THE ASSIGNMENT OF GRAMMATICAL RELATIONS 
IN NATURAL LANGUAGE PROCESSING 
Leonardo Lesmo, Vincenzo Lombardo 
Dipartimento di Infom~atica - Universila' di Tofino 
C.~ Svizzera 185 - 10149 Torino - ITALY 
e-mail: lesmo,vincenzo@di.unito.it 
1. Introduction 
One of the main goals of an interpreter is to map 
the syntactic descriptions found in the sentence 
into the correct roles that the elements 
(described by the nominals) play in the situation 
at hand (described by the verb). For instance, 
we must be able to state that in 
1) The cat ate the mouse 
the cat is the "eater" and the mouse is the 
"eaten thing". Of course, if we only talk about 
roles and situations we miss some significant 
generalizations. In 
2) The boy drank the water, 
if we say that the boy is the "drinker" and the 
water is the "drunk thing", we disregard the 
evident similarity of the roles of "eater" and 
"drinker" in the two situations. The notion of 
deep case arises as the common ground 
underlying a number of "apparently" different 
roles. Upon this notion some frameworks, that 
stand at the core of semantic representation and 
natural language processing, are built (see 
\[Fillmore 68\], \[Bruce 75\] and ISomers 871). 
The hard task is to devise a mapping 
between the surface descriptions and these deep 
cases. The complexity of some syntactic 
phenomena, like passivization, subject and 
object raising, long distance dependencies, has 
led many researchers to pose an intermediate 
level between the linear string of words and the 
case system. The concept involved is that of 
"grammatical relation", such as "subject", 
"direct object", "indirect object". It is claimed, 
for example, that "passivizatiou" is universally 
(cross-linguistically) explained if one says that 
the "object" of an active sentence becomes the 
"subject" in the passive form, rather than by 
saying that the NP in the VP is moved to 
replace the NP in S (that is a direct mapping). 
In the latter case it is implicit that the partictdar 
language under examination has a Subject- 
Verb-Object structure (SVO), as it usually 
happens in configurational languages such as 
English. In the example 
3a) Lo hanno visto gli amici di Piero 
(Him &tve seen the friends of Piero) 
3b) E' stato visto dagli amici di Piero 
((He) has been seen by l'iero's friends) 
the passive form does not obey tile law of direct 
mapping. The example is, however, easily 
accounted fbr by the relational theories. The 
passivization rule induces only changes of 
function: the SUBJ becomes the BY- 
complement and the OBJ becomes the SUBJ. 
The importance of grammatical relations, 
taken as primitives for a universal grammar, is 
stated by a number of formalisms often 
collected under the label of Relational 
Grammar. The problem is to map the surface 
constituents into their correct roles. With 
languages as Italian, which stands in the middle 
between configurational and freely ordered 
languages \[Stock 891 some flexibility is 
required to accomplish this task. One possibility 
is to adopt 11 neutral syntactic structure, open to 
several alternatives in the interpretation process. 
The head & modifier approach seems to feature 
this kind of neutrality, and has effectively been 
used for dealing with free word order 
languages, like the Slavonic languages \[Sgall et 
al. 861 and Finnish \[Jappinen et al. 86\]. 
The dependency formalism we have adopted 
is presented in \[Lesmo, Lombardo 91\]. An 
example is reported in fig.l, and concerns the 
sentence: 
4) La ragazza ebe lavora al guardaroba 
fu persuasa da un cliente a 
comprare una enciclopedia 
(The girl who works at the wardrobe was 
persuaded by a customer to buy an 
encyclopedia). 
The daughter nodes that stand on the left of 
their head precede it in the linear order of the 
sentence, while daughter nodes on the right 
follow it. The arcs that link the nodes in the 
dependency tree are of three types: arcs of 
structural and logical dependency (D&S arcs, 
represented by bold arrows in the figure), arcs 
of only structural dependency (STR arcs, 
simple arrows in the figure), and arcs of only 
logical dependency (DEP arcs, dashed arrows 
in the figure). D&S arcs link two words that 
stand in a "both structural and logical" relation. 
STR and DEP split these two functions of arc: 
an STR individuatcs a purely superficial 
AcrEs DE COLING-92. NAN'~S, 23-28 ^o~r 1992 1 0 9 0 PROC. or: COLING-92, NANTF.S. AUG. 23-28, 1992 
kA 
PEFISUADERE at 
RAG~2~ A 
LAVORARE ~, CLIENTE COMPRb~E 
\,oo k ./ ~ \.et 
~', .' OEld % L~ .," 
%%. 
CI-E A ENCICLOPEDIA 
GtJ,N:I\[I~V~ UNA 
J 
IL 
Fig.1 - An example of dependency tree. Because of space constraints, the figure already in- 
cludes the grammatical relations that will be described below. The bold labels (e.g. 
agt, pat) refer to deep cases. The labels immediately below them refer to the initial 
stratum of grammatical relations. The lowest labels are the last stratum (surface 
relations). SUB-Goal stands for a GOAL relation expressed via a subordinate sentence. 
Cho-1 is a chomeur (see text), expressed in Itatian via a BY (DA) complement. 
attachment, DEP represents a deep dependency 
between two words that are structurally 
independent. DEP arcs enable us to represent 
long distance dependencies, the sharing of 
dependent nodes (i.e. multiple heads, see fig. 1) 
and to represent coordinative and comparative 
constructions without violating the adjacency 
principle \[Hudson 84\], that applies only to STR 
and D&S arcs l. An arc involving dependency 
(of DEP or D&S type) is labelled with the 
grammatical relation that exists between the two 
nodes that it links (the arrangement in strata is 
explained below). 
The goal of this paper is to show that the 
formalism of Relational Grammar can be 
integrated in a useful way in a gcneral NL 
interpreter, in particular if the surface stntctures 
are represented via the dependency formalism. 
The paper examines the problems associated 
with the use of RG in an interpretive (as 
opposed to generative) framework, where the 
phase of surface relation hypothesization is 
critical. The partial configurationality of Italian 
can be exploited as heuristic information aiding 
the interpreter in selecting the preferable initial 
1 The adjacency principle intuitively states that a word 
B, that stands between the words A and C in the 
sentence, results in the santo position if we project 
Ihe related node.~ in the dependency tree onto a line. 
hypothesis. On tile contrary, the RG rules 
governing the nlappiug between strata airu at 
confirming the hypotheses: they are applied on 
the basis of the lexical and morphological 
infonnatioo associated with tile verb, where the 
lexicon provides tbe first stratum and possible 
constraints on nde applicability. 
2. The assignment of grammatical 
relations 
We start this section by providing a short 
overview of the main ideas of RG. Such ideas 
are shared by many fornlalisms, bttt we. will 
mostly refer to the work described in 
\[Perlnlutter 83\] and \[l)erhuutter, Rosen 841, 
where it can be tound a comparison with other 
RG fonnalisms. 
Grammatical relations are arranged in a 
hierarchy and are usually referred to by 
numbers: 1, which is the highest, corresponds 
to SUBJECT, 2 to DIRECT OBJEC~I ', 3 to 
INDIRECT OBJECT. The key principle of RG 
is the promotion of relations to higher levels in 
the hierarchy. The passive Cml be described as a 
promotion of 2 to 1 (i.e. DIR-OBJ to SUB J), 
leaving the previous 1 element "unemployed". 
The relation "uuemployed", which is technically 
indicated by the corresponding French word 
AcrEs DE COLING-92, NANTES. 23-28 AO(n" 1992 l 0 9 1 I)~OC. OF COLING-92, NAN'rES, AUG. 23 28. 1992 
(a) (b) 
give Mary the book John give Mary the book John 
Fig.2 - The Relational Networks associated with the sentences "John was given the book 
by Mary" (a) and "The book was given to John by Mary" (b). 
chomeur, is assigned to an element that cannot 
be involved in any other promotion. Consider 
5a) Mary gave the book to John, 
where Mary is the 1-element, the book is the 
2-element and John is the 3-element. If we 
apply the rule for passivization described 
above, the book must be promoted to the 1 
relation, Mary becomes a chomeur, while 
John is still the 3 element. The chomeur-1 
element, i.e. a chomeur element that results 
from the "unemploying" of a l-element, 
assumes the surface form of a by-complement 
in English, thus yielding 
5b) The book was given to John by 
Mary 
where Mary cannot be involved in any other 
promotion, because of its chomeur condition. A 
similar rule applies to double-accusative 
constructions, as shown in fig.2a. 
At the same level of promotional rules we 
can posit the lexical rules, that account for the 
determination of grammatical relations within 
subordinate untensed sentences, as in 4). Such 
information is stored within the lexical entry of 
the verb that governs the subordinate clause. 
For example, to promise forces the SUBJ of 
the subordinate clause to be the SUBJ element 
of the governing clause, as we can see in 
6) Mary promised John to write him a 
letter, 
where the SUBJ of write is Mary, the same of 
promise. On the contrary, to persuade 
forces the SUBJ of the subordinate to be the 
OBJ element of the governing clause, as in 4). 
It must be noted that the lexical rules are related 
to the assignment of relations in the initial 
stratum, even if they are subsequently changed 
by promotional rules. For example, in 
7) The girl was persuaded by a 
customer to buy an encyclopedia, 
the girl is the element which is still shared by 
the two clauses, even if it is the SUBJ now. 
The semantic interpretation process takes 
advantage of the functional analysis, i.e. the 
analysis in terms of grammatical functions: 
relational structures are easily mapped onto 
logical representations, because of the 
resemblance between a Relational Network and 
a Predicate-Argument structure: the initial 
stratum states which are the grammatical 
relations (actually the elements at the sentence 
level) that act as arguments of the predicate 
identified by P. 
From a computational point of view, 
syntactic and semantic clues must be taken into 
account, in order to map the grammatical 
relations onto the surface descriptions. The 
mapping is carried out incrementally, i.e. as 
soon as the nominal head of the complement is 
parsed: it is highly language-dependent, and 
considers features like inflectionality, 
configurationality and deep underlying 
structures. The mapping must also take into 
account the changes on the surface form that are 
induced by the rules on grammatical relations, 
discussed informally above 2. 
Unfortunately the mapping raises some 
difficulties. The bias of the rules at the RG level 
is of a generative kind. Rules start from the 
initial stratum, the one which is closer to the 
deep cases (arguments of the predicate), to 
produce the final stratum of the surface 
arrangement; on the contrary, in an interpreter 
of language, the task is to trace what rules have 
been actually applied (and in what order) to the 
initial stratum (and the subsequent strata) in 
order to achieve the surface realization of the 
grammatical relations. Useful heuristics are 
devised to identify the surface clues that 
evidence tile application of a particular nile. For 
example, a passivization is accompanied by a 
passive form of the verb. 
Starting from the surface descriptions, the 
2 Of course, the application of such rules involves also 
changes in focus. Two expressions that result to be 
derivable from each other (3a and 3b). according to 
phenomena that are explained in terms of grammatical 
relations, are therefore not strictly equivalent, even if 
both of them involve the same roles to be played by 
the individuals in the ground sentence (or, better, in 
the sentence that has been claimed to be ground). 
Acids DE COLING-92, NANTES. 23-28 AO~r 1992 1 0 9 2 PRec. OF COLING-92. N^r,n'I~S. Auo. 23-28, 1992 
interpreter is not always able to uniquely 
determine the assignment of grammatical 
relations on the basis of syntactic features. 
Consider 
8a) Giovanni il vino Io ha bevuto 
(John the wine \[it\] has drunk) 
8b) II vino Giovanni Io ha bevuto 
(The wine John \[it\] has drunk) 
Only semantics allows a hearer to realize that in 
both cases the "drinker" is John and not the 
wine. Hence a flexible interaction between 
syntactic and semantic information (selectioual 
restrictions) must be devised. 
Since in our system the analysis is 
incremental, in the sense that parsing and 
interpretation are synchronous processes, as 
soon as the dependency tree is extended with a 
head of a substructure the semantic interpreter is 
triggered to interpret it: in the case of verbs, all 
the complements that precede the verb are 
interpreted when the verb is found, while each 
complement that follows the verb is interpreted 
as soon as it is attached to it. 
The association of the grammatical relations 
with the descriptions in the sentence is 
accomplished by the rules at the relational level 
(GR rules), which are divided into three 
groups: the first of them deals with the initial 
proposal of relations based on syntactic features 
(Proposal Rules - PR), the second concerns the 
movement across the strata (Stratal Rules ~ SR), 
and the third, of a lexical kind (then Lexical 
Rules - LR), accounts for the sharing of 
relations in unteused subordinate clauses, as in 
4. The verbal lexical entry contains, among 
other information, its initial stratum of 
grammatical relations. Once the verb has been 
found, the GR rules are triggered iu order to 
find out the roles that are played by the elements 
that precede the verb in the input sentence 
(incremental interpretation). It is the actual input 
that determines which of the three groups must 
be applied. For example, if we have a single 
active sentence (without subordinate embedded 
clauses or passive forms), the Proposal Rules 
are triggered. SR and LR rules are activated 
only in presence of special features: a passive 
form (was eaten), for instance, activates the 
Passivization rule (belonging to SR), if we have 
the pair <SUBJ,OBJ> in the current stratum, 
while lexical rules are associated with verbs that 
govern subordinate clauses (e.g. t o 
persuade). The result of the application of one 
or more rules is the final stratum, against which 
the assignment of relations guessed by the PR 
group is nmtched. 
In the PR, the first feature that is taken into 
account is the syntactic form of the participants. 
SUBJ and DIR-OBJ require that the 
corresponding nominals are not preceded by a 
preposition, and prououns be inflected 
appropriately. For example, in 3a the pronoun 
I0 features an accusative case, thus a DIR-OBJ. 
If two nominal descriptions are not inflected 
and, hence, they cannot be associated with 
a particular relation via this marking, as in 
9) 11 gatto mangio' ii topo 
(The cat ate the n~)use), 
the position of the nominals can be useful, 
since, in a partially configuratioual language 
such as Italian, grammatical relations are 
usually connected with the canonical positions 
of the SVO order: with a transitive verb, the 
SUBJ precedes the verb and OBJ follows it; 
witll iutransitive verbs, the position of the 
nominal without a preposition does not affect 
the grammatical relation assigned to, since it 
will be surely the SUBJ wherever it is. If the 
order too does not give an unambiguous 
assignment of grammatical relations, the last 
resources are the number agreement for the 
SUBJ relation and the semantic check. In 
situations such as 
10) Le ragazze Giorgio le ha viste 
(The girls Giorgio \[them\] has seen) 
even if the nominal descriptions staud on the 
same side of a transitive verb, the latter agrees 
only with Giorgio in number. On the contrary, 
only semantics can solve a situation as 8; 
moreover, the semantic check can also reject an 
assignment made on the basis of the syntactic 
features that we have described. Consider, for 
example, the sentence 
11) Un snsso calcio' il vitello 
(A rock kicked the calf) 
Even if the order rules assign an sasso the 
SUBJ relation and il vitello the OBJ, such an 
assignment is rejected on the semantic ground. 
Notwithstanding a system that works correctly 
cannot be based only on semantics, since a 
sentence like il sounds really strange to a 
native speaker, if we are not in a particular 
focussing situation. 
3. An example 
In figure 1, we can find the result of the 
interpretation of sentence 4. When the analysis 
arrives at lavora, in the relative sentence, its 
initial stratum <SUB J> is retrieved from the 
lexicon. Since ehe (who) is a nominal without 
a prepositional marker, it (or better the element 
refelTed to) is the SUBJ of lavora, as stated by 
the Proposal Rules. Lavora has also an 
adjunct, al guardaroba (at the wardrobe), 
a non-teml relation of type LOC. The structure 
for the nominal description la ragazza che 
lavora al guardaroba has already been built, 
ACTES DE COLING-92, NArCrES. 23-28 Aot~q 1992 l 0 9 3 PROC. Ot: COLING-92. NAN'IF.S. At ¢;. 23k28 1992 
when the input word is the verb persuadere. 
Its lexieal entry provides the parser with an 
initial stratum of grammatical relations that 
consists of: a SUBJ, an OBJ and a subordinate 
sentential Goal (SUB-Goal) 3, i.e. a persuader, 
a persuadee and the persuasion. This basic 
assignment can be related to the deep cases of 
AGT, PAT and GOAL respectively. Moreover, 
a lexical rule is contained in the lexical entry: 
The SUBJ of the subordinate untensed clause 
governed by persuadere is the OBJ element 
of the governing clau.~e. 
Since the verb is in the passive form and the 
current stratum features a SUBJ and an OBJ 
relations, the Passivization rule in the SR group 
is triggered, in order to find the actual 
arrangement of relations in the input sentence. 
The new stratum is <SUBJ, Cho-I,SUB- 
Goal>, against which the proposals made by 
the PR group are matched. Since the nominal 
description already found is not inflected and is 
not marked by a preposition, the positional 
rules suggest that, since it precedes the verb and 
agrees with it, a possible assignment of relation 
is SUBJ. The semantic check, which is 
activated on tbe basic relation (i.e. OBJ) 
validates such an assignment, because a girl that 
works at the wardrobe may happen to be 
persuaded. The analysis proceeds to the next 
nominal description, with the set of relations 
(Cho-1, SUB-Goal) not assigned yet. Da un 
eliente (by a customer) has exactly the 
form of a Cho-1 in Italian. The Proposal Rules, 
whose hypothesis is confirmed by the semantic 
check, are sufficient to deal with this situation. 
When we find the verb comprare (buy), the 
PR group assigns to such a description the 
SUB-Goal relation and consequently the lexical 
rule associated with persuadere assigns the 
SUBJ relation of the initial stratum of 
comprare to la ragazza the .... The initial 
stratum of comprare features also an OBJ 
relation, that will be assigned to eneiclopedia, 
when it is found. 
The completeness of the set of found 
grammatical relations is checked when the node 
corresponding to the verb is "closed", i.e. when 
it cannot have further modifiers. 
4. Conclusions 
The paper illustrates how RG can be used to 
map in a principled way surface dependency 
relations into thematic roles. 
3 '/'he Goal relation, such as Instrument or Location, is 
a non-term relation and participates only to special 
kinds of promotional rules (see \[Perlmutter 83\] for 
details). 
The main feature of the approach is the strict 
cooperation among different knowledge sources 
(lexicon, RG rules and semantics) in carrying 
out the task: this cooperation is made necessary 
by the partial configurationality of Italian, 
where the ordering of constituents can only be 
considered as the basis for plausible 
suggestions, but not as the source of stricts 
constraints. The adoption of an unmarked input 
(an unlabelled dependency tree) makes available 
a flexible starting point that leaves the RG 
module the task of making the required 
inferences. 
The ideas expressed herein are implemented 
in the GULL system (see \[Lesmo, Torasso 83, 
85al for the syntactic part \[Di Eugenio, Lesmo 
871 for the basic ideas about semantics): both 
levels of grammars are represented via 
condition-action rules. The system is 
implemented in Common Lisp and runs on 
SUN workstations. 
References 
\[Bruce 75\] Bruce B., Case Systems for Natural Lang- 
uage, Artificial Intelligence 6, 1975, pp, 327-360. 
IDi Eugenio, Lesmo 87\] B.Di Eugenio, L.Lesmo, 
Representation and Interpretation of Determiners in 
Natural Language, Proc. 10th 1JCAI, Milano, Italy, 
1987, pp.648-654. 
IFillmore 68\] Fillmore C., The case for case, in 
"Universal in Linguistic Theory", (Bach and Harms 
eds.), Holt, Rinehart and Winston, 1968. 
\[l-tudson 84\] Hudson R., Word Grammar, Basil 
Blackwell, Oxford, 1984. 
\[Jappinen et al. 861 Jappinen H., Lehtola A., Valkonen 
K., Functional Structures for parsing dependency 
constraints, Proc. COLlNG 86, Bonn, Germany, 
1986, pp.461-463. 
\[Lesmo, Torasso 83\] Lesmo L., Torasso P., A Flexible 
Natural Language Parser based on a two-level Re- 
presentation of Syntax, Proceedings of the 1st Con- 
li~rence ACL Europe, Pi~, Italy, 1983, pp.114-121. 
\[Lesmo, Torasso 85a\] Lesmo L., Torasso P., Analysis 
of Conjunctions in a Rule-Based Parser, Proceedings 
ACL 85, Chicago, USA, 1985, pp.180-187. 
\[Lesmo, Lombardo 91\] Lesmo L., Lombardo V., A 
Dependency Syntax for the Surface Structure of 
Sentences, Pro,:. of WOCFAI, Paris, July 1991. 
\[Perlmutter 83\] Perlmutter D.M. (ed.), Studies in Rel- 
ational Grammar 1, Univ. of Chicago Press, 1983. 
\[Perlmuner, Rosen 84\] Perlmutter D.M., Rosen C.G., 
(eds.), Studies in Relational Grammar 2, The Univ. 
of Chicago Press, 1984. 
\[Sgall et al. 86\] Sgall P., Haijcova E., Panevova J., 
7"he Meaning of the Sentence in its Semantic and 
Pragmatic Aspects, D. Reidel Publ. Co., 1986. 
\[Shiners 87\] Somers H.L., Valency and Case in Comp- 
utational Linguistics, Edinburgh Univ. Press, 1987. 
\[Stock 89\] Stock O., Parsing with Flexibility, Dynamic 
Strategies, and Idioms in Mind, Computational 
Linguistics 15, 1989, pp.l-18. 
ACl'l~.S DE COLING-92, NANTES, 23-28 Aot;r 1992 1 0 9 4 PRoc. OF COL1NG-92, NANTES, AUcI. 23-28, 1992 
