Inheriting Verb Alternations 
Adam Kilgarriff 
Longman Dictionaries 
Burnt Mill 
Harlow, Essex CM20 2JE 
England 
Abstract 
The paper shows how the verbal lexicon can 
be formalised in a way that captures and 
exploits generalisations about the alterna- 
tion behaviour of verb classes. An alter- 
nation is a pattern in which a number of 
words share the same relationship between 
• a pair of senses. The alternations captured 
are ones where the different senses spec- 
ify different relationships between syntactic 
complements and semantic arguments, as 
between bake in "John is baking the cake" 
and "The cake is baking". The formal lan- 
guage used is DATR. The lexical entries it 
builds are as specified in HPSG. The com- 
plex alternation behaviour shared between 
families of verbs is elegantly represented in 
a way that makes generalisations explicit, 
avoids redundancy, and offers practical ben- 
efits to computational lexicographers. 
1 Introduction 
The paper shows how the verbal lexicon can be for- 
malised in a way that captures and exploits gener- 
alisations about the alternation behaviour of verb 
classes. An alternation is a pattern in which a num- 
ber of words share the same relationship between a 
pair of senses. The kinds of alternations to be cap- 
tured are ones where the different senses specify dif- 
ferent relationships between syntactic complements 
and semantic arguments, as in the relation between 
bake in "John is baking the cake" and "John is bak- 
ing", or between melt in "the chocolate melted" and 
*I would like to thank Gerald Gazdar and Roger Evans 
for their many valuable comments, and SERC for the 
grant under which the work wasundertaken. 
"Mary melted the chocolate" .1 Given that compact- 
ness and non-redundancy are a desideratum of theo- 
retical descriptions, the different usage-types for bake 
and wipe should not require us to introduce different 
primitives into the lexicon. Moreover, as the alter- 
nations are shared with other verbs, they should be 
described at some general node in a hierarchically 
organised lexicon, and inherited. 
DATR is a formal language in which the such rela- 
tionships and generalisations can be simply stated. 
Much has been written about verb alternations 
and their syntactic corollaries. Here we do not 
add to the evidence or construct new theory, but 
simply formalise other people's accounts: those of 
\[Atkins et al., 1986\] and \[Levin and Rappoport Ho- 
vav, 1991\]. The first investigates the range of al- 
ternations between transitive and intransitive forms 
of verbs. The second, titled Wiping the Slate Clean, 
explores the relations between meaning and subcate- 
gorisation possibilities for 'wipe' verbs, 'clean' verbs, 
and related groupings. The language used is DATR,. 
a default inheritance formalism designed for lexical 
representation. We follow Levin and Rappoport Ho- 
vav in taking a distinct subcategorisation frame as 
defining a distinct word sense, and also in work- 
ing with commonsense verb classes such as 'cooking 
verbs', since classes such as this serve to predict the 
alternations a verb participates in with some accu- 
racy. 
An important constraint is that the lexical entries 
are of a kind specified by a grammar formalism, so 
can be used for parsing and semantic interpretation. 
The formalism chosen in this paper is HPSG \[Pollard 
~The morphosyntactic distinctions between, for exam- 
ple, bake and is baking are not addressed here. Extensive 
DATR treatments of morphology are provided in various 
papers in \[Evans and Gazdar, 1990\]. 
213 
and Sag, 1987\]. 
Below we present detailed formal accounts for 
alternations involving cooking verbs and physicab 
process verbs. After motivating the DATR treatment 
and considering related work, we describe how verb 
entries appear in HPSG, then represent alternations 
as mappings between HPSG lexical entries, then in- 
troduce the main constructs of DATR and define a 
translation from HPSG notation to DATR. Finally 
we build a DATR inheritance network which repre- 
sents the alternate verb forms by inference, without 
the lexicographer having to explicitly say anything 
about them. 
The analysis presented in this paper is a part 'of 
a larger lexicon fragment which describes a further 
five alternations relating seven verb classes and for- 
malises much of the structure described in both ar- 
ticles. The complete fragment, illustrated in Fig. 1. 
is presented in full in \[Kilgarriff, 1992\]. 
WORD-CLASS 
V\]~B 
UNS 
• X 
o • • PH'Y~PR(JC "IRANSrrlVR 
, // , /I \ ~ ' 
X \ I ~u~ ~ ~. 
C-OF-S DITRANSrrlvB SURP-Cq~NT 
COOK.INO-VB i OIVI~ \[ ~ WH~'B 
/ &..." I \ ,-..!"w,. 
Cook "''- ~ RHMO~ ~...a Pluck 
BakB I ..r ~ Prune 
ll~n Ih~: defer ~. 
Broken Ibis: ~ label an l~lic~ 
frows po\[mt from ,.,b!na, e~a to panm~. 
Figure 1: Verb taxonomy 
1.1 Why DATR? 
As 'lexicalism' -- the doctrine that the bulk of the 
information about the behaviour of words should be 
located in the lexicon -- has become popular in com- 
putational and theoretical linguistics, so formalisms 
for expressing lexicM information have been devel- 
oped. The syntax, semantics and morphology of 
most words is shared with that of many others, so 
the first desideratum for any such formalism is to 
provide a mechanism for stating information just 
once, in such a way that it is defined for large num- 
bers of words. Inheritance networks serve this pur- 
pose. If words are arranged into a taxonomy or some 
other form of network, then a fact which applies 
to a class of words can be stated at a nonterminal 
node in the network and inherited by the words to 
which it applies. Work in knowledge representation 
has addressed questions of different kinds of network, 
and the kinds of machinery needed to retrieve inher- 
ited information, in detail (see, e.g., \[Brachman and 
Levesque, 1985\]). 
The next requirement is that exceptions and sub- 
regularities can be expressed. It must be possible to 
describe concisely the situation where a word or class 
of words are members of some superclass, and share 
the regular characteristics of the superclass in most 
respects, but have different values for some feature or 
cluster of features. SeverM lexical representation for- 
malisms addressing these desiderata have been pro- 
posed, e.g. DATR \[Evans and Gazdar 1989a, 1989b, 
1990\]; LRL \[Copestake, 1992\]; \[Russell et al. 1991\]. 
The work described here uses DATR. 
DATR has certain desirable formM and computa- 
tional properties. It is a formal language with a 
declarative semantics. Retrieving values for queries 
involves no search. Multiple inheritance specifica- 
tions are always orthogonal, so a word may inherit 
from more than one place, but any fact about that 
word has the place it is to be inherited from uniquely 
specified. The problem of different ancestors provid- 
ing contradictory values, often associated with mul- 
tiple default inheritance, is thereby avoided, yet the 
kinds of generalisation most often associated with 
the lexicon can still be simply stated. To date 
it has been used to express syntactic, morphologi- 
cal, phonological and a limited amount of seman- 
tic lexical information \[Evans and Gazdar, 1990; 
Cahill and Evans, 1990; Gibbon, 1990; Cahill, 1993\]. 
Verb alternations have not previously received a 
DATR treatment. 
1.2 Related work 
The work described here is at the meeting-point of 
lexical representation languages (as discussed above), 
lexical semantics (as in Atkins et al. and Levin and 
Rappoport Hovav; see also \[Levin, 1991\]) and for- 
mal accounts of alternations (see particularly \[Dowty, 
1979\]). 
Recent work which aims to bring these three 
threads together in relation to the lexical repre- 
sentation of nouns includes \[Briscoe et ai., 1990; 
Pustejovsky, 1991; Copestake and Briscoe, 1991; 
Kilgarriff, 1993 forthcoming; Kilgarriff and Gazdar, 
1993 forthcoming\]. (The latter two are companion 
papers to this, also using DATR in similar ways.) A 
paper addressing verbs is \[Sanfilippo and Poznanski, 
1992\]. 
This covers some of the same alternations as this 
214 
paper, and has similar goals. The formalism it uses 
is LRL, the typed default unification formalism of 
\[Copestake, 1992\]. Unlike DATR, this is both a lex- 
ical representation language and a grammar formal- 
ism. Whereas, in this paper, we represent the lexicon 
in DATR and then construct HPSG lexical entries, 
Sanfilippo and Poznanski need deal with only one 
formalism. This has a prima facie advantage but 
also a cost: the formalism must do two jobs. DATR 
is designed specifically for one, and offers more flexi- 
bility in the representation of exceptions and subreg- 
ularities. In LRL, multiple default inheritance is re- 
stricted to the cases where there is no clash, with the 
condition enforced by a checking procedure, in con- 
trast to DATR where the orthogonal nature of inheri- 
tance required by the syntax means that the problem 
does not arise. Also, LRL default inheritance must 
operate within the constraints of a type hierarchy, 
and the formalism requires two kinds of inheritance, 
default and non-default. In DATR, inheritance is not 
constrained by a type hierarchy, and inheritance, de- 
fault or otherwise, invokes a single mechanism. 
2 An HPSG-style lexicon 
The alternations to be addressed in detail here are 
the ones relating the transitive, which we treat as the 
base form, to the ergative ("The cake baked") and to 
the unspecified object ("John baked"). 
WORD bake 
MAJ 
SYN SUBCAT 
SEM 
RELN 
BAKER 
BAKED 
V 
(NP\[NOM\] 
NP\[ACC\] SEM ) 
Figure 2: AVM for transitive bake. 
Fig. 2 shows a simplified version of the HPSG lex- 
ical entry for transitive bake, in attribute-value ma- 
trix (AVM) notation. NP abbreviations and angle- 
bracket list notation, where a comma separates list 
elements and there is no separator between the con- 
i uncts of a feature-structure within a list, is as in Pollard and Sag, 1987\]. The boxed variables indi- 
cate the roles the semantic arguments play in the 
syntactic structure. 
For ergative bake, the same BAKE relation holds 
as in the base form, but now between an unspecified 
BAKER and a BAKED which is the subject of the 
sentence. The unspecified role filler is not 'bound' 
to a complement (i.e. any item on the SUBCAT list) 
but is existentially quantified (EX-Q). The ergative 
form is intransitive so has only one item on its SUB- 
CAT list and the SEM of that item unifies with the 
BAKED, so the AVM for ergative bake will be as in 
Fig. 3. For unspecified-object bake in "John was 
WORD bake 
\[MAJ V \] SYN 
SUBCAT (NP\[NOM\] SEMIS\]) \[RELN BAKE\] 
SEM BAKER EX-Q 
BAKED \['i-\] 
Figure 3: AVM for ergative bake. 
baking", the subject is matched to the BAKER and 
it is the BAKED which is unspecified, so existentially 
quantified, as in Fig. 4. 
WORD bake 
\[MAJ V \] SYN 
SUBCAT (NP\[NOM\] SEM \[~ ) 
SEM BAKER 
BAKED -Q 
Figure 4: AVM for unspecified-object bake. 
For bake and other cooking verbs, we are able to 
represent the extended senses directly in terms of the 
same predicate that applied in the base sense. We 
now move on to a case where this does not hold. 
For melt, the intransitive ("The ice melted") is ba- 
sic and the transitive ("Maria melted the ice") is ex- 
tended, and it is not possible to define the extended 
sense directly in terms of the base. The transitive 
can be paraphrased using cause, "Maria caused the 
ice to melt" and we call the alternation 'causative'. It 
is clearly closely related to the ergative, and it would 
be possible to treat the transitive form as basic, with 
the ergative alternation applying. That route has 
not been followed for two reasons. Firstly, melt is a 
member of a class of physical-process verbs, also in- 
cluding evaporate, freeze, dissolve, sublime and coa- 
lesce. They all clearly have intransitive senses. They 
all might, in the right setting, be used transitively, 
but in cases such as coalesce the transitive is not a 
standard use and it would patently be inappropriate 
for it to be treated as a base form. If we are to stand 
by the intuition that these verbs form a class, and 
all participate in the same alternation, then all must 
have an intransitive base form. 
Secondly, transitive melt introduces an aspect of 
meaning, call it CAUSE, which is not in any sense 
present in the intransitive. For bake, CAUSE is al- 
ready a component of the meaning, whether or not 
the verb is being used ergatively. A default en- 
tailment of CAUSE is that its first argument, the 
CAUSER, has proto-agent properties \[Dowty, 1991\]. 
If intransitive melt were treated like ergative bake, 
215 
CAUSE would be a component of the meaning of in- 
transitive melt. Its semantics would have an existen- 
tially quantified MELTER argument, which would 
he a CAUSER and which we would expect to have 
agent-like properties. Ifi ergative uses of bake, the 
baking scenario still includes an agent who is doing 
the baking and fills the BAKER role, even though 
they are not mentioned. (We concern ourselves here 
only with cooking bake, not '~rhe stones baked in the 
sun" and other usage-types where bake is behaving as 
a physical process verb.) In '°the ice melted" there is 
usually no agent involved. While it might always be 
possible to assign a filler to the MELTER slot, per- 
haps "the hot temperature" or "the warm climate", 
they do not fit readily into the agent, CAUSER role. 
So we do not treat causatives as ergatives. 
A standard analysis of causatives after \[Dowty, 
1979\] as presented by \[Chierchia and McConnell- 
Ginet, 1990, chapter 8\], is 
AyAzM ELT/2(z, y) = Ay)~zCAU SE(z, M ELT/I(y) ). 
The semantics of the causative has the predi- 
cate CAUSE, with MELT/1 re-appearing as its sec- 
ond argument. In addition to intransitive melt as 
shown in Fig. 5 we have causative melt as shown in 
Fig. 6. (The relation between lambda expressions 
and feature structures is discussed in \[Moore, 1989; 
Kilgarriff, 1992\].) 
WORD 
SYN 
SEM 
melt 
MAJ V \] 
SUBCAT (NP\[NOM\] SEM ~\] ) 
RELN MELT/I \] 
MELTED E\] 
Figure 5: AVM for intransitive melt. 
WORD 
SYN 
SEM 
melt 
SUBCAT (NP\[NOM\] SEM , 
NP\[ACC\] SEM ) 
OA SER  \]REL  MELT/, 
O USED L MEL EDF1 
Figure 6: AVM for causative melt. 
3 DATR: a gentle introduction 
A simple DATR equation has, on its lhs, a node and 
a path, and, on its rhs, either a value: 
Nodel:<a b c> Iffi value. 
or an inheritance specification. Nodes start with cap- 
ital letters, paths are sequences enclosed in angle- 
brackets, anything on the rhs that is not a node or a 
path is a value. The primary operation on a DATR 
description is the evaluation of a query, that is, the 
determination of a value associated with a given path 
at a given node. Where a value is not given directly, 
it may be inherited by following a trail: the inheri- 
tance specification on the dis at step n becomes the 
lhs for step a-/-l. The specifications may state both 
node and path, node only or path only. They may 
also be local or global. Where they are local, the 
unstated node or path is as it was on the lhs, so if 
we have the node: 
Node1: <a> -- Node2: <x> 
<b> ~, Node3 
<c> B <y>. 
then 
Node1: <a> inherits from Node2: <x> 
Node1: <b> inherits from Mode3: <b> 
Node1: <c> inherits from Node1: <y>. 
(Where a number of node-path specifications for a 
given node are stated together, the node need not 
be re-iterated. The full stop is delimiter for either a 
single equation or such a cluster of equations.) 
Where inheritance specifications are global, with 
the node or path on the rhs in double quotes: 
Node4: <a> -- "NodeS" 
<b) Im "<Z>". 
then the 'global context' node or path is picked up 
to complete the specification. For the purposes of 
this paper, the global context node and path are the 
initial query node and path. 
When there is no lhs to exactly match a node-path 
pair to be evaluated, the mechanism which gives rise 
to DATR's nonmonotonicity comes into play. This is 
the 'longest leading subpath' principle. The node- 
path pair inherits according to the equation at the 
node which matches the longest leading subpath. 
Thus, with Node1 as defined above, 
Nodel:<a ax ay> inherits from Node2:<x ax ay> 
Node1 : <b bx by> inherits from Node3 : <b bx by> 
Node1: <c cx cy> inherits from Node1: <y cx cy> 
If there were any more specific paths defined at 
Node 1, for 
<a ax>, 
<a ax ay>, 
<b bx>, etc., 
then these inheritances would be overridden. Note 
that the match must be with the longest leading sub- 
path. In this fragment, the queries 
Node 1 : <d>, 
Node1 : <ax a>, and 
Node1 : <> 
216 
all fail to match and are undefined. (The other 
queries may also be undefined, if the trail of inher- 
itance specifications terminates without reaching a 
value at some later stage, but they are not found to 
be undefined at this stage.) 
Two particular cases of inheritance used in the pa- 
per are: 
NodeS: <> == Node6 
<e> == Node7:<>. 
In the first, the leading subpath to be matched is 
null, so this is a default of defaults: no queries will 
terminate at this point, since any query which does 
not make a more specific match will match this line 
and get passed on from Node5 to }lode6, path un- 
changed. This is the simplest form of inheritance, 
usually used to specify the basic taxonomy in a DATR 
theory. In the second, path element e is 'chopped' 
from the beginning of the path, so: 
Node5 :<e ex ey> inherits from Node7: <ex ey>. 
4 Translations into DATR 
Now we move on from describing the alternations, 
and describing the inheritance formalism, to repre- 
senting the alternations within the formalism. The 
DATR translation is straightforward: AVMs can be 
rewritten as sets of equations which then become sets 
of DATR equations. DATR paths must be associated 
with nodes, so a node for the paths to be located at is 
introduced. FIRST and REST have been shortened 
to fi and re. DATR is not a unification formalism, 
and all the theory will do in relation to re-entrancies 
will be to mark them with matched pairs of variables, 
here vl, v2 etc., to be interpreted as re-entrant pairs 
outside DATR. We introduce the feature binding for 
the variables to be the value of. 2 In order that gener- 
alisations covering BAKERs, COOKERs and FRY- 
ERs can be stated, we replace verb-specific names 
such as BAKER for slots on a semantic args list. 
(This does not represent a change in the semantics: 
the first member of the argument list of the bake 
predicate will continue to be the BAKER whatever 
lexical entry it occurs in. It simply allows us to ex- 
press generalisations.) We use pred for the predi- 
cate rather than RELN. Following these changes, the 
(simplified) DATR lexical entry for transitive bake is: 
Bake : <word> = bake 
<syn maj> = v 
<syn subcat fi sem binding> = vl 
<syn subcat re fi sem binding> = v2 
<syn subcat re re> = nil 
<sem pred> = bake 
<sere args fi binding> = vl 
2The feature also makes it possible to use the fact that 
a semantic argument has an existential-quantification 
(ex-q) binding to override the default that it is bound 
to a complement. 
<sere args re fi binding> = v2 
<sem args fi binding> ffi nil. 
5 An inheritance hierarchy 
The next task is to place the verbs in a hierarchy so 
generalisations need stating only once. DATR allows 
different kinds of information to be inherited from 
different places, and also allows generalisations to be 
overridden by either idiosyncratic facts or subregu- 
larities. The hierarchy is illustrated in Fig. 1. At 
the top of the tree is WORD-CLASS, then VERB, from 
where all verbs inherit. They all have a subject, and 
by default this unifies with the first item on the axgs 
list. There will be no call for an INTRANSITIVE node 
because all the positive information that might be 
stated there is true of all verbs so can be stated at 
the VERB node, and the negative information that 
intransitive verbs do not have direct objects is ex- 
pressed by the termination of the subcat list after 
its first item at VERB (via ARG and NIL; see below). 
TRANSITIVE inherits from VERB, adding the default 
binding between second complement and second ar- 
gument. 
VERB: <>  ffi WORD-CLASS 
<syn maj> == verb 
<syn subcat fi sere binding> == vl 
<sere args fi binding> == vl. 
TRANSITIVE: <> == VERB 
<eyn subcat re fi sere binding>  - v2 
<sere args re fi binding> == v2. 
List termination involves a measure of ingenuity, in 
order that nil is the value of <syn subzat re> and 
<sem args re> at VERB and <syn subcat re re> 
and <sere args re re> at TRANSITIVE, but nowhere 
else: 3 
VERB: <sere args> == ARG: <> 
<syn subcat> ffi= COMP:<>. 
<syn subcat fi syn case> ffi= nom 
<sem args fi semfeats>  = AGENT:<>. 
TRANSITIVE: <syn subcat re> =ffi COMP:<> 
<sem args re> ffi= ARG:<>. 
ARG: <f i semf eats> == PATIENT: <> 
<re> ffi= NIL:<>. 
COMP:<fi syn> == NP:<> 
<re> == NIL:<>. 
NIL:<> == nil 
<fi> == UNDEF 
<re> ffi= UNDEF. 
The COMP and ARG nodes provide a location for de- 
fault information about syntactic complements and 
semantic arguments. Complements are, by default, 
accusative noun phrases. Following \[Dowry, 1991\], 
we have a default expectation that subjects will have 
'proto-agent' semantic features and objects, 'proto- 
patient' ones. The role of Dowty's approach in this 
analysis is that it gives us a way of marking the dif- 
ference between agents and patients which says more 
3This treatment is due to Roger Evans. 
217 
than simply using the labels 'agent' and 'patient', 
and has the potential for subtler distinctions, with 
different subsets of proto-agent and proto-patient 
features applying to subjects and objects of different 
verb classes. AGENT and PATIENT set up the expected 
values for four of the characteristics Dowty discusses. 
NP:<maj> == n 
<case> == ace. 
AGENT:<volition> == yes 
<sentient> == yes. 
PATIENT: <changes-state> == yes 
<causally-affected> == yes. 
The default accusative case and proto-patient seman- 
tic features must be overridden in the case of the 
subject: 
VERB:<syn subcat fi syn case> == nom 
<sam args fi semfeats> == AGENT:<>. 
To this skeleton, we add some smaller classes 
based on meanings. Once we introduce them we 
can start expressing generalisations about alterna- 
tion behaviour. To distinguish alternate forms from 
base forms, we introduce the alt prefix. To re- 
quest information about a non-base form, we start 
the query path with alt x, where x is a label identi- 
fying the alternation under consideration. We adopt 
a convention whereby all-upper-case nodenames are 
used for nodes for classes of words, such as cook- 
ing verbs, while lexical nodes have only initial letters 
capitalised. 
Bake:<> == C00KING-VB 
<word>  ffi bake 
<sam pred>  ffi bake. 
C00KING-VB:<>  ffi C-OF-S 
<sam arSS re fi semfeats edible>  = yes. 
C-0F-S: <> == TRANSITIVE 
<alt erg>  = PHYS-PROC:<> 
<alt erg sam> =ffi "<sere>" 
<alt erg sam args fi binding> == ex-q 
<alt erg sam args re fi binding>  ffi vl. 
Bake is a cooking verb, and cooking verbs are, in 
the base case, transitive change-of-state verbs. Thus 
Bake inherits, by default, from C00KING-VB which 
inherits from C-0F-S (for 'change of state') and then 
from TRANSITIVE, so acquiring the default specifica- 
tions for semantic features for its subject and object, 
and the re-entrancies between subject and first argu- 
ment, and object and second argument. The DATR 
fragment now represents all the information in the 
DATR lexical entry for bake presented above, and 
case and proto-agent and proto-patient specifications 
in addition. 
The first generalisation about alternations that 
we wish to capture is that change-of-state transi- 
tives such as bake undergo the ergative alternation 
to become change-of-state intransitives, or 'physical 
process' verbs. We access the lexical entries for the 
ergative forms of verbs with DATR queries with the 
path prefix alt erg, which work as follows. The 
semantics of the ergative will be the same predicate- 
argument structure as the base form, and this is im- 
plemented in the third line of the ¢-0F-S node which 
tells us, with the double-quotes, to inherit the erga- 
tive's semantics from the semantics of the node for 
the base form of the verb. The two further speci- 
fications for ergatives are that the first argument is 
existentially quantified, and the second unifies with 
the first complement via vl. 
In all other matters, as the second line of the 
C-0F-S node tells us, the ergative form is diverted 
to inherit from a node for physical-process intransi- 
tives: 
PHYS-PR0C:<> == VERB 
<sam args fi semfeats>  = PATIENT:<>. 
The first semantic argument of a physical-process in- 
transitive has proto-patient semantic features and 
otherwise inherits from VERB. This is a case where 
the default - that first semantic arguments (realised 
as subjects in the base case) have proto-agent fea- 
tures - has been overridden, but the reader will note 
that this has been entirely straightforward to express 
in DATR. 
We now have almost all the information needed to 
build the lexical entry for ergative bake. One item we 
do not yet have is the intuitively obvious fact that 
the word for the alternate form is the word for the 
original. This is true by definition for all alternate 
forms. All alternate forms will eventually have their 
alt x prefix (or prefixes) stripped and inherit from 
WORD-CLASS at the top of the tree. So we add the 
following line: 
NORD-CLASS:<word> == "<word>". 
Now all alternate forms will inherit their .ord from 
the word at the global context node, which will al- 
ways be the node for the base form. 
Many cooking verbs undergo the 'unspecified ob- 
ject' alternation, for which we shall use the label 
unspe¢. All information relating to this form is gath- 
ered at an UNSPEC node: 
UNSPEC: <> == VERB 
<sam> == "<sam>" 
<sam args re fi binding> :ffi ex-q. 
This simply states that the form is a standard intran- 
sitive, with the semantics of the base form except 
that the second argument is existentially quantified. 
Cooking verbs with alt unspec prefixes are diverted 
here by the addition of: 
C00KING-VB:<alt unspec>  ffi UNSPEC:<>. 
Now we move on to melt, a physical-process verb 
with a causative form. The ergative alternation led 
from C-0F-S to PIIYS-PROC. This makes a similar 
journey in the opposite direction, from PIIYS-PROC 
to CAUSE and then TRANSITIVE. The alternation la- 
bel is cause. 
218 
Melt:<> == PHYS-PROC 
<sem pred> == melt 
<cord> == melt. 
PHYS-PROC:<> == VERB 
<alt cause> -= CAUSE:<> 
<alt cause sem args re fi> == "<sem>" 
<alt cause sen ares re fi 
ares fi binding> == v2. 
CAUSE:<> == TRANSITIVE 
<sem pred> == cause. 
Causative melt, with the alt cause prefix, is 
a regular verb of causing, and inherits its syntax 
and most of its semantics including the predicate 
cause/2 from CAUSE. Its first argument will have the 
usual characteristics of a CAUSER, and its second, 
the predicate-argument structure of the base form of 
the verb. As the predicate melt is now identified as 
the second argument of cause, the item that melts 
is identified as the first argument of the second ar- 
gument of the causative form of the verb, and it is 
this which is re-entrant with the second item on the 
subcat list, as specified in the final line of PHYS-PR0C. 
The reward for this superstructure is that lexical 
entries can now be very concise. By adding a three- 
line entry, e.g., 
Bake: <> == COOKING-VB 
<gord> == bake 
<sem pred> == bake. 
to the lexicon, we make available, for cooking verbs 
such as bake, a set of eighteen specifications for the 
base form, and fifteen each for the ergative and 
unspecified-object, and for physical process verbs, fif- 
teen for the base and eighteen for the causative, all 
complete with case, subcategorisation, proto-agent, 
proto-patient and re-entrancy specifications, as be- 
low: 
Bake: 
Bake: 
Bake : 
Bake: <syn 
Bake: <syn 
Bake: <syn 
Bake: <syn 
Bake : <syn 
Bake : <syn 
Bake: <syn 
Bake : <sem 
Bake: <sem 
Bake: <sem 
Bake : <sem 
Bake: <sen 
Bake : <sel 
Bake: 
Bake: 
<lexical> = true. 
<eord> ffi bake. 
<synmaj> = verb. 
<sem 
<sem 
subcat 
subcat 
subcat 
subcat 
subcat 
subcat 
fi syn maj> = n. 
fi syn case> = nom. 
fi sen binding> = vl. 
re fi syn ~aj> = n. 
re fi syn case> =acc. 
re fi sem binding> - v2. 
subcat re re> = nil. 
pred> = bake /2. 
args fi binding> = vl. 
ares fi 
ares fi 
ares re 
ares re 
ares re 
ares re 
semfeats volition> = yes. 
semfeats sentient> = yes. 
fi binding> = v2. 
fi semfeats 
changes-state> = yes. 
fi semfeats 
causally-affected> - yes. 
re> = nil. 
6 Summary and discussion 
First, HPSG-style verbal lexical entries, and the 
mappings between them corresponding to alterna- 
tions, were described. But at this stage, the gener- 
alisations were not captured. So then these entries 
were translated into DATR, and arranged into a tax- 
onomy so an alternation only needed expressing once, 
at a non-terminal node from which the verbs to which 
it applied would inherit. Information about syntax, 
semantics, and patterns of polysemy was concisely 
expressed in a manner both theoretically and com- 
putationally appealing. 
The lexicon fragment described in detail is part 
of a larger fragment which also formalises the rela- 
tions holding between transitives and intransitives 
of 'care' verbs such as wash, where the intransitive 
means the same as the reflexive; between transi- 
tive, intransitive, and two ditransitive forms of the 
'clear' verbs ("clear the desk"; "the skies cleared"; 
"clear the desk of papers"; "clear the papers off the 
desk"); and between transitive and ditransitive forms 
of 'wipe' verbs ("wipe the shelf"; "wipe the dust off 
the shelf"). The complete fragment thus covers a 
number of the common transitivity alternations of 
English. 
The paper aims to present both a study of lexical 
structure and an approach to practical lexicography. 
On the latter score, the ideal to which the paper con- 
tributes sees the lexicographer only ever needing to 
explicitly enter information that is idiosyncratic to 
a word and inheritance specifications, as everything 
that is predictable about a word's behaviour will be 
inferred. Maintaining consistency in the lexical rep- 
resentation, and updating and revising it, will also be 
quicker if a generalisation is located just at one place 
in the lexicon rather than at every word to which it 
applies. 
Transitivity alternations defy classification as ei- 
ther syntactic or semantic phenomena. They are 
clearly both. The generalisations are associated with 
semantic classes of verbs, and have both syntactic 
and semantic consequences. The verb taxonomy of 
Fig. 1 may be used for conveying specifically linguis- 
tic information, as explored in this paper, but also 
potentially forms part of an encyclopedic knowledge 
base, with knowledge about any type of cooking held 
at the COOKI~G-VB node and knowledge specifically 
about frying and baking at the Fry and Bake nodes. 
It might be argued that this is to confuse two differ- 
ent kinds of information, but, as illustrated in this 
paper and argued in \[Kilgarriff, 1992\], the lexicon 
of English holds both the syntax and semantics of 
lexical items. The approach offered here indicates 
how linguistic and encyclopedic generalisations may 
be attached to the same taxonomic structure. 
\[Boguraev and Levin, 1990\] show that an expres- 
sively adequate model for the lexicon must incorpo- 
rate productive rules so that novel but rulebound 
uses of words can be captured. Thus "the her- 
219 
ring soused" is interpretable by any English speaker 
who has come across soused herring, but intransitive 
souse will not be added by any lexicographer to any 
dictionary: it is most unlikely that any corpus will 
provide any evidence for the form, and if it did, it 
would be of insufficient frequency to justify explicit 
treatment. The ergative form of souse must there- 
fore be in the lexicon implicitly. Its availability to 
speakers and hearers of English can be inferred from 
knowledge of the kind of verb which souse is and the 
kinds of processes, or alternations, that verbs of that 
class can undergo. The DATR analysis demonstrates 
how such implicit availability of verb forms can be 
formalised. 
6.1 Further work 
A further question that the question of productivity 
invites is this: how are we to represent which verbs 
undergo which alternations? First, we might wish 
to develop devices within DATR or a related formal- 
ism for identifying which alternations apply where, 
and two such mechanisms are presented in \[Kilgar- 
rift, 1992\]. But as we look closer, and consider the 
difficulty of placing many verbs in a semantic class, 
or the role of metaphor, analogy, and simple familiar- 
ity in determining which alternations are applicable 
in a given context of language-use, so the idea of a 
yes/no answer to questions of the form, "does this 
verb undergo this alternation?" loses plausibility. 
This reasoning applies also to verb classes. The 
analysis offers an account of verb behaviour which 
is premised on verb classes, but their only justifica- 
tion has been by appeal to commonsense and an ill- 
defined notion of their ability to predict which alter- 
nations a verb participates in. Nothing has been said 
about how the classes might be identified, or how 
decisions regarding where a verb should be placed 
might be made. 
The questions, "what class does a verb belong 
to?", "what are the relative frequencies of the dif- 
ferent patterns it occurs in?", and "is this pattern 
grammaticalT' are intimately connected. Alterna- 
tion behaviour is a major source of evidence as to 
how a verb should be classified, and grammaticality 
judgements are premised upon the patterns a com- 
petent speaker has frequently encountered in their 
experience of the language. The further develop- 
ment of computational lexical semantics of the kind 
described in this paper requires foundational work 
on the relation of corpus-based statistical findings to 
formal knowledge representation. 
References 
\[Atkins et al., 1986\] B. T. S. Atkins, Judy Kegl, and 
Beth Levin. Explicit and implicit information in 
• dictionaries. In Advances in Lexicography: Proc. 
Second Ann. Conf. of the UW Centre for the New 
OED, pages 45-65, Waterloo, Canada, 1986. 
\[Boguraev and Levin, 1990\] Branimir K. Boguraev 
and Beth Levin. Models for lexical knowledge 
bases. In Electronic Text Research: Proc. Sixth 
Ann. Conf. of the UW Centre for the New OED, 
pages 65-78, Waterloo, Canada, 1990. 
\[Brachman and Levesque, 1985\] Ronald J. Brach- 
man and Hector J. Levesque, editors. Readings 
in Knowledge Representation. Morgan Kaufmann, 
Los Altos, California, 1985. 
\[Briscoe et al., 1990\] Edward J. Briscoe, Ann A. 
Copestake, and Branimir K. Boguraev. Enjoy the 
paper: Lexical semantics via lexicology. In COL- 
ING 90, volume 2, pages 42-47, Helsinki, 1990. 
\[Cahill and Evans, 1990\] Lynne J. Cahill and Roger 
Evans. An application of DATR: The TIC lexicon. 
In Proc. ECAI-90, pages 120-125, 1990. 
\[Cahill, 1993\] Lynne J. Cahill. Morphonology in the 
lexicon. In ACL Proceedings, 6th European Con- 
ference, Utrecht, 1993. 
\[Chierchia and McConnell-Ginet, 1990\] 
Gennaro Chierchia and Sally McConnell-Ginet. 
Meaning and Grammar. An Introduction to Se- 
mantics. MIT Press, Cambridge, Mass., 1990. 
\[Copestake and Briscoe, 1991\] Ann A. Copestake 
and Edward J. Briscoe. Lexical operations in 
a unification-based framework. In James Puste- 
jovsky and Sabine Bergler, editors, Lexical seman- 
tics and knowledge representation: ACL SIGLEX 
Workshop, Berkeley, California, 1991. 
\[Copestake, 1992\] Ann Copestake. The ACQUILEX 
LKB: representation issues in semi-automatic ac- 
quisition of large lexicons. In Proc. Third Conf. on 
Applied Natural Language Processing, pages 88- 
95, Trento, Italy, 1992. Association of Computa- 
tional Linguistics. 
\[Dowty, 1979\] David R. Dowty. Word Meaning in 
Montague Semantics. Reidel, Dordrecht, 1979. 
\[Dowty, 1991\] David R. Dowty. Thematic proto- 
roles and argument selection. Language, 
67(3):547-619, 1991. 
\[Evans and Gazdar, 1989a\] Roger Evans and Gerald 
Gazdar. Inference in DATR. In ACL Proceedings, 
4th European Conference, pages 1-9, Manchester, 
1989. 
\[Evans and Gazdar, 1989b\] Roger Evans and Ger- 
ald Gazdar. The semantics of DATR. In An- 
thony G. Cohn, editor, Proe. Seventh Conference 
of the AISB, pages 79-87, Falmer, Sussex, 1989. 
\[Evans and Gazdar, 1990\] Roger Evans and Gerald 
Gasdar. The DATR papers. Technical Report 
CSRP 139, School of Cognitive and Computing 
Sciences, University of Sussex, Falmer, Sussex, 
1990. 
\[Gibbon, 1990\] Dafydd Gibbon. Prosodic associa- 
tion by template inheritance. In Proc. Workshop 
220 
on Inheritance in Natural Language Processing, 
pages 65-81, Tilburg, 1990. ITK. 
\[Kilgarriff and Gazdar, 1993 forthcoming\] Adam 
Kilgarriff and Gerald Gazdnr. Polysemous rela- 
tions. In Frank R. Palmer, editor, Festschrifl for 
Sir John Lyons. CUP, Cambridge, England, 1993, 
forthcoming. 
\[Kilgarriff, 1992\] Adam Kilgarriff. Polysemy. PhD 
thesis, University of Sussex, CSRP 261, School of 
Cognitive and Computing Sciences, 1992. 
\[Kilgarriff, 1993 forthcoming\] Adam Kilgarriff. In- 
heriting polysemy. In Patrick St. Dizier and Eve- 
lyne Viegas, editors, Computational Lexical Se- 
mantics. CUP, 1993, fortheoming. 
\[Levin and Rappoport Hovav, 1991\] Beth Levin and 
Malka Rnppoport Hovav. Wiping the slate clean: 
A lexieal semantic exploration. Cognition, 41:123- 
151, 1991. 
\[Levin, 1991\] Beth Levin. Building a lexicon: The 
contribution of linguistics. International Journal 
of Lezicography, 4(3):205-226, 1991. 
\[Moore, 1989\] Robert C. Moore. Unification-based 
semantic interpretation. In A CL Proceedings, P7th 
Annual Meeting, pages 33--41, Vancouver, 1989. 
\[Pollard and Sag, 1987\] Carl Pollard and Ivan A. 
Sag. An Information-Based Approach to Syntax 
and Semantics. Volume I: Fundamentals. CSLI 
Lecture Notes, No. 13. Chicago University Press, 
Chicago, 1987. 
\[Pustejovsky, 1991\] James Pustejovsky. The gen- 
erative lexicon. Computational Linguistics, 
17(4):409-441, 1991. 
\[Russell et al., 1991\] Graham Russell, Afzal Ballim, 
John Carroll, and Susan Armstrong-Warwick. A 
practical approach to multiple default inheritance 
for unification-based lexicons. In Edward J. 
Briseoe, Ann A. Copestake, and Valeria de Paiva, 
editors, Proc. ACQUILEX workshop on default in- 
heritance in the lezicon, Tech. report 238, Univer- 
sity of Cambridge Computer Laboratory, 1991. 
\[Sanfilippo and Poznanski, 1992\] Antonio Sanfil- 
ippo and Victor Poznanski. The acquisition of lex- 
ieal knowledge from combined machine-readable 
dictionary sources. In Proc. Third Conf. on Ap- 
plied Natural Language Processing, pages 80-87, 
Trento, Italy, 1992. Association of Computational 
Linguistics. 
221 
