
causal actant (C.A.) 
lqcus 
goal 
soarce 
14 
The be3 is the mos;t;: neutral ectant end the one obligatory 
actant in .A aimple sentence. It is the person thing, fact, or. 
state of affair8 about which samctbing is baing predicated such 
se mo'vamant, locatton, ownership (change of location or etate 
af ownirship) , class mecabership, etc. 
The functional selatl~n causal actant refers to tho person, 
object, natural force, fact, or state of affairs etc., ths t; 
functions to cause or bring about the action, change, or situation 
described or implied in the predicate. In addittan, all nouns 
that may fqnctiort as the causal actant in a sentence are marked 
in the lexicon with the feature (+potent)--directly or by re- 
dundancy rules, following Chafe (1970) . Not all sentenpes 
will have a causal actant, 
Where present, the locuC4 is the object or being that the 
'~herne is a syntactically and semantically defined fuactioflal 
reldtion introduced by Gruber (1965). We follow Grubet ' 8 cnpice 
of terminology and note that a sirnilqr functional relation was 
referred to as the '~bjecti~ve case' by Fillmore (1968b) and 
the 'neutral case' by the UCL4 English Syntax Project (1968) 
4~he functional relation locus includes under one category 
the two functional relations that ~ilhare and others have called 
dative and locative. This is qain more closely in line with 
G~ubor's analysis than with otter auailable analyses 
theme is in, at, on, near, etc. Its use also implies that there 
rs no motion of the theme being expressed in the sentence 
(i4eap 
the func tional relation ' locus ' typically occurs with 
s tatlve predicates) . 
The two other elementary Eunctional relations that may in- 
teract with themes, causal ac tants and loci are the directional 
frmctLonal relatLons 'source' and 'goal'5--the notion of source 
of ten being marked by ttte use of the prepositions ' (away) frolh' 
'out of', and goal being expressed by the prepositions ' to(wards) 
' in(to) ' , and 'on(to) ' 
The following six sentences give several examples showLng 
how marked ahd unmarked noun phrases can be reinterpreted as one 
of the f lve elementary functional relations defined aboye. 
(7) The object descended. 
Theme 
C8) John gave the obiec t to Mary, 
Source Theme Goal 
(9) Mary has the obiec t. 
LOCUS P heme 
5~heae functional relations have been used with great 
success by Gruber in his lexical inves tigations (1965) . We 
see the need to employ these notions as primitive functional 
relations in the type of gramar we have built. It is possible, 
however, that they should be treated as motional variants of 
locus. This further possible generalization is not utilized 
here. 
(10) Mary br&a the oblect 
Causal Theme 
Actant 
(11) The object is on the tabla. 
Theme Locus 
(12) - John carried the ,object from New York to Los AnpL%les 
Causal. 
Ac tant 
Theme Source Gas1 
The basic types of sentences that we w111 discuss shortly 
in terms of 'paradigms1 have, as a minimum requirement, one 
noun phrase functianing as a therbe; and at most, one occurrenc 
of any of the five functional relations described above. In 
other words, a simple or basic sentence type will not have two 
noun phrases functioning ag themes, or two noun phrases func- 
tioning as causal actants, e tc. 
However, English has evolved a number of derived sentence 
types where hro occurrences 0-5 a given functional relation are 
possLble. Consider the following sentences: 
(13) Hank broke the window. 
C.A. Theme 
(14) A rock broke the wir~dow. 
C.. A . Theme 
(15) The noise broke the window. 
C.A. Theme 
(16) Hank broke the window with a rock. 
C.A. Theme C.A. 
Direct Indirect 
In sentence$ 3) through (15) , the surf ace subjects , regard- 
less of semantic differences, ate functioning as causal actante 
in our analysie. 
It is only in sentence (16) where two causal 
ectante occur, that it becomes necessary to distinguish two 
kinds of causal actants: direct and indirect. 
6 
Whenever two caueal actanta occur in an active sentence, 
the unmarked one in subject position is the obligatorily anim- 
ate, d'lrect causal actant and the one marked by 'with' that 
occura later in the sentence is the inanimate tndirect causal 
ectant. In Fillmore (1968b) and other caee-type grammars the 
preposition 'with1 hae of ten been analyzed as indicating the 
inatrmental case--whl.ch we have reanalyzed as marking an ' in- 
direct causal actant. ' We feel that 'with' does not indicate 
any given functional relation consistently; rather one of the 
things Iwith' indicates is a derived sentence type having more 
than one occurrence of a given functional relation. In (16) 
'with' indicated en indirect celleal actant. In sentence (17), 
- 
C~illmore in his cage grammar system dis tingutshes between 
agentive. and ine trumental causal ac tants . For our purposes, the 
functional relation causal actant is sufficient--'agent1 and 
' ins trument ' being some thing close to, though not identical to, 
what we describe as direct and indirect causal actants respectively 
A given sentence may contain both types of causal actants--in 
which case the direct cauaal actant is subject. If only one is 
present, its direct or lndirect status is irrelevant; it is 
simply the causal actant subject of the sentence. 
1 8. 
'with' indicates a co-theme, and in (18) 'with' indicates a 
eo-causal actant: 
(17) John went to New York with Peter. 
~liziz 
I 
Goal 70- theme 
Primary 
(18) 
Bill played tennis with EIer 
~eusal T-Kzir dl 
Acttjnt 
Primary 
Note that the co-theme in sentences such as (17) and the 
co-causal actant in (18) must be semantically parallel to the 
primary theme or causal ac tant with respect tcr humanness or 
animacy, In sentences such as (19) where 'with marks a seman- 
tically non-parallel .theme, two sentences are always involved-- 
the surface object Eunc tioning simultaneously as the rheme of 
S1 and the locus of S2: 
(19) Mary saw the man with the cane 
Locus Theme, Theme Sq 
J- L 
s1 
+ Locus S2 
The fact that two sentences are involved in (19) is reinforced 
by a two-sentence paraphrase of sentences such as (20) first 
suggested by Lees (1960) : 
(20) Mary Jaw the man who has the cane. 
The error of assigning one semantic function to 'with' 
(or to most other prepositions for that matter) is further 
emphasized by the ambiguity inherent in sentences such as the 
fbllowing, which have been discussed in another context by 
Hall (1965): 
(21) The detective broke the window with the buralar. 
C.A. 
Primary 
Theme CO-causal/Second- 
acy C.A. 
A c t anst 
In one interpretation, (21) means the detective used thc bur- 
glar's body to break the window,' in which case, the 'burglar' 
la functioning as an indirect caueal actent. In the other in- 
terpre tation, (2 1) meana ' the detective and the burglar broke 
the window together, ' in which case ' the burglatt is a co- 
causal actant. 
Obviously, a good deal of semantic information about the 
verb as well as semantic information about the noun phrases 
involved i~ needed in order to specify the exact role of a 
nqun phrase following 'with. 
Before moving on to the next section of the paper, some 
cemarks are in nrder as to why we use only one category 'locus' 
for what ethers may dietinguish as 'dative' and 'locative' on 
grounds of differences in animacy. Consider the following 
sentences : 
(22) John haa a new car. 
- 
Locus Theme 
(23) New York has skyscrapers. 
Locus Theme 
(24) - John knows the answer. 
Locus - Theme 
Primary 
(25) The garden knows the answer. 
(*~r imary Locus) Theme 
20 
Sentences like (22) and (23) show us that there is a deep fnnt- 
I 
tional similarity bctwcen so-calltld 'dntivcs ' and lacat ivcs 
which justifies treating them ur~ifarmly 3s @Xt)ci in prccisclv 
--- 
I 
the wev that 'agents' and il~struents ' were seen tu function 
similarly and were uniformly labeled as causal - actants. Scntcnccs 
-- 
(24) and (25) illustrntc that some vcrhs requirt? a priruerv (i.e.. 
mental) locus as thc subjcct. This su~gcsts rhat tkrc is one 
I 
function.21 ca tegorv locus ' --hut that on occasion 1 t is neccs- 
sary to distinguish primary (mental) loci from secondary (phys- 
ical) loci. 
In this presentation of our basic and derived functional 
relations, we have tried to point out that there is no need to 
posit many different functional relations in order to dis- 
tinguish among the kinds of causal actants, themes, loci, etc. 
When and where features of animacy, purposiveness, etc., rJisc 
the need for dis tinguishing either primarv loci (animate) from 
other loci or direct causal actants (agents) from other causal 
actants, the apparatus is available in the form of lesical fea- 
tures of nouns, 
Classification of Verbs into Paradigms 
In several of Fillmore's papers (see 1968a and 1968b) it is 
I I 
suggested that each English verb be assigned a case frame" that 
would indicate what functional actants occurred obligatorily and 
optionally with a given verb. Translating from Fillmore's cases 
tpto our functional relations, this would mean tBPt the lexical 
entry of verbs like die, open, and - kill, would contain infor- 
mation much as 19 represented in terms of the case frames shown in 
(26) , (27) , and (28) re~pectively. 
(26) - die +[ ] thcsg 
r 
(C.A.) 
(27) +[ , ] theme 
(direct C. A. (inditec t 
c A=) ( 
-1 
(28) - kill +[ ] theme 
+ (indtrec 
C.A. 
Notation: 
{ 
) 
= disjunction 
( ) = optionalfty 
1 ] = case frame 
C.A. = causal actant 
What these lexical entries meaa is that - die occurs only 
with one obligatory ac tent ' theme' (29). The verb* open, like 
die, has an obligatory theme (30a) but also an optional causal 
- 
actant (30b),which may even be realized as direct causal actant 
plus an indirect causal actant (30c). The verb kill again has 
the obligatory theme, but, in this case, the causal actant is 
also obligator(y (31a) ; there is also the possibility of expres- 
sing the required causal. actant as a direct causal plus an 
indirect causal actant (31b). 
(29) John died. 
TEEiz 
(30) a. The door opened. 
7Sheme 
b. John opened the door 
C.A. Theme 
c. 
- John opened the door with e stick. 
direct theme indirect 
C.A. C,A* 
(31) a. - J~hn killed Roger 
C,A. theme 
b. 
Johh killed Rogor with a knife. 
- 
direct theme indirect 
C,A, C.A. 
While agreeing in spirit with Fillmore, we propose to take 
the , , case frame suggestion one step further and make it more use- 
ful in terms of grammar recognition. Since there are numerous 
verbs that behave exactly like -9 die many others that behave exact- 
ly like open, and a good T mber of others that behave like -# kill 
we propose to establish paradigms that summarize or recap4 tulate 
the functional felations and syntactic and semantic featdzes of 
large classes of verbs. Then the only feature needed in the 
lexicon would be a specification of which ~aradigm(s) a given 
verb sense belorged to. 
In other words, for verbs that are like - die. (i. e. , require 
a theme which occurs in surface subject position and do not per- 
mit a causal actant), we establish the intransitive paradigm. 
For recognftion purposes, the paradigmatic feature intransitive 
tells us that the surface subject is functioning as e theme, that 
B causal actant is impos~ible, and that any other actant must 
be marked by an appropriate preposition. 
The verb - kill, however,, is a membel' of. the transitive7 
paradim which requires. both a theme and a causal actant. 
7~ote the ~3i.f ference between our use of the terms - tran- 
sitive and in_tlan_e_itive and the traditional and generatL~e 
- 
uses. In traditlonal terms, ' transitivet means merely that 
a verb takes both a surface slibject and a surface object; in 
generative terms, ' transitive has meant. that a verb i capable 
of being passfvized. h nur system. transitive means that a 
verb takes a causal actant surface subject and theme surface 
object in the active voice. We stress this distinction because 
many of the verbs that allow. the passive transformation do not 
have causal actant subjects in the active voice (e.g., Mary 
claw the men/ the men were seen by Mary .) . This is why we stress 
that our use c>E the term transttive t:~ describe the behavior 
of a particuLar verb paradigm is more specific and functionally 
odented than other usages of the term. While it is true that 
all of our [+transitive] verbs may occur in the passive as well 
as the active voice, this fact is not e defining property of 
[+transitive] verbs in our system but a redundancy rule of sorts. 
Likewise, intransitive means not only that the verb takes a 
surface eubject and no unmarked surface object but that this 
surface subject is functioning as a theme. 
Furthermore, from the recognition point of view, we know that 
for all verbs marked [+transitive] tho surface subject is the 
causal actant and tho surface object is the thcmo iT tho ve'rh 
is in the active voice. 
A large class of verbs like open belohg to what we refer 
to as the ergative pbradigm. 
8 
'Ergative' is a surface srac- 
kre typological term long used by linguists to characterize the 
grammar of those languages that seem to assign the same syntactic 
role or case to both the subject NP of an intransitive verb and 
the object NP of a transitive verb. ('Ifgo languages typically 
described as ergative are Basque and Eskimo.) In the lidht of 
this definition, the behavior of open and the other two verbs in 
the examples below appear to be 'ergative': 
(32) a. John opened the door. 
b. The door opened. 
(33) a. The heat melted the ice. 
b. The ice melted, 
(34) a. The pressure increased the temperature of the air. 
b. The temperature af the air increased.. 
8 
We prefer this traditional term to Lakoff 's rwhet uncon- 
ventional use of the term 'inchoative' in his dissertation, - The 
Nature of, Syntactic Irregularity (1965), and elsewhere to de- 
scribe similar phenomena. 
What happens in thegs sentences is that if che causal actant 
subject of the (a) exemplee ie preeent, it is the surface subject; 
and the surface object, which mu8 t also occur, is the theme. 
If no causal actant is present, the surface crubjec't is the theme 
end there is no surface object. Also the verb ib an active, 
declarative English eentence mast, aa e rule, come second. 
Thus if there ie no causal actant present filling the eubject 
elot, the theme (or a noun phraae beating some other functional 
relation) must fill the subject alot. 
There ere &any ergative paradigm9 verbs in English. In 
par ticu'ar, the vecabulary of the meteoro,logy lessons that our 
project was centrally concerned with contains many verbs belong- 
ing to this ergative paradigm (e.g., move, (re)distribut@, re- 
duce, relate, replace, skew, slant, spread (throughout), s tbrt, 
a top, decrease, transfer (in) to, weaken, break(up) , boil, begin, 
dry (out), dissolve, decompose, divide (into), condense, form 
(out of), etc.) . 
9~orne readers will question the need for an ergative par& 
digm and suggest tbt each such verb be assigned to both the 
intransitive and transitive paradigms. However, both the 
economy gained by srating just one paradigmatic feature for verbs 
like open and the apparatus which then permits us to explain 
the relationship between raiselrose -9 kill/die a9 suppletive 
erga tive forms convinces us the t this paradigm feature is 
extremely useful if not necessary. 
26 
The above ergative paradigm verbs retain the same lexical 
shape whether or not a causal actant is prcscnt in suhj,ecr pn- 
sition. However, there are also o large numhcr of pairs of 
transitive-intransitive verbs that, by virtue oE their sement ic 
(and sometimes phonological or orthographical) s imilerity , should. 
perhaps bc cbsified as suppletivc members of an ergative. verb 
pair. 
The transitive membct of thc pair occurs vthcq 3 causal 
ac tant is present; the intrans it ive member sccufs when no causal 
actant is present. For examplet: 
(35) a. The pressure ra,.ised the temperature of the air. 
b. The temperature of the air-\rose. - 
(36) a. John killed the rabbit. 
b. The rabbit died. 
The full ergative paradigm of some verbs contains a stntive 
level as well as the two nonstative levels we have been discuss- 
ing : 
(37) a. Henry opened the doow. (-+€.A,, -state) 
b. The door opened. (-LA., -state) 
c. The door is open. (-C.A., +state) 
(38) a. John killed the rabbit. (Kd., -state) 
b. The rabbit died. (-C.A., -state) 
c . The rabbit is dead. (-C .A. , *ate) 
The functional relatiol .of the subject in the (c) sentences 
above is the same as in the (b) sentences--the differenoe between 
2 7 
them being thw statiw vs. the nonstative nature of the verb 
phrase. 
Thie d.iff$rence some-times has no effect on the surface 
forq of the verb in irregular cases such a9 (37c), but it is 
more generally the case that a different qurfa~~e form occurs 
dead of (384) -+he regularly different form being not a 
(e.Et., . 
1ex;icelly related adjective like - dead but a paec ~art~icipla 
(e.g., th,e ddor ie cLo8ed). The (a) Pentences above contain what 
Chafe (19703 oalls ,activity predicates, the (b) sentences process 
predicates, and ehe (c) sentences s tative predicates. 
Crriber (1965) has discussed transfer verbs at length, and 
based on his discussion we have found it desirable to establish 
three transfer paradigms. First of all, there are the two-way 
transfer verbs which allow ovest expression of both a source and 
a goal: 
\ trans ferked 
Secondly, there are source-subject transfer paradigms which per- 
mit goals but not sources to be overtly stated in the predicate 
since the subject is the source of the transfer. 
to Mar . 
d 
Thirdly, we also have the goal-subj ec t trans fes paradigm which 
permits sources hut not goals to be directly stated in the 
predicate, and the subject is the goal of the transfer since s 
from phrese is implicitly or explicitly e-reesed. 
8 watch from Jane. 
-Tnezno Source 
Somewhat similar to the transfer paradigm is the one-way 
putltake paradigm that includes additive verbs like put in(to)/ 
on(to) (42) and privative verbs such as take out, of /off 
took 
Source 
In this paradigm virtually all of the verbs can be used in con- 
junction with either goals or sources--denoting the theme's corning 
to or going from a locus respectively. 
The preposition makes 
clear which case is intended. A few of these verbs even permit 
the expression of both source and goal in the predicate, and 
these verbs form a separate two-way putjtake paradigm (44): 
(44) 
Mavis out of the house tnta the garden. 
pushed Theme Source &a1 
There is another paradigm for verbs like 'have' which re- 
qutre a locus as surface subject and a theme in object position 
(45) : 
c.5) gci; ) a f;cgtune 
posse e8 erne 
A paradigm for the converse of (45) includes verbs such as 
I 
occupy which have surface subject themes and take loci in 
surface ob j ect position (46) : 
Theme Locus 
There are several other possible paradigms which will not be 
mentioned and illustrated hece. The point we want to make is 
that large numbers of verb senses can be classified in a way 
that facilitates recognition: the semantic function of surface 
subject3 and objects of verbs can be easily ascertained by ref - 
erence to a paradigmatic feature on the verb. 
Paiadigms and Trangforma t ions 
The paradQms discussed above interact with a nusber of 
movement and deJetion transformations that should also be die- 
cussed, 
Firstly, a number of English verbs permit a transforma- 
tional deletion in the surface structure of e lexically un- 
specified yet semantically delimited surface object theme. For 
exanple, a verb such as - eat may occur either with a lexically 
specified object theme (47) or with a lexically unspecified 
one (48) 
(47) Bill ate a sandwich. 
c?r Theme 
(48) Bill ate. a 
CK Theme 
(= food/something edible) 
1 
We describe sentences such as (48) as having undergone an ec- 
cusative deletion. The deleted indefinite theme must be 
lO~he term 'accusativef has been usea by language typologists 
to refer to languages that treat the subject of an intransitive 
verb and the subject of a transitive verb with the same inflec- 
tion. Also, im languages commonly described in terms of case 
grammar the accus,ati.ve case generally corresponds closely to 
our notion of a surface object in English. We have borrowed 
this term and are using it to specify a particular type of 
deletion that occurs. in English. 
reconstructed if all the functional relations expressed in the 
sentence are to be explicitly staged in the deep structure. 
Other examples of verbs with potential accusative deletion in 
the data are: conduct, evaluate, gain, lose, observe, read, re- 
call, under8 tand, consume, etc . 
Another type of transformational deletion occurs with vir- 
tually all motional verbs such as -9 run move -9 - stir, walk, - jump, 
etc., whenever the surface subject of such verb is [+ animate] . 
Consider the following senteqces : 
(49) a. The mechine ran. 
Theme 
b. The door moved. 
Theme 
(50) a. John ran. 
Theme and -C.A. 
b. The cat moved 
'heme and C,A 
The sentences in (49) which have inanimate surface subjects con- 
form perfectly to the intransitive variant of the ergative para- 
digm discussed previously ( . , the surface subject functions 
as the Theme) The sentences in (50), however, are somehow dif- 
ferent in that the surface subject is not only functioning as a 
theme but also as a causal actant since no external causal ac- 
tant has been explicitly mentioned. We can roughly paraphrase 
this type of sentence as follows: 
(51) a. John cawed himself to run. 
b. The cat caused itself to move. 
Since such paraphrases are not possib'le for the sentences in 
32 
(491, we assume that a reflexive' deletion takes place in sen- 
tences with transitive-paradigm verbs of motion or verbs of 
self-oriented habit (e.g., wash, dress, shave, etc.) When an 
animate surface subject functions both as the causal actant and 
the theme of the [+transitive] variant of an ergative verb; then 
the coreferential theme/causal actant Fs not stated b~icc in the 
surface structure if tho vorb may undergo reflexive deletion. 11 
The accusative and reflexive delet,ions discussed above help 
to explain an interesting ambiguity that occurs in sentences 
such as (52) : 
(52,) Mary washed. 
which may have either of the interpretations in (52') d : 
(52 I) a. Nary washed (the laundry) 
b. Mary washed (herself). 
ll~his is different, of course, from what happens to verbs 
that permit overt reflexivization. However, these no~d@letable 
reflexives tend not to include verbs of motion or verbs af self- 
oriented habit. 
(i) John cut himself. 
(ii) Mary wore herself out. 
(iii) The dog prdte'cted himself. 
Since the transitive paradigm verb wash will be marked in the 
lexicon as additionally permitting either the accusative dele- 
tion or the reflexive deletion, the two ambiguous readings OX 
(51) given in (52) will be automatically predicted by the re- 
cognition grammar since the accusative deletion rule will re- 
copstruc t (52a) and the reflexive deletion rule (52b) . 
In addition to deletions, the grammar recognizes and effec- 
tively ' reversesT various movement transformations or permutations 
that may have operated to produce the surface structure of cer- 
tain sentences. 
A well-known movement transformatibn is the passive trans- 
formagion which operates on a sentence such as (53) and produces 
(54) : 
(53) Dick purchased the car yesterday.. 
(54) 
The car was purchased yesterday by Dick. 
A subsequent transformation may optionally delete the 'by' phrase 
in (54) yielding (55). 
(55) 
The car was purchased yesterday. 
Given a sentence such as (54) our rules will 'reverse' the pas- 
sive transformation and yield (53). Given a sentence such as 
(55) our rules will reverse the passive transformation and pro- 
duce aq active structure similar to (56). 
(56) 
Someoae purchased the car yesterday . 
The procedure sews simple enough yet th~re are many prob- 
lems involved in che accurate recognition and rovcrsel of p@s- 
sive sentences First of all, not all 'by' phrases co-occurring 
with a passive verb call be reanalyzed as the surface subject of 
the corresponding active sentence: 
a. by John. 
(57) The rations were increased by the sea. 
by noon. 
by 50 
Of the above four 'by' phrases, only (a) could function as sur- 
face subject in the- active voice version of (57) ; semantic 
features are needed to determine whether the noun object in a 
'by' phrase is a causal actant, a location, a time, or a neasure- 
ment. In our lexicon John is [+potent] , - sea is [+,s tationaryl , 
noon is and percent is 
fe-atures will allow 
our passive reversal rule to construct a reasonable active voice 
var tant for all the passive sentences in (57) . 
Another problem in passive reversal concerns the verbs in 
the various transfer paradigms. For several of these verbs most 
speakers of English recognize two different passive permutations 
(e., (59) and (60)) of the same active sentence (58): 
Active (58) - John gave Mar the book. 
Source 'd Theme 
- 
12~ny unit preceded by a cardinal number gets reanalyzed as 
a measure phrasef 
35 
Passive (59) Mary was aven tHe book (by. John). 
(60) The book was given (to) Mary (by John). 
In such cases. semantic information about the surface subjects 
and surface objects of the passive sentences is required if the 
'theme' and the 'goal' are to be properly labered in the deep 
structure. In such caees tho following feature hierarchy--or 
something like it-+seems to operate: 
human 
animate 
(6 1) s ta tionary 
concrete 
abstract 
The two nouns involved as ' goal ' and ' theme ' seldom are at the 
same level on the hierarchy in terms of their lexical features, 
and the one that is higher thhn the other is always the goal. 
Thus in (59) and (60) --irrespective of the surface order in the 
passive--'Mgryl is analyzed as peal and I the book as themo . 
Some readers no doubt may wonder why we have bothered to 
write rules that depassivize13 sentences. The answer is that 
we wish to get optimum efficiency and accuracy out of our para- 
digms without indefinitely prolif @rating their number. The 
paradigms are set up to assign functional relations to the 
marked and unmarked noufl phrases occurring in active declarative 
structures. By first reversing the results of deleti.on trans- 
format ions, question trans£ ormations and other movement trans- 
formations 
(e.g. passive), we are able to use a minimum nun\bes 
of paradigms to assign accurate functional labels to all of the 
noun phrases in a sentence. 
The paradigms for transfer verbs just discussed above with 
respect to the passive transformation are also involved in 
another movement transformation which we refer to 8s 'goal focus' 
13we are for the moment ignoring potential ambiguities and 
changes of meaning caused by changes of quantiFier ordering in 
active and passive sentences such as those which Chossky has 
pointed out (1965) . 
e.g. (I) Everyone in the room speaks two languages. 
(can be Interpreted as being different languages 
for each person) 
(2) Two languages are spoken by everyone in the room. 
etends to be interpreted as two specific languages 
that everyone speaks) 
To explain thia transformation be must first establish the 
difference between iumediate source8 and goals on the one hand, 
and remote sources and goals on the other. With verbs of trans- 
fer ip the active voice, the surface subject is a source, goal, 
or causal actant, The theme is the surface object. Other 
aspects of the motLon or transfer are usually implied. The 
following examples illustrate this : 
(62) Johd sold the house. 
7 
(Implied: to some 'Goal ' ) 
7 
Source meme 
(63') 
Harry bought a dune buggy. 
zeal Theme 
(64) Sam brought the beer. 
~azl Theme 
Ac tant 
(Implied: from some 'Source') 
(Implied: from so- 'Source' 
and to some 'Goal') 
Verbs like 'buy' (e.g., get, obtain, acquire, fetch, borrow, 
etc,) and verbs like 'sell' (e.g., teach, serve, donate, give, 
lend, etc.) have an ilmnedtate goal and an immediate source in 
subject poaition respectively. In addition, they may exgrass 
a remote goal or a remote source, bvth of which get marked with 
the preposition "for " The following examples show that, if 
I1 
the subject embodies the function source," the optional "for" 
phrase will also embody (remote-) the function of source. If 
the subject expresses the function "goal, then the optional 
"for" phrase may express (remotely) elther the funotion "goal" 
11 
or the function source. 
11 
(65) John sold the-house far Harr , 
1m7 meme dur c e 
source 
(66) Harr bought a dune b 
-8 
for his son 
1-e Remote goal 
goal 
- 
or 
Remote source 
Transfer verbs like 'bring1 that oxpress B causal actant 
in subject position and LqpLy both source and goal are also 
It 
ambiguous when occurring with an optional f or1' phrase unless 
it is clear from the contest that the causal actant is func- 
tioning on behalf of a lremote source or for the benefit of a 
remote goal. 
(67) Sam brought th beer for us. 
c a= T#ZKE- EZFE goal 
ac tant 
- 
or 
Remote source. 
In the above sentence "for us" can meenl'for our use or benefit, 
I I 
in which case "for us" represents a remote goal. It, also can 
mean "at our request, acting on our behalf," in which case "for 
US" represents a remote source. 
One way of disambiguating sentences like (67) is to apply 
'Goal Focus' movement (68) whenever the remote goal sense of 
"for NP" is the sense intended. 
&8) Sam brought us the beer. 
C aux GO= Theme 
'Goal Focus' can move either a goal (70) or a remote goal that 
immediately follows the theme (69) to a p~sition of focus 
between the verb and the theme. 
14 
(69) a. John bought the house for Mar 
1iZiia -a 
goal or Source 
b. John bought Mary the house. 
I-. Remote Theme 
goal goal 
(70) a. John gave the book to Mar . 
- 
Source meme + 
John gave Mar the bo&. 
b. - 
Source d Theme 
Whenever god. L,KJS movement occurs, it seems that the remote 
goal loses its preposition end becomes a goal object more in- 
timately associated witti the verb than it had been as a 
140ne apparent condition on "Goal Focus" movement is that 
th theme may not be a pronoun (the goal may be either e noun 
or pronoun). In other words, it can not occur with sentences 
auch ea the following: 
a. John gave it to 
fR:l] 
b. *John gave it. 
However the possibility (in fact the grarnmaticality of utter- 
ances such aa "ginme it!" (%. em, Give me it) renders this con- 
dition dubious. 
prepositional object. 15 
Our grammar makes use of several other deletion and mve- 
ment reversiqg tr'ansf~rm~ationa that will not be discussed horc. 
What we have tried to illustrate in this section of the paper 
is that both paradigms and transformations arc useful componen to 
in a recognition gramar. 
l%e have reascn to believe that an obj eo tivalization move- 
ment rule like Focus' in Eng1,ish is also widely used in 
other languages. French, for example, appears to have a move- 
ment operation even more general than the goal-res tr ic ted 
objec tivalization transformation of English. It is for this 
reason that speakers of French say sentences such as "open me 
the door" when speaking English. English does not permit the 
movement of "me" to object position in this sentence because 
"me" is a remote causal actant or source and not a remote goal 
In "Open the door for me. " In this sentence "(you)" is the 
11 
immediate causal actant and source and the verb open" belongs 
to the ergatiwe paradigm and not to the movement-transfer class 
of verbs. Thus the rules oE English do not permit "~oal FOCUS" 
movement to apply in such a case. 
In the course of the development of this recognition gram- 
mr we noted that ccrmp1e.tive aspect (.. inference, of com- 
pletion ve. no inference of completion) sometimes serves to 
di$ t inguish highly similar verb paradigms% 
In such cases, both 
the lexical ahape of the verbs end the prepoeitione, as well as 
the order of ectanta in the surface structure, play a role in 
signaling the presence of completive aspect. Recall the para- 
digm for verbs of putting and taking (71) discussed earlier in 
the paper: 
(71) a. John 
put 
water in the pool. 
lZ7X7 [-comp etive1-e Goal 
took the groceries out of the bag. 
be W [-completive~ ~hemb Source 
Note that the verbs of the sentences in (71) are not marked 
as having completive aspect: one cannot infer that the pool in 
(71a) ie full of water nor that 'the bag' in (7Lb) is empty of 
groceries. There is, however, another paradigm similar to the 
put-take' paradigm but different from It in that all of ite 
verbs indicate completive aspect. We refer to this [+completive] 
paradigm aa the 'join-separate' paradigm (72.), which is the para- 
dim to which the verbs fill and empty, for example, belong. 
(72) a. Jch filled che pool with water. 
F.x [+completive] Locus Theme 
emptted 
[hompletive] 
the ba of roceries, 
d* 
42 
The sentences. in (72) --unlike those in (71) --do allow us 
to infer that ' tho pool' in (72a) is full and that ' the bag' in 
(72b) is empty This is boceuse.varbs of joining--which mart 
their theme with the preposition '~Fth.'--end verbs of separating- 
which mark their theme with 'of1--are elway8 [+crmplctioc] 
Other examples of j6ining verbs are: fill, supply, provide, 
s tock, cover, present, furnish, plant, smear, sprinkle, e tc . 
Additional examples of separating verbs are: relieve, empty, 
rob, deprive, withhold, rid, clear, drain, deplete, etc. 
Some English verbs--= well as being used to express the 
completive notions of joining and separation--may also be uged 
to express nm-completive activities such as putting or taking. 
This dual: function of the verbs has been the source of much 
confusion and unsatisfactory analysis, and is a topic we should, 
therafose, like to pursue in some detail. In Rall (19165), for 
example, the following sentences were cons-idered more or less 
equivalent and thus relatable vi.a a transformational rule which 
considered the objec-c paint in (73) as basic and the object 
'wall' in (74) as derived: 
(73) John smeared paint on the wall. 
(74) John srneare'd the wall with paint. 
Fillmore (1968b, p. 48) e-xplained these sentences by 
suggesting that both 'wall' and 'pint' were originally supplied 
with prepositions reflecting the lorcatisre and instrumental cases 
43 
respectively. This yields 'on the Gall' and 'wlth paint' in the 
deep structure. Fillmore then analyzed the verb - smear as having 
the following property: whichever of the ttwo deep structure 
elements concerned is chosen as the 'direct object', it must 
fall next to the verb and must lose its preposition. 
While more ar less agreeing with Hall's intuftion that 
'paint' is a basic 'objectv--in our system 'theme1--Ln (73) and 
a derived'object' or displaced theme in (74), we disagree with 
her implicit assumption that sentences such as (73) and (74) are 
syntactically or semantically equivalent or the explicit pro- 
posal that they should be related by a transformattional rule. 
We also disagree with Fillmore's decision to analyze 'with' in 
sentences such as (74) as an ins fance of the instrumental case. 
In our view, aspec tual differences between (73) and (74) call 
for a different assignment of Jeep structure functional relations 
and necessitate the postulation of distinct paradigms. In our 
analysis the surface object of (73) is a theme; the verb "smear" 
expresses an activity similar to "p~tt~ing" in (73) and ' the wall 
ie the goal of the theme 'paint'. In (74) 'smear' is being used 
to express the completed joining of a displaced theme--i. e. , 
'paint1-- to the locus--i.e. , 'wall'. In other words, our SOB- 
ution will analyze verbs such as 'smear '. as belonging to two 
1 
dirfererrt paradigms.. In one of the paradigms, smear' behaves 
like 'putr and in the other it expresfses joining af a location 
and a theme. In this latter sense, the displaced theme is 
expressed optionally, though always implied as something present 
but unspecified if not overtly expressed: 
(75) a. John smeared the wall with paint. 
b. John smeared the wall. (Implicit: with something) 
(76) a John filled the glass with water. 
b.. John filled the glass. (Implicit: with scitnething) 
In the non-complbtivc sense of 'smear' where 'well' functions as 
goal, this 'goal' element is obligatorily expressed and not 
deletable, 
ran his brother 1 
(77) a. John smeared paint on the wall 
b. *John smeared paint. 
Z 
( in the box 7 
(78) a. John put the bohk j (over) thOre l 
b *John put the book. 
Verbs of joinibg and separating are not the only ones 
belonging. to a paradigm that signals completive aspect. 
There are 
two paradigms for verbs of contact that also can be used as 
examples. First consider the non-completive contact paradigm 
stones at the wall. 
kicked Goal 
In sentences auch as those in (79), we cannot infer that the 
theme made contact with the goal; we assume the theme moved 
toward the goal, but we do not know whether contact was -made. 
Contrast theee sentences with verbs occurring bn the completive 
contact paradigm (80) : 
(80) John 
{:: } 
the wall with stones. 
K?r ombarded Locus Theme 
(i-comple tive) 
These sentences in (80) force us to infer that the theme has 
made contact with the locus; no other interpretation is possible. 
In the two sets of paradigms discussed above, some unusual 
phenomena have taken place. In those cases where the verb 
signals completive action, the surface order of ac tants in 
active surfaces is as follows: 
Surface Surface 
(81) Subject Object Marked NP 
+ W? +I NP 
[tcomple tive] 
1 + prelw 
C.A. Locus Theme 
46 
This is not ,the normal order of elements. The usual surface 
order for active sentences containing B cau~a1 actant, a theme, 
and a locus is the following: 
Surf ace Surface 
Subject 
& 
Obj ec t 
V 
Theme 
Prep 9 NP 
Thus is appears that we get completive aspect only when a locus 
occurs in the position where one wauld normally expect the theme 
to occur. 
This hypothesis is confirmed by a completive gal~digrn with 
no unmarked surface objects that has a locus in subject position 
instead of a theme (83). This paradigm is similar to the non- 
completive paradigm which has the theme in subject position 
(84) and foll~ws the normal order of constituents for sentences 
having only a theme and a locative actant: 
(83) The arden is swarming with bees. 
* [+co13pletive] Theme 
are swarming in the garden. 
T erne [-completive] Ldcus (84) i? 
The meanings of these sentences have been debated by Fill- 
more (1968b) and Chomsky (1972), among others. Fillmore assigns 
the same case representation to (83) and (@+) sugges tins that 
there is a difference of focus with perhaps some corresponding 
difference in cognitive content Chonsky points out that the 
sentence8 are not synonymous--that in (83) the bees are neces; 
earily all over the garden but that in (84) the bees might be 
only around their hive. In light of our analyses of sentences 
(71) - (80) we suggest that both Fillmore' e remarks and Chomaky ' 8 
remark8 are correct but indomplete. The functional relations 
exhibited by the noun phrases in (83) and (84) are the same, 
the verb aspect ie different. This auggeets that there are two 
paradigme involved: a completive paradigm accounting for (83) and 
(85) , and a non-completive paradigm accounting for sentence (84) 
and similar eentences like (86) : 
Theme + verb + Locus 
- 
[ -completive] 
(85) a. Passengers are riding in the bus. 
b. 
Fish swim fn the stream. 
c. Groceries were in the bag. 
Locus + verb + Theme 
[+completive] 
(86) a. The bus is sagging with passengers. 
b. The stream teems with fish. 
c . The bag bulged with groceries. 
With lexically different verbs and prepositions as in (85) 
qnd (86) the two paradigms are eaeily distinguishable. When 
there is lexical overlap in different verb penses as in (83) and 
(84), the prepositions as well as semantic features will serve 
to distinguish [+completive] paradigms from [ -completive] paradigms 
Paradigms and xncotporation 
In the second section ~f Gruber (1965: 5-27) there fe a 
discussion of the gramnatical process of inc~rporation. The 
term incorporation has been used most frequently by 1ingui.s ts 
working on American Indian languages; however, (Sruber's discus- 
sion of incorporation along with the following discussion of cer- 
tain ~aradigma supports the position that the concept of incor- 
poration is needed to describe in full generality certain lexical 
and grammatical facts about English. 
Gruber discusses verbs that obligatorily or optionally in- 
corporate certain adver-bs or pre~ ositions . For example, Gruber 
claims that the verb cross obligatorily incorporates the adverb 
across. He illus trates this with paraphrase relations in accept 
able sentences such as the following: 
(87) John crossed the stteet. 
(88) John went across the street. 
Note also that the sentence below is impossible: 
(£39) *John crossed across the street. 
The lexical and grammatical facts cannot be explained with maxi- 
mum generality unless the verb cross is analyzed as a motional 
verb similar to a yet also incorporating Me adverb across. 
Another example of incorporation that Gruber provides con- 
cerns ve~bs co-occurring with the preposition - for. The verb 
want _. obligatorily requires and incorporates -9 fbr while thd verb 
yearn requires but never incorporates - for. 
The verb wish, how- 
ever, requires - for end allows both poesibilitiee (is incor- 
poration of - for is optional) . 
Gruber 's examples illustrate this : 
(90) a. John wants a book. 
b, *John wants for a book; 
c. *John yearns a book. 
d. John yearns for a book. 
e. John wishes a book. 
d. John wished for a book. 
These verb-preposition co-occurrences would be very difficult to 
explain economically without making refereace to the notion of 
incorporation. 
Likewise, in setting up the paradigms needed to account 
for a large number of verbs in English we realized 
that certain 
verbe could be best described as belonging to a particular para- 
digm but, in addition, incorporating a specific noun actant. 
For example, the paradigm describing source-subj ec t transfer 
verbe like ~ive, -S send sell, etc. can also include verbs like 
help, aid, Bupp 
- ort, etc. if we recognize that such verbs incor- 
porate the theme but are otherwise exactly like the source- 
subj ec t transfer paradigm: 
Normal 
Source (91) - John the car to Mar 
Transfer Source sent Tlizird 
Goal 
Focus 
(92) - John gave Mar the car 
Source d- 
Theme - # - 
incorporating %elpad 
Source (93) John 
Wansf es Source supported 
The fact that these incorporating verbs can be paraphrased as 
follows further reinforces this analysis. 
(94) - John gave to blsrv 
S ounce %T 
Goal Focua 
(95) John gave Mary 
Source 
{ %ip } 
support 
The theme is not the3 only actant that may be incorporated. 
Another such example of incorporation is found in verbs such as 
%, chart, graph, stratify, etc,. which resemble two-way trans- 
fer verbs like bring, trans.fer, take etc. except that the goal 
actant has been incorporatad in the verb Thus we have the 
regular two-way transfer paradigm: 
e 
(96) John 
and th'e goal-incorporating version of the two-way t~ansfer 
paradigm: 
(97) John the statistics. 
Theme 
This analysis is again reinforced by the possibility of para- 
phrasing the goal-incorporating two-way transfer verbs as follows: 
(98) Jdhn transferred the s tatis tics onto the 
causal Theme 
Ac tant 
Goal 
Once it is recognized that part of the lexical description 
of a verb may .Include the fact that it belongs to a particular 
paradigm but also that it incorporates a particular functional 
relation more typically expressed by a surface noun, the para- 
digms will apply to many additional verbs, and thus many un- 
usual phenomena can be explained more consistently and adequately. 
Consider the verb 'surface' as it occl3rs in the two follow- 
ing sentences--which on a superficial level appear grammatically 
similar : 
(99) Phil surfaced the treasure chest. 
(100) The men surfaced the sidestreet. 
In eaob of these sentences the verb has incorporated the 
noun 'surface' as an actant; however, in (99) the goal actant 
has been incorporated and in (100) the theme actant has. Thus 
the verb 'surface' is being used in two very different ways 
The following paraphrases of the abave sentences will help 
clarify this. 
(99') Phil brought the treasure chest to the surface 
C v Theme Goal 
v. 
mrr -*# 0 4 
is 
Without the notion of incorporation such pairs of sentences 
would pose sevcra dif ficul tics for our paradigmatic analysis 
of verbs. Given the notion of incorporation as a working hypo- 
thesis, however, virtually every verb in English will hc ex- 
plainable in terms of a finite set of paradigms. 
One final axample of incotputation that wc would like to 
discuss has to do specifically with verbs of precipitation. 
ining 
owing 
izzlin 
eeting 
Such sentences are unusual because the surface subject is a 
lexicallv empty dummy elcment it, so it appears that this 
sentence-type has no real theme. This could be somewhat dis tres- 
sing since we agree with Gr'uhcr that a predicating element and a 
theme are the minimal and obligatory elehlents in every sentence. 
Again, the notion of incorporation proves to be useful. If we 
analyze the verb in the above sentences ,as being very much like 
the minimal verb "fall" yet incorporating the theme, a plausible 
a'nalysis is achieved in terms of both patapbrasability and para- 
digm as s ignmen t . 
(102) - It is raining. 
b" 
dummy +Theme 
subject 
(103) Rein is fallinp, 
TliEz 
- 
The related historical fact is that English sentences re- 
quire the verb in second position; thus verbs that for some 
reason (e.g., incorporation) are without an overt nominative 
subject to function as theme have come to require an empty it' 
(referred to as the impersonal or expletive - it) in subject 
position. 
From the Surface Structure of a Sim~le Sentence to its Deer, 
Functional Structure 
In the recognition grammar that we have developed, the 
first step in making a correct deep structure analysis of a 
simple sentence is the surface grammar's parsing of the noun 
phrase (8) , prepositional phrase(s) (i . e , marked noun phrases), 
the finite verb (phrase), and other possible surface structure 
constituents such as conjunctions and adjectives or adverb 
phrases. 
At this point, it musf be established whether the form of 
the verb bhrase) is active or passive. The next step consists 
of looking up the surface verb in the lexicon to determine 
(i) what paradigm class it belongs to and (ii) what movement and 
deletion operations may apply to the verb. 
Next, it ie possible, and in fact very simpla, to taks 
each unmarked noun phrase and to ascertain Nherhor its surfacc 
structure role is subject or object and to mark the noun phrases 
ancordingly . 
Finally, by using the lexical information, a set of 
hnuristics is consulted to assign the proper deep structure 
functional relation to each noun phrase occurring in the sen- 
tence. On= this has been done, rules of interpretation or 
"understanding1' my be applied to the deep str8ucture which has 
been recons eructed by the recognition gramar . 
We shall now apply all these rules and strategies (except 
for the rules of interpretation) in an admittedly oversimplified 
fom to a sentence in order to demonstrate the sequence of the 
procedure, The simple sentence we shall consider is the following: 
The hea~ slowlv evaporated the water. 
(104) Step I: Apply surface grammar rules and parse the sur.-= 
face s truc ture into labeled constituents . 
The heat + slowly + evaporated + the water 
NP ADV v(p> NP 
Step 11: Assign a surface role to each NP not preceded 
by preposition. 
The heat + slowly + evaporated + the water 
SUBJECT OBJECT 
Step 111: Determine the form of the finite verb (phrase) 
active, (simple past) 
Step IV: Look up the verb 'evaporate' in the lexicon 
for paradigm classification and its features 
of movement, dele tion, and incorporation. 
er~a tive 
Step V: Look up the predetermined heuristics for surface 
sentences with active-ergative deletion verbs. 
Heuristic : For sentences with active, ergatf ve- 
paradigm verbs : 
(i) If there is a subject and an object, the 
subject is a causal actant and the object 
is the theme, 
(ii) Pf there is a subject but no object, the 
subject is the theme. (The stative form 
of an ergative- transitive verb always uses 
this s tra tegy .) 
Step VI: Apply appropriate heuristic of Step V to surface 
information de tennined during Steps 11-IV and 
transform the surface s tructure into the 
appropriate deep s truc ture . 
(105) a. Surface structure: 
The heat -+ slowly + evaporatkd + the water 
NP ADV 
v(p) 
WP 
SUBJECT ACTIVE OBJECT 
ERGATIVE 
[ +TRANS ITIVE 
$lQ%?k 
CAUSAL 
ACTANT 
/ 
the heat the wateq 
7 
Thus the recogni,tfon system is conceptually complete as far 
as simple sentences are concerned, and col.nuplex sentences cm also 
be handled, given oertain modifications that will be discussed 
briefly in the following sec tions . 
Extension of the System to Cowlex Sentences 
It would be impossible to review all our procedures con- 
cerning complex sentence types in English; thus we shall merely 
exemplify our techniques f acus ing mainly on inf ini tival--and 
marginally on gerundive-gsentential cdements in the process. 
Embedded sentences typjcally contain a number of optional 
and obligatory surface deletions Which must be recovered. Re- 
cons truc tions involve those sub j ec tless predicates that are a 
result of deletion transformations taking place as part of the 
embedding process. For the sake of comparison, we have provided 
in (106) wtences which have undergone subject deletions or 
movements of this type, and sentences in (107) which have not. 
(LOG) a. Sid wants to go there, 
Sub j ec tless predicate 
b. To ~6 there would be unwise. 
Subjec tless predicate 
C. Camping is en j oyable . 
Subjectless predicate 
d, The children started playing. 
Subj ectless predicate 
(107) a. Sid waited fox John to come. 
embedded 
subject 
b. It would be unwise for Eric to go there 
embedded 
subject 
c. Walter's insulting us came as no surprise. 
embedded 
gubject 
d. The mother regretted her son's having stolen the money. 
embedded 
subject 
In order to capture the fact that the infinitival and 
gerundive verb phrases in (106) have embedded subjects that have 
been deleted, moved or left lexically unspecified, our grammar 
makes use of the following highly general rules: 
(108) a. Insert a 'for + A ' subject before any infinitival 
phrase not already preceded by a' (for) + NP' subject. 
b. Insert a ' subject before any gerundive (i.e. -ing) 
verb phrase not already preceded by either 'NPt or 'NP' 
[+ possl 
Let us consider spec-if ically four types of 'for -st embedding 
situations that move or delete subjects. 
Equi sub1 ec t-subj ec t dele tion: 
Equi obi ec t-aubj ec t deletion: 
('lo) 
John told B 1 
--+ John told Bill to go. 
I 
Subject-to-subject raising: 
John begalif to run. 
A 
John run 
Sub 1 ec t-to-objec t raisin&: 
Mary wants NP .-p Mary wants Jahn to go. 
I 
Note that in all of the above cases, when the lower sentencu 
loses its subject through either movement or deletion, the 
lower predicate becomes infinittvalized. l6 
-om the txmognition poht of view these movements and 
deletions mug t be reversed if we wish to reconstruct the subject 
wd ptedicate of lower sentences accurately. The first step in 
our procedure is the application of the rule stated in (108b) 
since rhas will identify subjec tless predicates, thas permitting 
the rules to build tentative deep s eruc tures . 
(113) John wants to go: -) John wants for 2 to go 
l7 _) 
tentative deep 
syntactic structure: 
John wants NP 
1 
for -S, 
161t should also be pointed out that gerundivization as we1 
as infinitivalization can result from either of these operations 
(em&, John began running; I like playiag volleyball, etc.) 
17we are tentatively inserting a deleted ' for a ' in all 
such cases ; however, this is a s implication which may eventually 
turn out to be infelicitous (em in some cases it may not be 
necessary to insert a 'for'). 
(114) 
John told Bill to go.-) John. told Bill for A to go.+ 
tentative deep 
syntactic structure: CIp 'S %-*\ 
JO~A - tbld-~iir~p 
for -S 
/ 
L ---^-I 
A go 
(115) 
John began to run. -+ John began for iA to run.. . ...--+ 
tentative deep as\ 
--- ,, --, 
syntactic erructure: .-. 
John began NP 
I 
for -S 
/ ' 
*--- 3 
A run 
(116) 
mry wants John to go.-) Mary wants John for A to go. + 
tentative deep 
syntactic structure:/ 
Mary wants John NP 
I 
for -S 
\ 
LL-LL 
The next thing we must do is classify verbs in such a way 
that we can correctly ftll the deltw in all of the above strut.- 
tures. Thus instead of describing a verb as a 'subject-subjec t 
deletion' verb (log), we will describe it as a 'copy subject' 
verb - (ll3h so we know that the higher sub j ec t must be copied 
onto the delta in the embedded sentence in order for the syntactic 
deep s'ttucture to be complete and accurate. The following table 
shows how the features in our recognition grammar correspond to 
the features that might be used in a generative grammar. 
- --... -- --- --- & - - *I - - -- --- ---- _L 
7 Recognition 
Feature 
I 
--- J--- Feature I 
/ 0b j ec t-subj bc t deletion 
Subjec t-to-s'ubJect 
raising 
I ,I, 
; hope, ask... 
I ! 
Copy Object 1 I tell, order, 
I 
i 
comand.. . I 
Y 
Drop Subject began', con- 
I tinue, tend, 
B 
seem... 
i I* 
Drop Object I want, expect, 
I 
~9ish, , , 
Once these features have been assigned to the appropriate verbs, 
the tentative deep syntactic structures can be finalized. 
There is, however, one trans formation which must take place 
before the analysis of embeddings begins, and this is the 'pas- 
sive-to-activ.el transformation. Thus a sentence such as (117) 
is first changed to (118) before the embedding procedures are 
effected, 
(117) John was told to leave. 
(118) A told John to leave. 
If this transformation were not carried out, all features of 
ob j ec t-embedding- verbs would have to be re-analyzed in the 
passive as follows: 
- - 
1 
1 4 
Active Voice i Passive V~ice 
. ---. 4 +- - - -I-- 
I 
Copy Object I Copy Subject 
i 
Drop Obj ect I 
1 
Drop Subjecr: 
- ----- - - -.-----I--- ---- - --- -- - - - - - ---.--..----. ..- . - s 
63 
While it would not be particularly difficult to change these 
epecificatione, it seeme quite unnecessary in light of the Pact 
that de-paasivization must, in any caee, be done before the 
appropriate paradigm heuristic applies; thus it can easily be 
carried out before embedding procedures apply as well. Para- 
digm analysis then begine with the lowest (i.e. moat embedded) 
sentence in the tree and m'ovvo up with the result that an 
entire embedded sentence (via an intermeaiate NP node) is 
functioning as theme or causal actant in a higher sentence. 
The procedures outlined above for dealing with certain 
types of complex seritences--however partial and sketchy-- do 
indicate that it is possible to extend the verb paradigm re- 
cognition technique to complex sentences, i.e., its use is 
not limited to sFmple sentences. 
Concluding S tatemen t 
Although many details have been omitted from this paper, 
we believe that such an approach is as linguistically and psy- 
chologically valid as other theoretical or pragmatic recognition 
procedures developed to date. Innovations in linguistics end 
psycholinguis tics such as Bever and Langendoen's use of "perccp- 
tual strategies , " which explain as well as account for certdin 
I81n addition to the automated IBM Recognition Gramar of 
Culicover, et ale (1969) and Winograd's work at MIT on Project 
MAC (1971), both of which we mentioned in the introduction, there 
is, of course, also the Halle-Stevens "Analysis-by-syn thesig" 
model (1964), which must be considered as a candidate for a 
recognition-grammar model theoretically--even though Halle 8pd 
Stevens were concerned with speech recognition per - se. Although 
we can see that the "Analysis-by-synthesis" model is useful in 
explaining hallucinatory reconstruction of speech and related 
phenomena, we feel that neither their fully active model (nor 
a fully passive model for that matter) will prove to recon- 
s,ruct accurately the complexities of normal human speech per- 
ception or to explain human language recognition and comprehension. 
65 
facts of historical. syntax (1972) , tend to con£ irm our working 
assumption, which is that there are perceptual cues in the sur- 
face structure of English that are critical in grammatical re- 
cognition. We hypq thesize that these perceptual cues, when 
combined with lexical informatian and information regarding the 
patterned movements and deletion8 that take place in syntactic 
con£ ismations i . paradigms, trans£ orma tions and incorporations) 
make up much of the "knowledge" that the speaker of a language 
gradually acquires and then uses in recognizing and understand- 
ing the sentences of his native language (or any language he learns 
and knows, for that matter). We are convinced that this know- 
ledge overlaps wLthY yet is somehow 3 different from, the knowledge 
and skills that are required if one wishes to produce grammatical 
sentences in a Language, 
Kelley (1968) , for example, has postulated that comprehen- 
sion is baaic in language acquisition and that rules of pro- 
duction are not essential for comprehension but may develop 
alongside of the necessary comprehension rules to satisfy other 
goals and purposes. If Kelley's model is correct, it indicates 
that human recognition graannars--an essential component of com- 
prehens ion--are ,not mere inverses of human production grammars. 
This, in turn, suggests that the most efficient computer-based 
recognition procedures will be based, not exclusively on gen- 
erative production- type grammars, but also on principles and 
66 
rules similar to those that human beings would seem to employ 
expressly in language recognition ana ramprchcnsi~n. On tho 
othe,r hand, the model also suggests that the most realistic 
tecognition procedure will overlap to a considerable extent 
with a produc tion-type grammar--utilizing, in different ways, 
a groat deal of tho same information. In the development of 
our recognition model, such psgchoZinguis tic considerations 
proved to be aos t useful. 

REFERENCES 
Beveq, T. and D. T. Langendoen, "The Interactior. ~f Speech 
Perception and Grammatical Structure in the Evolution of 
~anguage," paper psesented at the UCLA Conference oa His- 
torical Linguis tic8 in the Perspective of Transformational 
Theory, February, 1969. This paper is publishsd in 
Lin~uis tie Change and Generative Theory, R. S tockwell and 
R. Macaulay, eds., Bloomington: University of Indiana Press, 
1972. 
Chafe, Wallace, L., Meaning and the Structure of Langbage, 
(Chicago: University of Chicago press) , 1970. 
Chomsky, Noam, Aspects of the Theory of Syntax, (Cambr>idge, Mass 
MIT Press) , 1965. 
Chamsky, Noam, "Same Empirical Issues in the Theory of Trans- 
formational Grammar. " Reproduced by the Indiana Univer- 
sity Linguistics Club, 1970. 
Culicover, P., et al., "An Automated Recogni tion Grammar for 
English," Technical Report FSC 69-5907, IBM, Boston 
Programming Center, Cambridge, Mass. 02139, July,1969. 
Fillmore, Charles J. ,"Lexical Entries for Verbs, I' FoundB t ions 
of Language, Val. 4 (pp. 373-393) , 1968a. 
Fillmorb, Charles J., "The Case for Case, " in Universal$ in 
Linzuistic Thory, Bach and Harms ,eds. (New York: yolt, 
Rinehart, and Winston) , 1968b. 
Gruber, Jeffrey, "Functions of the Lexicon in Formal Descriptive 
Grammars, " Sys tern Development Corporation, Santa hhnicn, 
Calif. , TM-3770, December, 19.67. 
Gruber, Jef frep, "Studies in Lexical ~elations," bIIT Ph.D. 
diss~rtation, Sep t~lber 1965. (Reproduced by fhe Indiana 
Lrniversity Linguistics Club, January, I970 .) 
Hall, (Par tee) , Barbara, "Subject and Obj oc t in Hodcrn English, " 
NIT Ph.D. dissertation, J~ne~l.965, 
Halle, M. and K. N. Stevens, "Speech Recognition: A ?lode1 and 
a Program for Research," in The Structure of Langukge, 
Fodor and Katz, eds. (Englewaod Cliffs, N. J.: Prentice- 
Hall), 1964. 
Halliday , M.A .K., "Categories of the Theory of  ramm mar ,I1 WORD, 
17, 1961. 
Halliday, b1.A .K., "Some Notes on 'Decp ' Gramnar , " Journal of 
Linguistics 2, 1966. 
Halliday, M.A.K., "Notes on Transitivity and Theme in English, 
Journal of Linguistics, 3. 1967. 
Kelley, Kalon, "The Role of Semantics in Early Syntactic Ac- 
quisition," unpublished paper dated July, 1968. 
Lakoff, George, "The Nature of Syntactic ~rregularit~," The 
Computational Laboratory, Harvard University , Cambridge, 
Piass., Report No. NSF-16, December, 196 5. 
Lees. R. B., Grammar of English Nominalizatioris, IJAL, Vol. 26, 
No. 3. Part 11, July, 1960. 
UCLA English Syntax Project, (principal investigators: R. P. 
StockueLL, P. M. Schachtet, and B. H. partee), Integration 
of Transformational Theories of English Syntax, 2 Vols,
prepared under contract No. AF 19 (628-6007, for the U. S. 
Air Farce, La G. Hanscom Field, Bedford, Mass., October, 
1968, Report ESD-TR-68419). (Later publiehad in a revised 
and somewhate abridged form as The Malor Syantactic Sttuc- 
tures of English by Stockwell, Schachter, and Partee, 
New York: Holt , Rinehart and Wins ton, 1973.
Winograd, Tersy, Procedures as a Represkntatibn for Data in a 
Computer Pro~ram for Unders tandLnp; Natural Language, Project 
MAC, MAC TR--84, Mass. Institute of Technology, 1971. 
