Lexical Rules: What are they?

Andrew Bredenkamp, Stella Markantonatou and Louisa Sadler

Department of Language and Linguistics
University of Essex, UK
fandrewb,marks,louisag@essex.ac.uk

Abstract ive Alternation, etc) and word formation phenom-

ena (inectional and derivational morphology). In
Horizontal redundancy is inherent to lex- fact, Pollard and Sag also refer to declension class
ica consisting of descriptions of fully membership and similar facts as horizontal rela-
formed objects. This causes an unwel- tions, and as we shall see, the boundary between
come expansion of the lexical database vertical and horizontal relations is not immutably
and increases parsing time. To eliminate xed once and for all.
it, direct relations between descriptions The notion of lexical rule is often given some
of fully formed objects are often dened. status at the level of linguistic or psychological
These are additional to the (Typed Mul- theory. (Pollard & Sag 1987) make reference to
tiple) Inheritance Network which already a generative or procedural interpretation of lexical
structures the lexicon. Many implement- rules as a deductive mechanism which can be de-
ations of horizontal relations, however, ployed on a needs only basis, for example, to gen-
fail to generate lexical entries on a needs- erate words from a single base form. The concep-
driven basis, so eliminate neither the tion of lexical rules as essential generative devices
problem of lexicon expansion nor that of (rather than static statements expressing (sub-)
inecient parsing. Alternatively, we pro- regularities), is shared in much inuential work
pose that lexical entries are descriptions (e.g. (Bresnan 1982), (Pinker 1989)), although
of objects open to contextual specica- it is by no means universal, even within HPSG.
tion of their properties on the basis of Viewed from an implementational perspective, on-
constraints dened within the type sys- the-y application of lexical rules brings with it a
tem. This guarantees that only those number of distinct advantages which follow from
grammatical lexical entries are infered the drastic reduction in the size of the lexical data-
that are needed for ecient parsing. The base (lexical construction is less time consuming
proposal is extremely modest, making and parsing time should be reduced as lexical look
use of only basic inference power and ex- up is less ambiguous, etc). At rst sight then it ap-
pressivity. pears that the benet of adding an external Lexical
Rule component outweighs the disadvantages (ex-
ternal powerful mechanisms). We will rst show
1 Lexical Rules: what are they? that their role is less clear than this suggests and
Within the strongly lexical framework of HPSG, certainly more problematic, before suggesting in
lexical rules are used to express relations among Section 2 an alternative which eschews any extra
descriptions|-a kind of indirect \horizontal re- mechanisms.
latedness" (Pollard & Sag 1987, 209) which can be
contrasted with the vertical relations between the 1.1 Horizontal and Vertical Redundancy
type(s) of lexical elements. Type relations are, of The parallel drawn above between vertical related-
course, captured directly as the monotonic (typed) ness (expressed with the type system) and ho-
multiple inheritance network itself, which struc- rizontal relatedness among descriptions of fully
tures the lexicon. formed objects is however rather misleading.
Typical examples of horizontal redundancy in Monotonic multiple inheritance networks are most
the hierarchical lexicon thus conceived are the Al- naturally used to represent generalisations over
ternation phenomena (e.g. Dative Shift, the Locat- the properties that (groups of) linguistic objects
share | inspection of any network will conrm constrained to account for `exceptional' behaviour,
that they are usually deployed to express what that is, for those words which do not participate
is essentially a componential analysis of objects to a given horizontal relation despite the fact that
and of the relationship between them (dened on their description makes them appropriate candid-
the basis of this analysis). On the other hand, ates for the relation (verb alternations oer several
horizontal relations among descriptions (very of- examples of these situtation, for instance, `giving'
ten modelled by means of lexical rules) are es- verbs which do not exhibit the so-called `dative
sentially relations holding directly between objects shift' phenomenon).
themselves. While this intuition is clear, this is Modelling of `exceptional' behaviour leads either
much less adequate an approach for morpholo- to an extreme complexity of the type system or
gical relatedness, where a componential approach to non-monotonic solutions (Flickinger 1987) be-
may often appear just as natural as an object re- cause it turns out that certain horizontal relations,
latedness view, especially if the formalismincludes usually dened over types, must be blocked for in-
functionally dependent values, permitting the ex- dividual objects.
pression of allomorphic variation and the like. In
fact, many putatively horizontal relations may be 1.3 Implementing Horizontal Relations
simply re-expressed within a type hierarchy by Several dierent implementations of horizontal re-
viewing them from a componential perspective, lations exist. All of them add extra machinery
obviating the need for expressing them on the \ho- and some add extra expressive power to the core
rizontal" dimension which may lead to the use of mechanism.
lexical rules. But this is only possible once one Most frequently, horizontal relations are imple-
frees oneself from a view of lexical relatedness as mented as unary rules operating at parsing time
something which holds essentially between words within a derivational component. Such a compon-
(objects which correspond to maximal types, that ent is added to the inheritance machinery for in-
is types at the bottom of the type hierarchy). dependent reasons, mainly because of the limited
Horizontal relations are perhaps most naturally expressivity of the type system. With LRs, some
captured by an extra device (LRs) external to the lexical entry is considered as `basic' and all other
lexical network and associated inference mechan- lexical entries are derived from it introducing oth-
ism { see (Krieger & Nerbonne 1993) and (Cal- erwise unjustied directionality to the grammar.
cagno 1995) for recent HPSG proposals. Some re- In addition, the derivational implementation of ho-
cent work ((Meurers 1995) and (Riehemann 1994)) rizontal relations fails to produce lexical entries as
partly departs from this view by expressing rela- needed, instead, it produces lexical entries accord-
tions between objects using the vertical axis (that ing to the system's internal algorithm of searching
is, using the type system), but again the starting the rule space. Considerable ambiguity is intro-
point is `complete' lexical objects. duced with unpleasant results for parsing time.
1.2 Why avoid Horizontal Relations? Extra machinery for blocking these rules in or-
der to account for exceptional behaviour is also
Horizontal relations have a number of undesirable necessary.
features as well as requiring an external mechan- Alternatively, LRs may be compiled out but,
ism. Horizontal relations (between objects) are under this approach too, problems like direction-
in principle pretty much unconstrained. Vertical ality and the blocking of LRs as well as expensive
relations are more constrained because they are ambiguity at parsing time remain unsolved.
based on componential analysis, starting out from
the set of properties that objects have. On the
other hand, any object can be related to any other
2 An alternative proposal
object by stipulation in an external mechanism. In In this paper we explore an alternative to hori-
architectural terms, it is simply accidental (if for- zontal relatedness which exploits the idea that it is
tuitous) that lexical rules are often used to relate often possible to conceive of the linguistic objects
minimally dierent objects | they are capable of in such a way as to eliminate potential sources
much more promiscious behaviour. of ambiguity and additional external mechanisms.
This state of aairs is amply demonstrated in To illustrate our approach we will propose an ac-
the literature, which abounds with attempts to count of a subset of Verb Alternation phenom-
constrain horizontal relations by appeal to subsi- ena which rely on what are essentially underspe-
diary principles (predicate locality in LFG, con- cied lexical entries. The lexicon will then con-
straints of a psycholinguistic nature in the work of tain one (verbal) entry and the system will rely
(Pinker 1989), etc). Horizontal relations must be only on the existing resources (the type hierarchy)
to provide the dierent interpretations of the pre- pointer is included as an extra feature of the
dicate which license the distinct complementation value of synsemjlocjcontjnucleus. This fea-
patterns. Analysis is incremental and determin- ture we name sem(antic) cons(traints) and
istic and the procedure relies mainly on what we we make it appropriate for the same values that
will call `trivial type inference'. In the sections the prepositional synsemjlocjcontjnucleus is
that follow rst we discuss the linguistic approach assigned. The lexical entry for to load would look
underlying our proposal, second we compare our as in (6).
proposal to existing underspecication approaches (6) 2phon l,o,a,d  3
and nally, we give some details of the implementa- 2 2   337
tion which relies on no special features or external 6
6 subj np: 1 7
devices. 6 6cat 6  h i 777
6
6 6 4 comps   577
6 6 np: 2 ,pp: 4 comps np: 3 77
2.1 Underspecication 6 6 77
6 6 2 3 77
6synsem j loc 6 rel load 77
We will exemplify our approach by treating a 6 6 77
7
6 6 6 arg1 1 7 77
subset of verb alternations which conform to 6 6 6
6content 6arg2 2 7
7
7
77
the following general schema (1). These include 6 6 6arg3 3  77
6
6 6 6 7
7 77
the so-called spray/load (locative) alternation, the 4 4 4sem cons 4 arg1 2 5 57
5
wipe/clear alternation, the break/hit alternation contact
arg2 3
etc (Levin 1993).
(1) V NPj P1 NPk ] ! V NPk P2 NPj ] The lexical entry for the preposition with is
given below:
We adopt the view that verb predicates are open
to contextual information (which must be contras- (7) 2phon w,i,t,h  3
ted to the approaches whereby verb predicates are 6 2 2   337
6 comps np: 2 7
treated as fully formed objects which dictate the 6 "rel # 77
6synsem j loc 6 6
6cat 6 with 7
777
exact nature of their dependents). Consider the 6 4 4content 557
4 arg1 2 5
predicate load: with contact
arg2 3

(2) The peasant loaded the horses.

(3) The peasant loaded the horses on the boat. (6) is an underspeci
ed entry which gets fur-
(4) The peasant loaded the horses with hay. ther specied at parsing time when an appropri-
ate PP is attached. For instance, if a with-PP is
(2) is ambiguous between (3) and (4) each one encountered, then an interpretation according to
of which is not ambiguous. The contextual factor which the location surfaces as the direct object of
that resolves the ambiguity is the semantics of the the verb is infered.
head of the prepositional complement which here There are some theoretical reasons why we have
is taken to specify whether the direct object of the chosen to include a \pointer" to prepositional se-
verb is understood as the location and the oblique mantics rather than making it compatible (uni-
complement as the locatum or vice versa. The cru- able) to verbal semantics as Wechsler (Wechsler
cial assumption here is that prepositions have their 1994) has proposed. Firstly, if verbal and pre-
own semantics, an idea rst exploited in (Gawron positional semantics were uniable then we would
1986). not be able to explicitly state in the semantics the
We use HPSG to model our approach. (5) gives relation which each feature structure encodes as
the fragment of the type system constraining the there would be a clash of constants (relation names
values of the synsemjlocjcontjnucleus path in are constants). Secondly, identifying the semantics
the (word) description of prepositions which par- of verbs with that of prepositions does not allow
ticipate to the locative alternation phenomenon. for expressing certain types of diverse behaviour
(5) " contact # within the class of alternating verbs. For instance,
REL ? both load and stu show locative alternation, but
ARG1 content
ARG2 content only the former admits optional PP complements.
" with contact # " on contact # With to stu the interpretation under which loca-
REL with REL on _ in : : : tion is a direct object admits an optional PP com-
ARG1
ARG2
(location)
(locatum)
ARG1
ARG2
(locatum)
(location) plement (8) while the interpretation under which
locatum is a direct argument admits an obligat-
We furthermore assume that the semantics of ory one (9). Similarly, while both versions of to
the predicates include a pointer to the semantics of load are related to passive adjectives (loaded cart,
the prepositional complements they license. This loaded hay), only the \location" version is related
 contact optional obligatory inference power' is independently needed to deal,
for instance, with (10): if NP1 is a subtype of NP
on contact with contact then rule (10) will work only if trivial inference
power is available when the sequence NP1 , VP is
on con opt on con obl with con opt with con obl encountered.
(10) S ! NP, VP
Figure 1: Type system fragment encoding prepos- There are proposals in the literature which
itional alternation build on the idea of using underspecied entries.
However, several of them use additional, external
to such an adjective in the case of to stu (stued powerful mechanisms to simulate type inference.
pillow, *stued feathers). The exact treatment of (van Noord & Bouma 1994) use underspecied
such phenomena, however, goes beyond the scope verb entries and prolog delayed evaluation tech-
of our discussion here which concentrates on the niques to insert adjuncts in Dutch VPs without
use of underspecication. using lexical rules which would cater for the ne-
cessary variations of the subcategorisation list of
(8) Mary stued the pillow with feathers. verbs. In another proposal using underspecic-
(9) Mary stued the feathers into the pillow. ation (Sanlippo 1995) type inference (feature
Optionality of PP complements can also be cap- structure grounding) is simulated by relying on
tured easily with this proposal. With to load (2), an external mechanism as powerful as prolog.
(3) and (4), the PP complement is optional. The In dealing with dierent complementation pat-
grammar must have access to three dierent ver- tern phenomena, Sanlippo constructs type sys-
sions of to load, one with zero PP complements and tem fragments where the meet of the alternative
two with a PP complement participating in the al- complements is dened and subtypes verbs accord-
ternation discussed above. One approach would ing to complement types. Therefore, the informa-
involve dening two lexical rules an alternative tion about the alternation is duplicated in the type
would be to express all three possibilities directly. system as it is encoded both on the complement
Both are problematic, of course. Consider the situ- types and the verb types. The same information
ation when the grammar has two PS rules for VPs, is encoded again on a table of clauses which relate
one for discharging a NP,NP] subcat list and one a verbal \meet" type with a maximal complement
for discharging a NP,NP,PP] list. Without harm- type and a maximalverb type. Such type resolving
ing generality, assume that the bivalent version of clauses are provided for each alternation pattern.
to load is in the lexicon and two lexical rules gen- PS rules are annotated with procedures which pick
erate the trivalent versions. To process a trivalent up the correct verb type resolving clause when the
version, the parser will backtrack on the bivalent appropriate complement is encountered. Both the
version, will use a lexical rule and then, it will clauses and the searching procedures are mechan-
either succeed or it will backtrack again and use a isms external to the inferencing mechanism that is
second lexical rule. directly related with the type system. Sanlippo's
To avoid this, the following solution may be ad- approach, though powerful and exible, seems ex-
opted. First, the type system is augmented to al- travagant for phenomena like verb alternations of
low for declaring the property of being an optional the kind discussed here as well as inection phem-
or an obligatory prepositional complement, as in omena of the kind discussed in (Krieger & Ner-
gure 1. bonne 1993). In such cases the system can take
Second, a PS structure rule is introduced of the advantage of the fact that type inference can be
following sort: driven by the combination of the information that
VP ! VSUBCATNP,NP,P(optional)P]], NP is related to two separate strings (preposition and
Only one trivalent, underspecied version of to verb, verb ending and verb stem) as is exemplied
load is necessary. The parsing of a trivalent ver- in our proposal.
sion as before would involve backtracking on the Furthermore, in our approach no horizontal re-
rule dealing with optional complements but then lations exist as the lexicon contains only one entry
the rule dealing with obligatory ones would be and no other entry is ever generated. Instead, the
chosen and it would succeed anyway. single lexical entry is interpreted on the y, each
Only limited inference power is necessary for time according to well-specied constraints. Con-
this set up to work: the system must be able to sequently, no ambiguity problems result with a
infer that the unication of a subtype with its su- nice eect on parsing time. In this sense, using un-
pertype is of the type of the subtype. This `trivial derspecication dened in the type system is more
 economic than using lexical rules or a \static" ver- distinction.
sion of underspecication which is dened in the For the construction of the VP, a simple rule
lexicon. For instance, (Krieger & Nerbonne 1993) was used (VP ! V NP PP), of the following form
have used a specialised macro, the so-called dis- (14):
tributive (or named) disjunction, in a treatment (14)
of German verb inectional morphology:
While it is true that distributive disjunction ld:{sign=>phrasal:{
synsem=>synsem:{

does not add any expressive power to the sys- locl=>locl:{

cat=>cat:{
tem (though a piece of machinery, the specialised head=>HEAD,

macro, must be supported), if the macro is ever

subj=>SUBJ],
comps=>]},

called all the legal combinations are thereby gen- < 

content=>Content}}}}

erated and added to the lexicon. In this, the situ- ld:{sign=>lexical:{

synsem=>synsem:{
ation is precisely the same as with lexical rules, for locl=>locl:{

in each case, what is provided is simply a compact

cat=>cat:{
head=>HEAD=>verb:{},

representation of an ambiguity. subj=>SUBJ],

comps=>OBJ1,OBJ2]}
This can be also exemplied from the domain ld:{sign=>phrasal:{
content=>Content}}}},

of Verb Alternation phenomena. (11) will generate synsem=>OBJ1=>synsem:{

two lexical entries with an identical phon string. locl=>locl:{

cat=>cat:{

(11) 2phon l,o,a,d 

head=>noun:{}}}}}},
3 ld:{sign=>phrasal:{
synsem=>OBJ2=>synsem:{
6 2 8"   # 937 locl=>locl:{
6 >
> subj np: 1 >
> 7 cat=>cat:{
6 6 > >   _>
>77 head=>prep:{}}}}}}].
6 6 < comps NP: 2 , with PP: 3 =77
6 6cat "   77
6 6 > subj np: 1 # 77 The relevant lexical entries for the fragment
6 6 > >
>77
6 6 >
>   >
>77 were as follows. The verbal entry (load) subcat-
6 : 77
6synsem j loc 6
6
comps NP: 3 , on PP: 2
77 egorizes for a single NP subject and NP and PP
6 6 2 3 77
6
6 6 rel load 77 complements (15). This entry has underspecied
6 6 ;  77
6 6 6arg1 ;causer 7 77
1 semantics with respect to the semantic constraints
6 6content j nucleus 6 7 77
4 4 4arg2 2 location 5 57 on its second and third arguments (as suggested
;  5 in (8)). These are provided by (structure sharing
arg3 3 locatum
with) the sem constr feature of third argument,
Unlike lexical rules, our approach does not the prepositional phrase (the variable `Arg3').
face any blocking problem. A verbal pre- (15)
dicate that does not alternate (such as the
predicate to put (12),(13)), is assigned the
load ~
ld:{

apppropriate most specic semantics for its sign=>stem:{

synsemjlocjcontjnucleusjsem.cons attribute
m_PHON_LEXload],
synsem=>synsem:{
locl=>locl:{
|for to put that would be on-contact in order to cat=>cat:{

make sure that the locatum argument always sur- head=>verb:{},

subj=>synsem:{locl=>locl:{

faces as the direct object of the verb predicate. cat=>cat:{

head=>nom:{},

(12) John put his shoes on the shelf.

subj=>],

(13)
John put the shelf with his shoes.

comps=>],
spr=>]},
content=>Arg1}}],
comps=>synsem:{locl=>locl:{
cat=>cat:{

3 Implementation head=>nom:{},
subj=>],
comps=>],

The approach described in Section 2 can be imple- spr=>]},

content=>Arg2}},
mented in any environment that supports Typed synsem:{locl=>locl:{
cat=>cat:{
Inheritance because it is monotonic and demands head=>prep:{},

only `trivial inference power'. For the purposes of subj=>],

comps=>],

experimentation a grammar fragment was imple- spr=>]},

content=>Arg3}}]},
mented in the alep system - a lean formalism with content=>r_psoa:{

a simple inheritance type system, and a simple

psoa=>arg3_psoa:{
rel=>rel:{

context free rule backbone. Processing in this rel_name=>load},

arg1=>Arg1,
system is normally divided into separate structure arg2=>Arg2=>rel_psoa:{
sem_constr=>
building and feature decoration rule components, loc_alternation:{

however for our purposes no use was made of this arg2_state=>A2S}},

arg3=>Arg3=>rel_psoa:{
 sem_constr=>
loc_alternation:{
tion approach.
arg2_state=>A2S,
arg3_state=>A3S}}}}}}}}.

The prepositional entries now simply provide References

the \missing" part of the semantics, namely the Joan Bresnan (ed). 1982. The mental representa-
locum/locatum distinction: tion of grammatical relations. MIT Press
(16) Mike Calcagno. 1995. Interpreting lexical rules.
with ~ In AQUILEX Workshop on lexical rules. Cam-
ld:{
sign=>lexical:{
bridge, UK, August 9-11, 1995
m_PHON_LEXwith],
synsem=>synsem:{ Daniel Paul Flickinger. 1987. Lexical Rules in the
locl=>locl:{
cat=>cat:{ Hierarchical Lexicon. PhD. Stanford University.
head=>prep:{
pform=>with},
Jean Mark Gawron. 1986. Situations and Preposi-
tions. Linguistics and Philosophy 9, pp.327{382
subj=>],
comps=>synsem:{
locl=>locl:{
cat=>cat:{
head=>noun:{case=>acc}, Hans-Ulrich Krieger & John Nerbonne. 1993.
subj=>],
Feature-BAsed Inheritance Networks for Com-
comps=>]},
content=>inst_psoa:{ putational Lexicons. In (eds) Ted Briscoe, Va-
rel=> Rel}}}]},
leria di Paiva and Ann Copstake Inheritance,
Defaults and the Lexicon. Cambridge University
content=>inst_psoa:{
rel => Rel,
sem_constr => with_variant:{
arg2_state=>locatum, Press, 90|136
arg3_state=>locum}}}}}}.
Beth Levin. 1993. English Verb Classes and Al-
ternations. A Preliminary Investigation. The
4 Conclusion University of Chicago Press.
We have shown that horizontal redundancy is in- Walt Detmar Meurers. 1995. Towards a semantics
herent to a lexicon consisting of descriptions of for lexical rules as used in HPSG. In AQUILEX
fully formed objects. To eliminate horizontal re- Workshop on lexical rules. Cambridge, UK, Au-
dundancy, direct relations between descriptions of gust 9-11, 1995
fully formed objects must be dened externally to Steven Pinker. 1989. Learnability and Cognition:
the Typed Mulitple Inheritance Network or unin- The acquisition of argument structure. MIT
tuitive solutions must be pursued. Available im- Press
plementations of horizontal relations fail to satisfy
the reasons that dictate their implementation: the Carl Pollard & Ivan A. Sag. 1987. Information-
on-need generation of lexical entries and ecient based Syntax and Semantics. Volume 1. Funda-
parsing. Alternatively, we proposed that lexical mentals. Center for the Study of Language and
entries are descriptions of objects which allow for Information
further contextual specication of their properties Susanne Riehemann, Morphology and the Hier-
on the basis of clearly dened constraints. We archical Lexicon (ms.) Stanford University,
have shown that this is an easily implementable Stanford.
proposal even in environments with lean inference Antonio Sanlippo. 1995. Lexical Polymorphism
power and expressivity because it relies on very and Word Usage Extensibility. In aquilex
basic machinery which is available for independ- Workshop on lexical rules, Cambridge, UK, Au-
ent reasons. gust 9{11, 1995
This approach can be adopted whenever inform-
ation can be distributed among independent sur- Gertjan van Noord and Gosse Bouma. 1994. Ad-
face strings. Under the light of this proposal, many juncts and the processing of lexical rules. In
of the phenomena which have been argued in (Pol- Proceedings of the 15th International Confer-
lard & Sag 1987) to justify the horizontal related- ence on Computational Linguistics (coling),
ness approach can be viewed as dierent `inter- Kyoto, 1994
pretations' of a `core' lexical entry according to Stephen Mark Wechsler. 1994. Preposition Selec-
well-specied types of `context'. However, it must tion Outside the Lexicon. To appear in Proceed-
be noted here that this is not always a simple task. ings of the Thirteenth West Coast Conference
Roughly speaking, the less specic the contextual on Formal Linguistics
information is the more inference power and ex-
pressivity is needed to retain the underspecica-