SYNCtlRONOUS TAGs AND FRENCtt PRONOMINAL CLITICS 
Anne ABEILLE 
Ddpartement de sciences du langage, 
Universitd de P~u'is 8, 
2 me de la libertd, 93526 Saint Denis 
abedle@ frpSv I l.bimet 
Introduction 
Lexicalized Feature-based TAGs have been used for 
describing various syntactic phenomena across several 
languages, and for writing sizable grammars fer French 
and English (Abeilld 1988 and 199la, Abeilld et al. 
1990) which have been incorporated into a fully 
implemented pursing systenl (Schabes 1990) 1. The main 
linguistic advantages of such a forlnaiisnl are that: 
- it provides u larger donlain of locality than CFG-based 
formalisms such as HPSG or LFG, 
- it allows factoring recursion from tile domain of local 
dependencies, 
- as a consextuence of both above properties, it allows the 
grammar to be totally lexicalized (Schabes et al, 1988). 
However, a certain number of syntactic phenomena are 
difficult to represent in this framework. We focus here on 
French pronominal citrics which are cases of non 
canonical argufnen\[ realization. We show how they can 
be naturally handled using the "Synchronous TAG" 
extension of the formalism as a turther set oi 
wellformedness conditions on tbe language and we extend 
this treatment to uther cases of inisnlatch betv,,een 
syntactic attachment and senlantic role. 
1. Motivation 
French pronominal clitics have 'been subject to numerous 
studies in both thexwetical and computational linguistics. 
We restrict here our discussion 1o syntactic properties of 
clitics and thus ignore most of their morphological, 
2 phonological or binding properties , We show that they 
cannot be handled by existing TAG variants such as 
FOTAGs or MCTAGs (unless major cMnges are made 
to the existing French TAG grammar) but Ihut 
Synchronous TAGs provide an elegant and uncxpensive 
solution. 
1.1. Some linguistic properties of French 
citrics French pronominal citrics fall into 6 groups: tbe 
nominative ones @'e, or~, iL..), the 1st, 2d person and 
reflexivc ones (se. hie..), the accusative ones (le, la..), 
the dative ones (lui. leur..) plus en and y which each 
forms a class on its own. They follow each other in a 
rigid order: nora < refl < ace < dat < y <en, and are 
subject to numerous corcstrictions3: no identical 
l. It is also the basis of an interactive NLF' enviromnent 
(Paroubek et al. 1992). See Joshi 1987a and Schabes eta\]. 
1988 f~)r an introduction to TAGs. 
2. French pronominal clitics are subject to obligatory 
sandhi phenolnena such ms liaisolI and elision. 
3. "I'tmse properties can easily be described with a finite 
automaton; Cf Gross 1989. 
ACRES DE COLING-92, NANTES, 23-28 AoOr 1992 60 
ambiguous forms can cooccur even if they play different 
(and compatible) syntactic functions: 
(Ia) Paul songe c" Marie duns le jardin. (Paul is 
dreaming about M~u'ie in the garden) 
(lb) * Paul y y songe (Paul dreams of her there) 
(2a) Paul remptit un verre de vin (Paul fills one 
ghLss with wine) 
(2b) * l'aal en en remplit un (Paul fills one with it) 
No tst or 2nd person direct object may coc(:cur with a 
dative clitic: * Paul me leurprdsente (Paul introduces me 
to them). 
Citrics do not have tile same syntactic properties as NPs: 
they do not co~rdiuate with NPs, nor take NP modifiers, 
and are usually assigned a specific category (Cli0. It is 
commonly agreed upon that French pronominal ctitics 
(plus the negative marker ne) form a syntactic verbal 
cluster with the main verb (or with the reflected auxiliary 
4 if there is one) of the sentence 
In standard French, pronominal clitics are in 
complemenlary distribution with syntactic phrases as 
complements of various lexical heads (V, N, A). They 
may appear t)n a verbal head (rf which they are not an 
argument: 
(3) Jean en est aired (cplt of the past participle) 
(4~ Jean lui est fiddte (cplt of the adjective) 
They :ire also subject 1o nmnerous lexieal constraints. 
Object cliticization may be ruled out by certain verbs 
which impose a non clitic pronominal form: 
(5a) Jean (ressemble + pense) d Paul (Jean 
resembles/thinks about Paul) 
(5b) Jean lui (ressemble + * pense), 
(5£) .lean (?* ressembte + pense) d lui 
(6a) Jean (pense + aline) que Marie h~ti raconte des 
histoires. (Jean thinks/likes it that Marie tells him 
stories) 
(6b) .lean lie pense + * l'aime). (ruled out with 
intended mexming) 
Clitics en and y may also behave as adverbial modifiers. 
Several proposals have been made in computational 
linguistics lk)r handliug these phenomena. Bhs 1988 gave 
a GPSG treatnmnt of French citric objects, which 
essentially cnnsiders them as preverbal NPs (with a 
sDecial <Clit> feature), except for invert~ subject clilics 
which he considers verbal suffixes, lie does not consider 
adverbial uses nor cases of non kx:al dependencies or of 
auxiliary-verb combinations. Lexical entries for citrics in 
4 See for example Grnss 1968. Kayne 1975 for various 
argllnlellts t*or a Clilics Verb constittlent in Yrench. See 
Abeilld 1992 for arguments against a VP constituent. 
P~aoc. oF COL1NG-92, NANTES, AUG. 23-28, 1992 
dislocated constructions arc duplicated. Baschung et al. 
1987 present a UCG treatment with a more restricted 
coverage which considers objeet citrics as functors taking 
(on their right) a verb expecting an NP (or PP) argument 
on its right and returning a more saturated verb. They do 
not give a uuified treamlent of subject clitics which they 
consider as standard NPs not" do they handle non local 
dependencies. BSs et al 1989 give a slightly modified 
version of this treatment allowing for what they cu!l 
French "semi-free" word order. Miller 1991, using a 
HPSG-style framework, considers clitics as "phrasal 
affixes" on VP and uses optioual lcxical rules to update 
the subcategorization frame of the corresponding verbal 
entries and foot features to keep track of the presence of a 
clitic in the tree. He accnunts lot a lot of non local 
dependencies (including causative constructions) but 
needs extra constraints to handle locality constraints. He 
does not talk about inverted clitics nor dislocated 
constructions. This treatment is not, to otlr knowledge, 
implemented in a computational application. 
1.2. Difficulties with existing TAG vari,'mts 
We can first put aside "fi'ozen" clitics which are easily 
handled in Lexicalized TAGs: they do not play any 
semantic role and "co-head" the elementary trees of their 
predicate ("impersonal" il, se in "middle" constructions 
anti various idioms). Clitics with a semantic role 
(adverbial modifiers or arguments of a verb, an adjective 
or a noun) are more difficult to handle. 
One could think of adding to the existing tree families 
(associated with the predicams) an elementary tree (with a 
substitution node for the citric and an empty argument 
node) for each possible citric (or clitic combination) 
realiz,ation. This would be somewhat uuprincipled and 
5 lead to numerous extra trees , unless one generates the 
new trees by metarules or lexical rules (Vijay-Shanker 
and Schabes 1992, Becket 1990). It would also separate 
the syntax of adverbial ctitics from that of argumeutal 
ones attd disallow many non local dependencies. 
One might also considcr cxtensious of TAGs, such as 
FOTAG or MCTAG variants. In Multicomponent 
TAGs, used for extunple by Kroch and Joshi 1986, the 
elementary units of the grauunar are tree sets (notexl hcrc 
with braces) instead on single trees. Adjunction (or 
substitutimt) thus becomes simullaneous adjuuction (or 
substitution) of the different members of a tree set into a 
tree (or into the members of another tree set). The 
different members of such a set need not be of the same 
type, ~md we could use a set compristhg an auxiliary tree 
beaded by the clitic and an initial tree headed by tile 
empty striug for the ct)rrespondiug argument position: 
S ./- 
/ V <dotif>= x 
NO I' v b;°cc:+> NII' ! "~/"~V NI 
volt \ k la <refl>=- / e x, // 
Figure 1..lean la volt with MCTAGs 
Tile substitutien node corresponding to the NP 
realization of the ctunplement is thus filled with the 
empty string when the citric adjoins, ltowevel', this will 
not work for PP complements, since in this case, the 
whole PP subtree with the prel×mition "co-he4ul" woukl 
have to be "zeroed" when the citric adjoins, an opcratiou 
not provided by the forumlism 6. 
S (l~ce>=- 
NO ,I. PP j /~._ " V t<dat>=+./," 1 ! " "~" 
cc>=- 
ressemble a lui 
Figurc 2. llkx:king Jean lid ressernble with MCI'AGs. 
"Free Order" TAGs is a variant analogous to the ID/LP 
format for GPSG which was first introduced by Joshi 
1987b and developed by Becket et al. 1991. Argument 
clitics would thus N: treated as moved NPs (substituted 
at the same NP nodes) according to LI' statements such 
as the following: Nl(+Clit) < V. This representation 
might he attractive for handling cases of "clitic 
climbing" er non local dependencies but faces 
unescapable problems: 
I It is similarly unable to account for dm argument 
PP/clitic alternation, since the nixie to be substituted is 
an NP, not a PP in the non citric case. 
- It prevents from having an unified syntactic 
representation of the different tittles (it dcms not make 
any syntaclic distinction between NP and clitic subjects) 
- It does not regard tittles-verb as a constituent ill die 
syntax, and it is difficult to see how corestrictions 
between citrics could be handled (the ~uue warning heMs 
tk}r prohibiting adverb insertion between citrics and verb). 
Current TAGs versions thus do not provide a satisfactory 
account of die lion trltzen prmmminal clitics. We now 
turn to au alternate representation which will nut be 
strictly syntactic but involve the syntactico-scmantic 
interl)ce defined in the Synchronous TAGs fimnework. 
2. A Synchrnuous 3'At;, representation 
2.1. Synchronous TAGs Synchronous TAGs have 
been introduced by Shieher and Schabes 1991) to 
characterize correspondences between Tree-Adjoining 
lauguages. They can he used for relating two TAGs for 
two different hutguages for the purpose of machine 
translation (Abeilld et al. 1990), or lk~r relating a 
NI \ 
e 
5. Two different trees would lm needed for Jean le lui donne 
(/can gives it to him) where first clilic=N1, second clltic 
=N2, and for Jean me le donne (Jean gives it m me) where 
first clitic=N2, secemd clitic -N1. 
6. We recall that. due it) their lack of semantic autonomy, we 
consider argument marking t)repc~sitions as co-heads of the 
elementary tree of their predicate, contrary to prepnsitions 
heading adjLmcts wifich are autonomous heads of their 
auxiliary tree. 
ACIES Dr COLING-92. NArgrl~s, 23-28 Ao~r 1992 6 1 PRO(:. OF COL1NG-92, NAN'iI~S, AUG. 23~28, 1992 
syntactic TAG and a semantic one for the same language, 
.7 for the purpose of generation or semantic analysts . We 
consider here the latter case and assume that both 
syntactic and .semantic TAGs are lexicalized and feature- 
based 8. 
In Synchronous TAGs, TAG elementary trees are paired 
with semantic units expressed in a logical form language 
which is also a lexicalized TAG. The correspondences are 
stated as tree pairings with some of the nodes of the 
paired trees possibly linked. The following are examples 
9 of such pairings : 
5 F 
\1 I/ 11 
30an je'~,' \~/ 
Figure 3. Pairing with semantic flees for Jean and NO 
dorrmr 
The links between syntactic and semantic nodes mean 
that an operation at one node in the syntactic tree has an 
equivalent combination at the linked node in the 
semantic tree (and vice versa). More precisely, the 
semantic interpretation of a sentence is built 
"synchronously" with its syntactic derivation, by 
choosing a pair of elementary trees (a syntactic one:T1, a 
semantic one:L1) from the grammar and repeatedly 
updating it as follows (Shieber and Schabes 1990): 
- choose a link between two nodes nl and n2 in the pair 
(T1, L1), 
- choose another pai r of trees (T2 and L2), the root nodes 
of which match the respective categories of the two 
nodes chosen above, 
- form a new pair by combining T2 at node nl and L2 at 
node n2 and remove the link (if T2 and L2 have links, 
these are preserved in the resul0. 
The definition allows for the operations performed at 
each node to differ, one being an adjunction and the other 
a substitution for example. It also allows for a node to 
be linked to several other nodes: in this case, only the 
"consumed" link is removed (the others are preserved and 
7. See Shieber and Schabes 1991 for some arguments for the 
use of trees instead of flat structures in Logical Form, and for 
the use of adjunction as an alternative to substitution in LF. 
8. We refer the reader to Schabes ct al. 1988 for more details 
on LexTAGs. Suffice here to say that the TAG elementary 
trees are systematically associated with lexical "heads" 
anchoring them in the lexicon (and required not to be empty) 
and are combined together either by substitution or 
adjunction. Feature structures are also associated at the nodes 
of the elementary trees and constraints on combining trees 
can be expressed in terms of success or failure of unification 
(Vijay-Shanker and Joshi 1988). Not all featur/:s are shown 
here. 
9. The non terminal names in the semantic TAG are 
mnemonic for Formula, Term and Relation. Only the 
relevant links are shown. 
we adopt here the convention that they are "pushed" 
• . . 10 upwards m the case of an adjunctton) . 
Since multiple links are allowed, one syntactic node may 
be linked to .several semantic nodes, and conversely, one 
semantic node to several syntactic nodes. This allows for 
the ability of a given syntactic element to play different 
semantic roles and for differeut syntactic realizations of 
the same .semantic role. For example, as explained by 
Abeill6 and Schabes 1990, this formalism naturally 
accounts for file adverbial status of adjectives inserted 
into idiomatic expressions: Jean a pris une vraie veste = 
Jean a vratment pris une veste (Jean has really come a 
cropper). We want to show here that, provided it is taken 
as part of the well-formedness conditions of the 
grammar, it also accounts for the properties of French 
pronominal citrics. 
2.2. Augmenting TAG wellformedness 
conditions with synchronicity Sofar, well-formed 
sentences of a TAIL have been defined in the syntax only• 
In this respect, an input sentence is accepted by the TAG 
parser iff it obeys the following conditions: 
- it can be derived from an initial tree rooted in S, 
- all features unify at each node in the corresponding 
derived nee. 
It is however possible to allow for the definition of the 
well-formed sentences of the language to be given jointly 
by the syntax and by the semantics, so that among the 
ill-formed sentences certain will be accepted by the 
syntax but rejected by the semantic rules if they cannot 
assign any interpretation to them. Such semantic 
filtering is not without history (Cf Pullum 1984) but it 
seems especially fruitful with the Synchronous TAG 
formalism because: 
- the syntax and the semantics use extactly the same 
formalism, 
- the syntactic and semantic derivations are necessarily 
built up in parallel. 
The following well-formedness constraint is thus added 
to the parser: a sentence is acceptable iff it has at least 
one valid semantic interpretation built "synchronottsly" 
with it. By valid semantic interpretation, we mean that: 
- it can be derived from an initial semantic tree, 
- all features unify at each node in the corresponding 
derived semantic tree. 
Several linguists have also suggested such semantic 
filtering for cases usually thought of as more syntactic 
(e.g. Sag and Klein 1982). The purpose of this paper is 
to advocate this device in various cases which all involve 
the syntax-semantic interface among which French 
pronominal clitics, 
2.3. French pronominal elitics with 
Synchronous TAGs We rely on the existing 
elementary trees in the grammar to which we add 
substitution nodes for all possible clitics. Both clitics 
and corresponding NP, AP or PP nodes are optional in 
the syntax, their alternate realization is triggered by the 
10. We refer the reader to Shieber and Schabes 1990 for 
formal aspects of Synchronous TAGs (which are more 
powerful than TAGs). 
A~s DE COLING-92. NANTES, 23-28 AOt3X 1992 6 2 PREC. OF COLING-92, NANTES, AUG. 23-28, 1992 
associated semantic representation. We show how we get 
the following distribution: 
(7a) Jean int~resse Marie (Jean is interesting for 
Marie) 
(7b) Jean I'int~resse 
(7c) * Jean int&esse 
(7d) * Jean l'int&esse Marie 
S F /r---- A\ 
(~m <(NO)~ V2 (NI)~ ~VI i TO4, TI~ 
Io)+i, ..,..,""~VOXintrestl/ >= <Cl2~ I \" 
e V / 
ace>=÷ . I. ~ /_j/ 
Figure 4. Elementary tree pairing for NO intdresser N1 
Both NP argmuents of intiresser are linked with the 
corresponding Terms in the semantic tree. But the Clitic 
nodes are also linked to those Temts. They are thus 
prevented co~cur since only one substitution is allowed 
for each Term on the semantic side. Sentence (7b) is thus 
derived as follows: 
.of;7---" 
I/' I I ~F \]ean' pro" 
~ eregt' 
Figure 5. Synchronous derived trees for Jean l' int&esse 
The four sentences (Ta)-(7d) are allowed in the syntax but 
only ,sentences (7a) and (7b) are associated a synchronous 
interpretation. No interpretation is possible for sentence 
(7c) because its derived semantic tree is incomplete: dte 
T1 is obligatory in the semantics. No interpretation is 
possible for sentence (Td) because whenever the clitic or 
the NP tree is substituted, it synchronously fills the T1 
term and prevents the other to be substituted. 
A motivation tor treating cases (7c) and (Td) (i.e, cases of 
argument missing or cooccurrence between clitics and 
full phrases) as "semantically" deviant is that it seems to 
be always possible to construct a context in which they 
could be accepted 11. We thus consider all the argument 
nodes to be optional and compatible in the syntax, their 
realization will be incompatible in the semantics (and 
might be obligatory if the semantic representation 
specifies so), 
Handling elitic corresponding to PPs is now 
straightforward, as shown in the following pairing: 
S F Jr---- A\ 
(NO) ~I V2 (PP) 
..---.,o I/j 
dat>=+ I 6 N. / .-/ 
Figure 6. Elementary tree pairing for NO ressembler ~ Nl 
Notice 1hat although N! is the argument of the verb, it 
is the PP as a whole which as marked as optional (and 
will be prevented to cooccur with a dative clitic). 
The same result could be achieved if one considers the 
elitics to be adjoined (instead of substituted) on the 
syntactic side but this will necessitate a richer feature 
system to check clitic ordering and compatibility (see 
Abcill6 1991a for a previous accomtt along these lines). 
In order to keep the feature system as simple as possible, 
we provide in fact nodes for all possil)le clitic realization 
(argumeulal or adverbial ones) in the corresponding 
elementary trees. The complete tree for a transitive verb 
like voir is thus the following (with clitic numbering as 
in section 1.1) I 2: 
S 
Figure 7. Elementary trees for NO voir NI. 
We will titus get Jean yen voit, Il se volt etc... 
2.4. Ambiguity and haplohtgy Ambiguities are 
provided by the multiple links coming out of a clitic 
node: en for example can correspond to an accusative or a 
genitive complement, y can Ire a dative complement or a 
locative adjmtct... If one takes a verb with an optional 
complement (such as songer) and the ambiguous clitic y, 
sentence (Sa) below is parsed as ambiguous whereas only 
the adverbial reading is selected fory in sentence (8b)13: 
(Sa) Jean y songe. (1:Jean is dreaming there/2:Jean 
is dreaming about this) 
(8b) Jean y songe d ses soucis. (Jean is dreaming 
there about his problems) 
11. In fact (7d) is OK in spoken French, which can be shown 
to exhibit "clitic doubling", and (7c) may be improved as in: 
? Si Jean n'est pas intdressara c'est parce qu'il ne salt pas 
int~resser. 
12. See Miller 1991 for arguments for having the subject 
clitics separated from the complements ones, 
13. Contrary to Miller 1991, we do not consider that (84) 
has a third "haplology" interpretation (Jean is dreaming 
there about it). nor do sentences such as Jean en remplit. 
(Jean is filling some wiflx it) 
ACIES DE COLING-92, NANTES. 23-28 Aotlr 1992 6 3 PRO(:. Or: COLING-92. NANq'r~S, AUG. 23-28, 1992 
(I) S F F ..A /r" I 
~I R TO R NO 
an r vo dream' /jean' loc-pro' 
(2) s 
/' .2S.!n ( oor 0y, 
Figure 8. Two possible derivations for.lean y songe 
Notice that sentences (lb) and (2b) above are not 
generated since there is only one position for y and one 
for en in the syntax. 
2,5, Further constraints Clitic insertion is ruled 
out in extraction contexts: 
(9) Je sais qui Jean regarde (I know who Jean is 
looking at) 
(9a) * Je sais qui Jean le regarde 
(9b) C'est Marie que Jean regarde (It is Marie that 
Jean is looking at) 
(9c) * C'est Marie que Jean la regarde 
In these constructions, which correspond to distiuct 
elementary syntactic trees (in the tree family of their 
head) the obligatory syntactic realization of the extracted 
element naturally prevents the substitution of the 
corresponding clitic. 
This representation is also directly suitable for marking 
various constraints, e.g. structural ones (ruling out en 
direct object for PP complements) or lexieal ones (verbs 
which forbid cliticization of their complement such as 
pen.wr & Nhum or aimer que P). As for feature equations, 
certain links are structurally defined as part of the 
elementary trees reg,'u'dless of their lexical head (and there 
will be no link between en and the T complement node 
in the tree family for verbs with a PP complement), 
other links are brought by the lexical head and only 
certain verbs with a PP complement (ressembler but not 
penser) will define a link between the dative Clitic and 
their T complement node. 
We now show how the representation sketched above 
naturally extends to some cases of so-called "non local" 
dependencies and to cases of cooccurrence between the 
clitic and the corresponding argument. 
2.6. Locality constraints and non local 
dependencies. As noted by Shieber and Schabes 1990, 
locality constraints are inherent to the formalism of 
Synchromms TAGs. Contrary 1o Miller 1991, who runs 
the risk of allowing too many non local dependencies 
with the FFP, we titus do not need to add specific 
locality constraints. 
Notice first that some "non local" dependencies in a PSG 
are treated locally in a TAG framework. Examples of 
these are sentences with raising verbs (adjoining to VI) 
or attxiliaries (adjoining to V0) following the word order 
: Jean peut le voir, Jean l'a vu. Adjoining a raising or an 
auxiliary verb only updates tile links coming out of the 
Vl or V0 node and does not interfer with the links of the 
clitics. We straightforwardly get: Jean aurait dtd aimd de 
Marie= Jean en aurait dtF aimF (Jean would have been 
loved by Marie). The agreement equations are the 
following (with o-agr for object agreement): 
S 
(N0) 11, t:<agr>=x (NI)~, 
<~gr>=x V2b:<agr>=y 
(C 10 ~""~ vit:<agr>=y 
<egr>=x (El ~'''~'z)~,- b: <ngr>=w 
<o-agr>=z V0 t:<agr>=w 
<o-ogr>zz 
Figure 9. Agreement pattern for transitive verbs 
Cliticization of the complement of an adjective is 
directly allowed in copular consmtctions: Jean estfidele g~ 
Marie = Jean lui est fiddle (Jean is faithful to Marie). 
Copttlar constructions are (independently) treated as 
extended elementary trees in the TAG grammar, with the 
adjective co-heading the elementary tree with the verb and 
the chic and PP(de) complement nodes belonging to the 
i4 same flee . 
However, such cliticization is ruled out lin modifying 
adjectives. Sentences snch as: J'ai rencontrF une fille 
fidOle d Jean = * Je lui ai rencontrd une fille fidOle (I met 
a girl faithful to Jean) are not generated since the T node 
corresponding to the complement of the adjective does 
not belong to the same semantic tree as that of NO 
rencontrer N1 with which the clitic tree must be 
cmnbined. 
The same "local" treatment holds for cliticization of 
complements of nouns. It is allowed in light-verb 
constrnctions such as: Jean a fair le portrait de Marie = Il 
en afait le portrait (Jean made a picture of Marie), which 
are represented as exl~nded elementary trees with the light 
verb and the predicate nominal co-heading the slructure. 
It is rulod out by our treatment when the NP is in a 
15 productive argument position 
14. There are differences in acceptability for cliticization 
with verbs taking adjectival arguments : Je lui (sais + 
?*trouve) Jean fiddle. (I know\]find Jean faithful to her). The 
difference is the same for extraction out of tile AP : A qui 
(sais+ ?*trouves)-tu Jean fidt?le ? and is acounted for by 
different syntactic elementary trees (an extended one for 
savoir, one with die AP to be substituted for trouver). 
15. This is obviously too strong a constraint since there are 
cases where tile clitic corresponds to a complement of a 
noun at an arbitrary levet of entbedding, such as: Le bord de 
la couverture de ce livre est ddchir~e = Le bord de la 
eouverture en est ddchird (The ctmler of lhe cover of the 
book is tom out). 
AorEs Dr! COLING-92, NANTIiS, 23-28 AoI%r 1992 6 4 PROC. Or: COLING-92, NANTES, AU¢I. 23-28, 1992 
en, which allows a direct object with a null he.~l-noun: 
Jean achdte deux potatoes = .lean en aehdte deux (Jean 
boys two apples). In such cases, the determiner heads a 
syntactic N initial tree but its solnantic tree is an 
attxiliary T tree which adjoins to the T node filled by the 
clitic. We also account lot cases where the dative 
(humml) clitic is semantically equivaleot to a l}ossessive, 
a construction typical of nouns of inalienable possession 
(such as t×xly parts) combined with certain action verbs: 
Ce docteur selene les dents de Jean = Ce docteur lui 
soigne les dents (This doctor treats JeAufs teeth), llere, 
the clitic lui will paired with an auxiliary T tree (as that 
for possessive determiners) and its node will be linked 
wilh tile T node of the direct comlllement (for tile verbs 
allowing this): it may thus cooccur with Ihe NP 
COlOpleiIlcnt. 
For lack of space, we do uot develop here Ilclitic 
climbing" in causative constructions which require either 
multicomponent trees on the semantic side or lexical 
rules adding causative double-headed elementary trees to 
the existing tree families. 
2.7. Cuoccurcence between clitics and I'11tl 
phrases Such c{×}Ccorrences are exhibited by inverted 
contexts such its: Qui Jean voit-il ? Such inverted clitics 
are represented ia the syntax as auxiliary trees which 
trigger an mversiou feature (Ahei116 1991 a) and adjoin to 
the right of tile inflected verbal form. On the semantic 
side, they are reprcsented as ambiguous: they are 
associated witii Terms and may thus alternate with uon 
inverted clitics or NP suhjccts (Qui voit-il ?), provided 
the verbal nodes are linked with the subject T uede. But 
they are also associated with auxiliary trecs adjoining a 
question marking at the top F node (and thus allowed to 
cooccur with NP subjects). 
Otlrer cooccurrences are exhibited hy dislocated 
constructious such as : .lean, Marie l'a vu or Marie l'a 
vu, Jean (Jean, Marie saw) which tend to gencralize in 
spoken language. Right dislocation for complement 
clitics can be accounted lot straightforwardly with the 
cxisting elementary trees if one allows for an alternative 
semantic representation of the clitics, nan~ely an 
auxiliary emphatic scmantic trec (adjoining to the Term 
• . . 16 node already filled by tile NP) instead el a full rerm : 
5 F 
~o, ~w ,,. R TO < 
 o7, 2 o .t .X I/, 
\[ ~ I R'Y ;e~".z~l l0 VOlt IDt}FIO 
Figure 10, Derivation of a right dislocated construction 
Multiple dislocations are thus allowed : Je le lui ai 
confiE l'argent, ti Jean (I gave Jean the nloncy) :is well its 
"mediaiP' ones :,lean l'a eonfi£ l'argent, d Marie. We do 
not consider here left dislocations which have different 
16, This alternative represenuttion of clitic l}ronouns as 
semantically vacuous is similarly used by Grmlshaw 1982 
for Spanish clitic doubling (optional feature Pred in the 
clitic entry). 
syntactic properties and for which a purely semantic 
principle see, ms neces.~ary in order to bind the pre~l 
NP (Cf Hirschbuhler 1975, Fr'~lin 1988). 
3. F.xt ensions 
We show how tile synchronous 'FAG framework 
naturally handles other cases of discrepancies between 
syntactic attachment and semantic role, for various non 
canonical configurations, while keeping the semantic 
component very simple and strictly comlx}siti{mal. 
3.1 "Argument-adjuncts" It has often been noted 
that syntactic adjuncts may behave semantically as 
17 arguments . q't~ese adjuncts may be in complemenulry 
distribution with arguments, such as possessive 
determiners and PP(de) coulplement.s: la venue de Jean/ 
sa venue (Jean's coming, his coming) vs * la venue/*sa 
venue de .lean• Syuebronous TAGs uatorally allow to 
represent such possessives as auxiliary trees m the 
syntax and as initial trees in tile semantics. For example, 
tile above dislrihution with the noun venue is accounted 
for by linking the attactonent l×}tht of the determiner (the 
top N) and the substitution node {11' the PP complement 
tO one and the S~lllle snbstitlltiou node in tile 
18 corfespondillg senlantic tree : 
/I (~P~) ~" TO~ 
Figure 11. Elementary trees for venue 
The phrase * la venue is di~llowed because substitution 
is obligatory at "II) and the definite article la is not paired 
with a Term initial tree. Other such alternations involve 
"relational" adjectives such as voyage pr&~Mentiel/voyage 
du prdsident and are handled sinfilarly (with the relational 
adjective paired with a senumtic term). 
3.2. "F, xtended" adjuncts It has also been noticed 
that a(ljuncts inay \[lave a Selllantic SCOl~ well t}eyood 
their syntactic attachorent point. For examl}le, raising 
verbs, negative adverbials, quautifiers all llave a semantic 
sentential scope ahh{}ugh they attach to verbal {}r 
nominal items. These discrepancies are easily handled 
with Synchronous TAGs provided correst}nnding links 
are added to the elementary tree pairs (e.g. between VI 
and the F uode for raising verbs; Cf Abeilt6 199 l b). 
Convcrsely, we can ball{lie cases of I'narftlW" scol)~3 snch 
as extraposed relative clauses which attach to S allhongh 
they semantically modify an NP: tile syntactic S tree of 
tile relative chnlse is paired with an aaxiliary T rooted 
tree in tile scmaulics which adjoins t{} the T ilode 
17. We leca\[I lhat in TAGs, there is a structural distinction 
between modifiers (which are adj{}ined) 1ti1(.1 complements 
(which are substituted). 
18. For some linguistic argumePds \[or representing 
determiners as sylttactic adjullcts, see for example Abeill6 
1991a 
ACII!S \]}I!COLIN(I-92, NAN"rE8, 23 28 A{}UI' 1992 6 5 P~(}c. {}F COLING-92, NANrI!s, AI:{I. 23 2g. 1992 
corresponding to the modified NP (provided S nodes are 
linked with the argument T nodes) 19. 
Conclusion 
We have shown how some non canonical arguments can 
be naturally handled using the formalism of Synchronous 
TAGs, provided the syntax-semantic synchronicity is 
incorporated as a well-formednesss condition in the 
grammar. We have applied this treatment to French 
pronominal clitics and handle the; r complementary 
distribution with complements, without increasing the 
number of elementary trees in the grammar. Thanks to 
the extended domain of locality of TAGs, their locality 
constraints are handled (similarly as subjaeency) without 
specific stipulations. We also handle cases of non local 
dependencies, provided one adds alternative semantic 
representation for ambiguous clitic complements. The 
same idea can be extended to other cases of mismatches 
between syntactic attachment and semantic role, such as 
"extended" adverb scoping or exWaposition. 
19. This case has been handled by "local" MCTAGs (Kroch 
and Joshi 1986) with "empty" trees for coindexing tile NP 
and the extraposed relative. Due to the inherent locality of 
Synchronous TAGs, the same effect of clause boundedness as 
"local" MC-TAGs is achieved. 
ACTES DE COLING-92, NANTES, 23-28 AOt'rr 1992 6 6 PROC. OF COLING-92, NANTFs, AUG. 23-28, 1992 

References 

A. Abeill6 1988. A lexicalized Tree Adjoining grammar 
for French, Tech. report, Univ. of Pennsylvania, 
Philadelphia. 

A. Abeill6, K. Bishop, S. Cote, Y. Schabes, 1990. "A 
Lexicalized Tree Adjoining Grammar for English", Tech. 
Report, Univ. of Pennsylvania, Philadelphia. 

A. Abeill6, Y. Schabes, 1990. "Non compositional 
discontinuous constituents in TAG", Symposium on 
discontinuous constituency, Tilburg (forthcoming van 
Horck, Wiesjma (eds), Mouton de Gruyter). 

A. Abeill6, Y. Schabes, A. Joshi, 1990. "Using 
Lexicalized TAGs for machine translation", 13th Coling, 
Helsinki. 

A. Abeill6, 1991a. Une grammaire lexicalis~e d'arbres 
adjoints pour le fran#ais: application d l'analyse 
automatique, Th~se de Doctorat, Universit6 Paris 7. 
(forthcoming Editious du CNRS, Paris) 

A. Abeill6 1991b. "Raising verbs in Tree Adjoining 
grammar", 26th Linguistik Kolloquium, Poznan 
(Vetalani, Darski (eds), Springer Verlag). 

A. Abeill6, 1992. "Sur la notion de syntagme verbal en 
fran~ais", Lingvisticae lnvestigationes, XVI:2. 

K. Baschung, G. Bi~s, A. Corluy, T. Guillotin 1987. 
"Auxiliaries and clitics in French UCG", 3rd European 
ACL, Copenhagen. 

T. Becker, 1990. "Metarules lor TAGs", 1st TAG 
Workshop, Sarrebriick. 

T. Becker, A. Joshi, O. Rambow 1991. "Long-distance 
scrambling and Tree Adjoining Grammars", 5th 
European ACL, Berlin. 

G. B~s 1988. "Clitiques et constructions topicalis6es 
dans une grammaire GPSG du franqais", Lexique, n°6, 
Lille. 

G. B~s, C. Gardent, 1989. "French Order without 
order",4th European ACL, Manchester. 

J. Grimshaw, 1982. "On the lexical representation of 
romance reflexive clitics", in Bresnan, The mental 
representation of grammatical relations, MIT Press. 

B. Fradin, 1988. "Approches des constructions it 
d6tachement : la reprise interne", Languefraneaise, 78. 

M. Gross, 1968. Syntaxe du verbe, Cantil~ne, Paris. 

M. Gross 1989. "The use of finite automata in the 
lexical representation of natural language" in Gross & 
Perrin (eds), Electronic dictionaries and automata in 
Computational Linguistics, Springer Verlag. 

P. Hirschbuhler, 1975. "A non transformational source 
for left hand NPs", Linguistic Inquiry, 4. 

A. Joshi, 1987a. "An introduction to TAGs" in Manaster 
Ramer, The Mathematics of language, Benjamins. 

A. Joshi, 1987b. "The relevance of TAG to generation", 
in Kempen (ed), Natural Language Generation, 
Dordrecht. 

R. Kayne, 1975. French syntax, MIT Press. 

A. Kroch, A. Joshi, 1986. "Analysing Extraposition in a 
TAG", in Huck and Ojeda (eds): Discontinuous 
Constituents, Syntax and Semantics vol. 20, Academic 
Press. 

P. Miller, 1991. Clitics and constituents in Phrase 
structure grammar, PhD Thesis, Univ. Utrecht. 
(forthcoming Garland, New York) 

P. Paroubek, Y. Schabes, A. Joshi, 1992. "XTAG: a 
graphical Workbench for developing TAGs", 3rd Conf. 
on Applied NL Processing, Trento, Italy. 

G. Pullum, 1984. "Syntactic and Semantic parsability", 
lOth COLING, Stanford. 

I. Sag, E. Klein, 1982. "The syntax and semantics of 
English expletive pronoun constructions", in Barlow et 
al. (eds) Developments in GPSG, vol. 2, Indiana Ling. 
Club, pp. 92-136. 

Y. Schabes, A. Abeill6, A. Joshi, 1988. "Parsing 
strategies with lexicalized grammars: Tree adjoining 
grammars", 12th COLING, Budapest. 

Y. Schabes, 1990. Computational and Mathematical 
Properties of Lexicalized Grammars, PhD Dissertation, 
Univ. of Pennsylvania, Philadelphia. 

S. Shieber, Y. Schabes, 1991. "Generation and 
Synchronous TAGs", Computational Intelligence 7:4. 

S. Shieber, Y. Schabes, 1990. "Synchronous Tree 
Adjoining Grammars", 13th COL1NG, Helsinki. 

K. Vijay-Shanker, A. Joshi, 1988. "A Feature-based Tree 
Adjoining Grammar", 12th COUNG, BudapesL 

K. Vijay-Shanker, Y. Schabes, 1992. "Structure sharing 
in TAGs", 14th COLING, Nantes. 
