Selectional Restrictions in HPSG 
Ion Androutsopoulos Robert Dale 
Sottware and I(nowledge Engineering Laboratory Language Tcclmology Group 
Institute of Intbrmatics and Telecommmfications Department of Comlmting 
National Centre for Scientific Macquarie University 
Research "Demokritos" Sydney NSW 2109, Australia 
153 10 Ag. Paraskevi, Athens, Greece. e-maih Robert.Dale@mq. edu. au 
e-maih ionandr@±it, demokritos, gr 
Abstract 
Selectional restrictions arc semantic sortal con- 
straints ilnposed on the particil)ants of lin- 
guistic constructions to capture contextually- 
dependent constraints on interpretation. De- 
spite their linfitations, selectional restrictions 
have t)roven very useflfl in natural bmguage ap- 
pli('ations, where they have been used frequently 
in word sense disambiguation, syntactic disam- 
biguation, and anaphora resolution. Given their 
practical wtlue, we explore two methods to in- 
corporate selectional restrictions in the HPSG 
theory, assuming that the reader is familiar with 
HPSG. The first method eml)loys ItPSG~S BACK- 
GROUND feature and a constraint-satisfaction 
comt)onent t)il)e-lined after the parser. The 
second method uses subsorts of retbrential in- 
dices, and 1)locks readings that violat(', sole(:- 
tional restrictions during parsing. While the- 
oretically less satisfactory, we have Ibund the 
second method particularly useflfl in the devel- 
opment of practical systems. 
1_ Introduction 
~lPhe term selectional restrictions refers to se- 
mantic sortal constraints imposed on the 1)ar - 
ticipants of linguistic constructions. Selectional 
restrictions arc invoked, for example, to account 
tbr the oddity of (1) and (3) (cf. (2) and (4)). 
(1) ?Tom ate a keyboard. 
(2) Tom ate a banana. 
(3) ?Tom repaired the technician. 
(4) Tom repaired the keyboard. 
~IPo account tbr (1) and (2), one would typically 
introduce a constraint requiring the object of 
"to eat" to denote an edible entity. The odd- 
ity of (1) can then be attributed to a violation 
of this constraint, since keyboards are typically 
not edible. Silnilarly, in (3) and (4) one could 
postulate that "to repair" can only be used with 
objects denoting artifacts. This constraint is vi- 
olated by (3), because technicians are typically 
persons, and persons are not artifacts. 
We note that selectional restrictions attempt 
to capture contextually-dei)endent constraints 
on interpretation. There is nothing inherently 
wrong with (1), and one can think of special 
contexts (e.g. where Tom is a circus pertbrmer 
whose act includes gnawing on comtmter pe- 
ripherals) where (1) is felicitous. The oddity 
of (1) is due to the fact that in most contexts 
l)eople do not eat keyboards. Similarly, (3) is 
ti;licitous in a science-fiction context where the 
technician is a robot, |rot not in most usual con- 
texts. Selectional restrictions are typically used 
to capture flints about tlm world which are gen- 
c.r~lly, but not necessarily, true. 
In w~rious forms, selectional restrictions have 
been used tbr many years, and their limitations 
are well-known (Allen, 1995). For example, 
they cmmot account tbr lnetaphoric uses of lan- 
guage (e.g. (5)), and they run into 1,roblen,s in 
negated sentences (e.g. unlike (1), there is noth- 
ing odd about (6)). 
(5) My car drinks gasoline. 
(6) Tom cannot cat a keyboard. 
Despite their limitations, selectional restric- 
tions have proven very useflfl in practical appli- 
cations, and they have been employed in sev- 
eral large-scale natural  understanding 
systems (Martin et al., 1086) (Alshawi, 1992). 
Apart fl'om blocking pragmatically ilMbrmed 
sentences like (1) and (3), selectional restric- 
tions can also be used in word sense disanfl)igua- 
tion, syntactic dismnbiguation, and anaphora 
15 
resolution. In (7), for example, tile '))~qnter" 
refers to at computer peripheral, while in (8) it 
refers to a person. The correct sense of "printer" 
can be chosen in each case by requiring the ob- 
ject of "to repair" to denote an artifact, and the 
subject of "to call" (when referring to a phone 
call) to denote a person. 
(7) Tom repaired the printer. 
(8) The printer called this mornfilg. 
Silnilarly, (9) is from a syntactic point of view 
potentially ambiguous: the relative clause may 
refer either to the departments or the employ- 
ees. The correct reading can be chosen by speci- 
fying that the subject of "retire" (the relativised 
nominal in this case) must denote a person. 
(9) List tile employees of the overseas depart- 
nlents that will retire next year. 
Given tile value of selectional restrictions in 
t)ractical applications, we explore how they can 
be utilised in the HPS~ theory (Pollard and 
Sag, 1994), assuming that the reader is familiar 
with HPSG. Onr proposals are based on expe- 
rience obtained from using IIPSG in a natural 
 database interface (Androutsot)oulos 
et al., 1998) and a dialogue system for a mobile 
robot. To the best of our knowledge, selectional 
restrictions have not been explored so far in the 
context of HPSG. 
We note that, although they often exploit 
similar techniques (e.g. semantic sort hierar- 
chies), selectional restrictions costitnte a differ- 
ent topic from linking theories (Davis, 1996). 
Roughly speaking, linking theories explore the 
relation between thematic roles (e.g. agent, pa- 
tient) and grammatical thnctions (e.g. subject, 
complement), while selectional restrictions at- 
tempt to account tbr the types of world entities 
that can fill the thematic roles. 
We discuss in sections 2 and 3 two ways that 
we have considered to incorporate selectional 
restrictions into HPSO. Section 4 concludes by 
comparing briefly the two approaches. 
2 Background restrictions 
The first way to accommodate selec- 
tional restrictions in HPSG USeS the 
CONTEXTIBACKGROUND (abbreviated here 
as CXlBG) feature, which Pollard and Sag 
(Pollard and Sag, 1994) reserve tbr "Micity 
conditions on the utterance context", "presup- 
positions or conventional iml)licatures", and 
"at)prot)riateness conditions" (op cit pp. 27, 
332). To express selectional restrictions, we add 
qfpsoas (quantifier-flee parameterised states 
of atfairs) with a single semantic role (slot) 
iI1 CXIBG. 1 For exanlple, apart fi'om the eat 
qf'psoa in its NUCLEUS (NUt), the lexical sign 
tbr "ate" (shown in (10)) would introduce an 
edible qfpsoa in Bo, requiring \[\] (the entity 
denoted by tile object of "ate") to be edible. 
(10) "PIION (ate) 
CAT 
SS I LOC CONT \[ NUC 
cx I ~(~ 
\[ I,EAD lcTb l\]\] 
COMPS ( NP\[~J >ill 
c.t \[,,~A'r,~N FlJ // 
In the case of lexical signs for proper names 
(e.g. (11)), the treatment of Pollard and Sag 
inserts a naming (namg) qfpsoa in BG, which 
requires the BEARER (BRER) to by identifiable 
in the context by means of the proper name. 
(11) also requires the bearer to be a man. 
(11) "PIION <ToTrt> 
oa,l, \[,,,:AD 
CONT \[INDEX \[RESTR ~}\] 
Tile ItPSO t)rinciples that control the propa- 
gation of the BG feature are not fully developed. 
For our purposes, however, tile simplistic prin- 
co)Ic of contextual consistency of Pollard and 
Sag will suffice. This principle causes the BG 
value of each phrase to be the union of tile BG 
values of its daughters. Assuming that the lex- 
ical sign of "keyboard" is (12), (10)-(12) cause 
(1) to receive (13), that requires \[\] to denote an 
edible keyboard. 
1To save space, we use qfpsoas wherever Pollard and 
Sag use quantified psoas. We also ignore tense and as- 
pect here. Consult (Androutsopoulos et al., 1998) for 
the treatment of tense and aspect in our ltpsG-based 
database interface. 
16 
(12) 
(13) 
/ \[INI)EX \[\] 
"PlION 
SS I LOC 
(~/bm, ate, a, keyboard) 
-II\],;AI) verb 1 
CAT SUIL1 <> \[ 
COMI'S () J 
-QUANTS ( keybd \[INST \[~\]\])\] \[":a"E" Hm \] 
J CONT i NUC k I~A'rI':N eat 
|NAME :/bin|' / namg ~ " l> 
\[ 
\[lNS'l' \[~l / edible L J 
A(:('ording t,, (13), to accept (1), one has to 
place it; ill ;~ st)ecial conte.xt where edible key- 
bonrds exist (e.g. (1) is thlMtous if it reM's to ;~ 
miniature cho(:ol~te keyt)oard). Su(:h (:ontexts, 
however, are rare, ;rod hen(:e (1) sounds gen- 
erally odd. Alternatively, one has to relax the 
B(~ constraint that the keyboard 1111181; BC edi|)le. 
We assmne that special contexts ~dlow t)a.rticu - 
l~r BG constraints to be relaxed (this is how we 
wouht a(:(:omlt fin" the use of (1) ill ~L circus con- 
text), \])ut we (Io not ll;we any t'ornl~d lne(:hanisnl 
to sl)e(:itly exactly when B(~ (:(mstr;dnts ('~m l)e 
relnxcd. 
Similnr connnents apply to (3). Assuming 
th;Lt the sign of "req)aired" is (ld), nnd that 
the sign of "teclmi(:iml" is similar to (12) ex- 
cept that it; introdu('es a technician index, (3) 
receives a sign theft requires the repairer to 1)e a 
technician who is all artifact. U k~(:hnicians, how- 
ever, are generally not artifacts, which accounts 
for the oddity of (3). 
(14) -PIION <repaired) 
l / // 
L(.MP  < NI'  >J // 
"3SIL()C CONTINuc repair \[IIEPAI,\[I,3,) ~\]jl I 
Let us now (:onsider \]lOW it (:omtmter sysl;em 
~o,,ld ~e(:o,ll~t for (~)-(~). For ex~ulU,>,, how 
entity 
abstract physical 
animate edible inanimate "" 
"'" lllall tech.ician '/ X~ V "" keybd 
son animal non arfct ~lct 
edible animal edible_non afcl 
• " male tech "" 
......... ballalla 
Figure 1: A simt)listic semantic hierarchy 
WOilkl the system fig.re out fl'om (13) that (:1) 
is pragm~tieally odd? Among other things, it 
wouhl need to know that keyl)o~r(ts ~rc not edi- 
ble. Similarly, in (2) it would need to know that 
|)~m~m~s are edible, ~md in (3) (d) it; would need 
I;() 1)e nwarc that technicians are. llot artifacts, 
while keyboards m:e. Systenls that employ se- 
lc(:tionnl restri(:tions usunlly encode knowledge 
of this kind in the. fol:nl of sort hierarchies of 
worhl entii;ies. A siml)listic exmnt)le of such n 
hierm:chy is det)i('ted ill tigure 1. The hierarchy 
of tigure \] shows thnt nil lllell &lid technicians 
are 1)ersons, all 1)ersons are ~tniln;~|;e entities, all 
aninlate entities are t)\]lysi(:al ol)je(:ts, mitt so on. 
Some (1)ut not all) persons are 1)oth teehni(:ians 
:rod lnen at the same time; these t)ersons are 
nmml)ers of I;he male_tech sort. Similarly, all 
l)mlmms are edil)h; ;rod liot artifacts. No person 
is e(lil)le, because the sorts person and edible 
h~we no (:onnnon su|)sorts. 
It is, of course, extremely difficult to con- 
stru('t hierm'chies th~Lt include all the sorts of 
world entities. Ill natural bmguage systenls that 
target sl)ecifi(: and restricted dolmfins, however, 
constructing such hier;~rchies is feasible, because 
the relevant entity sorts and the possible hi- 
erarchical reb~tions between them are limited. 
In naturM lmlguage database interfimes, tbr ex- 
ample, the relevant entity sorts and the rela- 
tions between theln nre often identilied dur- 
ing the, (tesing of the database, in the tbrm 
of entity-relatiolMli 1) diagrams. We also note 
l;h;~t large-scah; smmmtic sort hierarchies are al- 
ready ill use ill artiticinl intelligence ~md natural 
17 
 gener~tion projects (tbr example, Cyc 
(Lenat, 1995) and KPML'S Upper Model (Bate- 
man, 1997)), and that the techniques that we 
discuss in this paper are in principle compatible 
with these hierarchies. 
To decide whether or not a sentence violates 
any selectional restrictions, we collect from the 
CONT and BO features of its sign ((13) in the 
case of (1)) all the single-role qfpsoas for which 
there is a sort in the hierarchy with the same 
name. (This rules out single-slot qt~)soas in- 
troduced by the CONTs of intransitive verbs.) 
The decision can then be seen as a constraint- 
satisthction problem, with the collected qfpsoas 
acting as constraints. (15) shows the constraints 
tbr (1), rewritten in a tbrm closer to predicate 
logic. HPSG indices (the boxed nmnbers) are 
used as variables. 
(15) kcybd(~\]) A man(~) A edible(~\]) 
Given two contstraints cl, c2 on the same vari- 
al)le, c~ subsumes c2 if the corresponding hier- 
archy sort of cl is an ancestor of that of c2 or 
if cl = c2. c~ and c2 can be replaced by a new 
single constraint c, if cl and c2 subsume c, and 
there is no other constraint d which is subsumed 
by cl,c2 and subsumes c. c and c' must be con- 
straints on the same variable as ct, c2, and must 
each correspond to a sort of the hierarchy. If the 
constraints of a sentence can be turned in this 
way into a tbrm where there is only one con- 
straint fbr each variable, then (and only then) 
the sentence violates no selectional restrictions. 
(15), and cdil, ,'am ot be re- 
p\]aced by a single constraint, because keybd and 
edible have no common subsorts. Hence, a se- 
lectional restriction is violated, which accounts 
for the oddity of (1). In contrast, in (2) the 
constraints would be as in (16). 
(16) banana(~) A man(m) A edible(~\]) 
banana(~\]) and cdible(F~) can now be re- 
placed by banana(F~), because both subsume 
banana(~\]), and no other constraint subsmned 
by both banana(~\]) and cdible(~) subsulnes 
banana(~). This leads to (17) and the conch> 
sion that (2) does not violate aw selectional 
restrictions. 
(17) banana(E\]) A man(D\]) 
This constraint-satisfaction reasoning, how- 
ever, requires a set)arate inferencing component 
that would be pipe-lined after the parser to rule 
out signs corresponding to sentences (or read- 
ings) that violate selectional restrictions. In the 
next section, we discuss an alternative approach 
that allows hierarchies of world entities to be 
represented using the existing HPSG framework, 
and to be exploited during parsing without an 
additional inferencing component. 
3 Index subsorts 
HPSG has already a hierarchy of feature struc- 
ture sorts (Pollard and Sag, 1994). This hierar- 
chy can be augmented to include a new part 
that encodes intbrmation about the types of 
entities that exist in the world. This can be 
achieved by partitioning the ref HPSO sort (cur- 
rently, a leaf node of the hierarchy of feature 
structures that contains all indices that refer 
to world entities) into subsorts that correst)ond 
to entity types. To encode the information of 
figure 1, rEf would have the snbsorts abstract 
and physical, physical would have the subsorts 
animate, edible, inanimate, and so on. That 
is, referential indices are partitioned into sorts, 
mid the indices of each sort can only be an- 
chored to world entities of the corresponding 
type (e.g. keybd indices can only be anchored 
to keyboards). 
With tiffs arrangement, the lexical sign for 
"ate" becomes (18). The Bo edible restriction 
of (10) has been replaced by the restriction that 
the index of the object must be of sort edible. 
(18) -PIION (at8) 
LCOVpS ( N~'CN 
L c°NTI \[I~aTEN INedible eat 
Similarly, the sign for "Tom" becomes (19) (cE 
(11)), and the sign for "keyboard" introduces an 
i,dex of sort k vbd as shown in (9O) (cf. (12)). 
(19) "PITON (Tom) o,T no,4 \]\] 
CONT \[,mSTI~ (} J l/ 
18 
\[ rm/N 1,:~fl, oa',.d) \] 
/ ' ' \[I{.I,,'STIt {} ill \[ ss I J,oc 
\[cxJ BG {} JJ 
Unification of indices pro(:eeds in the, s;lille 
maturer as unificatioll of all other typed feature 
structm:es ((Jarlienter , 1!)!/2). 'Fhe parsing of 
(\]) iIOW fails, 1)ecause it al, te, nq)ts to unilly an 
il dox or (i,lt,:o,hl(::ed t/y with 
an index of so,*; t,:eybd (introduced t,y (20)), and 
no Ill'SO sorl; is sul)sumed l)y both. in (:ontrast, 
the parsing o17 (2) would su(:('eed, because the 
sign of "bmuma" would introduce an index of 
sort banana, which is a sut)sort of edible (Iigur(~ 
1); hence the two indi(:es can 1)(', ratified. (3) and 
(4) would l)e l)ro('essed sinfilarly. 
in (7) and (8), there would 1)e two lcxi(:nl 
signs for "ln'illtcr": one inl;ro(lu('ing ml index of 
sort pri'nter_pe'r.s'o'n, and one im:o(lu(:ing an in- 
dex of sort pri'nte'r_periph,(~'ral. (printe'r4)er.~'on 
and l)rinter_periph, cral would t)e daughters of 
person and art'@tel respectively in tigure 1.) 
The sign for "repairc, d", would require the index 
of its ol)je(:t to be of sort arl,'l\[fact, and l;he sign 
of "(:ail(~d" wou\](l re(tuire its sul)je('l; index to t)e 
of sort per,so'n. This (:orre(:tly admits only the 
reading where the rel)aire(l entity is a (:Omlml;er 
peripheral, ml(t l;tm (:aller is ;t t)(',rson. Simil~tr 
llleC\]iallisnls (;;/,ll })e llse(t to (l(~,\[;(!lTillille t;tlP, (;of 
reel; reading of (9). 
With the al)proa(:h of this see, lion, it; is also 
possible to speciily seh;ctional restrictions in the 
declarations of qflIsoas in the Ill'SO hierarchy of 
feature structures, as shown in tigure 2, rather 
than in the lexi(:on. 2 When the same qft)soa is 
used in several lexical signs, this saves having to 
repeat tile same, selectional restrictions in each 
one of the lexical signs. For example, the verbs 
"rq)air" and "Iix" iiiay both introduce a repair 
qfpsoa. The restriction that the repaired entity 
must be an artifact can lie sl)eeified once in the 
declaration of repair in the hierarchy of feature 
structures, rather than twice in the lexieal signs 
of ~'cl)air" and "fix". 
2Additional layers can be included betwc,(m qfpsoa 
and the leaf sort;s, as sketched in section 8.5 of (Pollard 
and Sag, 1994), to group together qfpsoas with common 
selnalltifi roles. 
qfDso(  
\[EATI.H\[ (t,~i~r~ttt(;\] ... \[I~.E1)AIIIEIt l,('.7".~O,t \] 
cat\[EATI,ZN edible \] ,v.pctirkRl.ZPAnU.:D artiJhct\] 
Figure 2: Declarations of qfpsoas 
4 Conclusions 
We have presented two met;hods to incorpo- 
rate selectional restrictions ill IlPSG: (i) express- 
ing selectional restrictions as BACKGROUND con- 
straints, and (it) enq)loying subsorts of referen- 
trial indices. The first method has the advantage 
that it requires no modification of the cmTent 
IIPSO feature structures. It also lnzdntains Pol- 
lard and Sag's distinction bel;ween "literal" mid 
"non-literal" meaning (expressed t)y CeNT and 
I~ACKGI/OUN\]) respe, ctively), a distinction whi('h 
is lflm'red in the second approach (e.g. nothing 
in (18) shows th~lt requiring the obje('t to denote 
an edil)le entity is part of the non-literal mean- 
ing; of. (10)). Unlike the tirst method, however, 
the second apt)roach re(tuires no additional ta- 
ft;renting comtionent tbr determining when se- 
lecl;ional restrictions h~tve been violated. With 
sentences that contain several potentially aiil- 
lfiguous words or phrases, t;11(,, second apl)roat:h 
is also more etlicienl;, ~ls it blocks signs that 
violalx', selectionnl testa'tel;ions during parsing. 
In the tirsl; aplm)ach, these signs remain un- 
detected during parsing, and they may have a 
multiplicative effect, h;ading to a large nmnber 
of parses, which then have to l)e checked individ- 
ually by the taft;renting component. We have 
timnd the se(:ond at)l)roach t)articularly useful 
in the develolnnent of practical systems. 
There is a deet)er question here al)out the 
proper place to maintain the kind of intbrnla- 
lion encoded in selectional restrictions. The 
applicability of selectional restrictions is always 
context-dependent; and for any selectional re- 
striction, we can ahnost always find a context 
where it does not hold. Our second method 
above effectively admits that we cromer develop 
a general tmrlIosc solution to the problem of 
meaning interprel;ation, and that we have to at- 
cept that our systems alwws operate in specific 
contexts. By committing to a particular con- 
text of interpretation, we 'compile into' what 
was traditionally thought of as literal meaning a 
19 
set of contextually-determined constraints, and 
thus enable these constraints to assist in the 
HPSG  analysis without requiring an 
additional reasoning component. We take the 
view here that this latter approach is very ap- 
propriate in the construction of real applications 
which are, and are likely to be ibr the tbresee- 
able future, restricted to operating in limited 
domains. 

References 

J.F. Allen. 1995. Natural Language Under- 
standing. Benjamin/Cmnmings. 

H. Alshawi, editor. 1992. The Core Language 
Engine. MIT Press. 

I. Androutsopoulos, G.D. Ritchie, and 
P. Thanisch. 1998. Time, Tense and 
Aspect in Natural Lmlguage Database 
Interfaces. Natural Language Engineering, 
4(3):229-276. 

J.A. Batenlan. 1.997. Enat)ling Technology for 
Multilingual Natural Language Generation: 
the KPML Development Environment. Nat- 
ural Language Engineering, 3(1):15-55. 

B. Carpenter. 1992. The Logic of Typed Feature 
Structures. Number 32 in Canlbridge ~h'acts 
in Theoretical Computer Science. Cambridge 
University Press. 

T. Davis. 1996. Lczical Semantics and Link- 
ing in the Iticra~'chical Lezicon. Ph.D. thesis, 
Stanford University. 

D.B. Lenat. 1995. CYC: A Large-Scale Invest;- 
merit in Knowledge Infl'astructure. Cornm'a- 
nications of ACM, 38(11):33-38. 

P. Martin, D. Appelt, and F. Pereirm 1986. 
Transt)ortability and Generality in a Natural- 
Language Interface Systein. In B. Grosz, 
K. Sparek Jones, and B. Webber, editors, 
Readings in Natural Language PTvcessing, 
pages 585 593. Morgan KaufmamL 

C. Pollard and I.A. Sag. 1994. Head-Driven 
Phrase Structure Grammar. University of 
Chicago Press and Center tbr the Study of 
Language and Information, Stanford. 
