Proceedings of the 8th International Workshop on Tree Adjoining Grammar and Related Formalisms, pages 137–140,
Sydney, July 2006. c©2006 Association for Computational Linguistics
Handling Unlike Coordinated Phrases in TAG by Mixing Syntactic
Category and Grammatical Function
Carlos A. Prolo
Faculdade de Inform·atica - PUCRS
Porto Alegre, RS, 90619-900, Brazil
prolo@inf.pucrs.br
Abstract
Coordination of phrases of different syn-
tactic categories has posed a problem for
generative systems based only on syntactic
categories. Although some prefer to treat
them as exceptional cases that should re-
quire some extra mechanism (as for ellip-
tical constructions), or to allow for unre-
stricted cross-category coordination, they
can be naturally derived in a grammatic
functional generative approach. In this
paper we explore the ideia on how mix-
ing syntactic categories and grammatical
functions in the label set of a Tree Adjoin-
ing Grammar allows us to develop gram-
mars that elegantly handle both the cases
of same- and cross-category coordination
in an uniform way.
1 Introduction
Generative grammars that we commonly hear
about in computational linguistics are usually
based on syntactic categories. This is also the case
when the formalism used is the Tree Adjoining
Grammars (TAGs). Large scale handcrafted gram-
mars for many languages have been built based
on this paradigm (Joshi, 2001; XTAG Research
Group, 2001; Kroch and Joshi, 1985; Abeill·e
and Candito, 2000; Candito, 1998; Becker, 1993;
Frank, 2002; Joshi and Schabes, 1997; Abeill·e
and Rambow, 2000) as well as grammars extracted
from corpora (Chen and Vijay-Shanker, 2000;
Chiang, 2000; Hwa, 1999; Xia et al., 2001; Xia,
2001). The latter is partly due to the fact that large
scale annotated corpora such as the Penn Treebank
(Marcus et al., 1994; Bies et al., 1995) give pri-
macy to syntactic categories. After all this is the
most strongly sedimented generative approach at
least since (Chomsky, 1957).
Computational approaches of grammar based
on grammatical function such as that of Susumu
Kuno (Kuno, 1987) have been given less impor-
tance. Although we can think of simply inserting
functional labels in elementary trees or use them
in a meta-level to generate the grammar, such as
in (Candito, 1998; Kinyon, 2000; Cl·ement and
Kinyon, 2003), such tags are generally not seen
as an essential part of the derivational process.
Nevertheless coordination is such an inherently
functional phenomenon as we show next. Con-
sider the sentences in (1) and (2). These are ex-
amples of regular coordination between phrases
of the same category. They can easily be handled
in the traditional grammar approaches of syntactic
category.
(1) She  ew [PP on May 1st and on July 4th ].
(2) They sell [ADJP electric and electronic ]
products.
Now look at the cases in (3) and (4). They
are different in the sense that the coordination is
across categories. This poses a strong problem
to the traditional grammar of syntactic categories.
This has been noticed for TAGs in (Prolo, 2002).
Recently this has also been tackled in the HPSG
framework by (Sag, 2003) and (Abeill·e, 2004).
The Penn Treebank calls this constituents UCP for
 Unlike Coordinated Phrases (Bies et al., 1995).
The problem is that we would need rules of the
kind below (using context-free rules for short  
see (Prolo, 2002) for TAGs). Basically all pairs of
constituents can be coordinated but we can not as-
sign to the resulting constituents either of the sub-
constituent tags.
137
UCP → ADVP CC PP
UCP → PP CC ADVP
UCP → ADJP CC NP
UCP → NP CC ADJP
(3) She  ew [?? yesterday and on July 4th ].
(4) They sell [?? electronic and computer ] de-
vices.
However, UCP coordination is not random.
Two constituents can be coordinated only when
they are ful lling the same grammatical function
(with respect to a third head). In (3) they are play-
ing the role of adverbial adjuncts of went. Either
one can engage in that relation individually and
hence they can be coordinated while playing that
role. Likewise in (4) the adjective electronic and
the noun computer are both  ne as left NP modi-
 ers. Therefore they can be conjoined as such. As
a  nal example, consider the sentences in (5). Be-
cause the direct object of the verb know can be re-
alized as either an NP or a sentential complement,
they can be coordinated in that role as shown in
(6).
(5) I know the answer.
I know that you don’t know it.
(6) I know [ the answer and that you don’t know
it ].
Clearly the recursive process of conjoining the
constituents is at the grammatic functional level.
We show next how we can solve this problem el-
egantly by mixing grammatical function and syn-
tactic category in the set of symbols for the tree
nodes of a TAG.
2 A Grammar of Grammatical
Functions and Syntactic Categories
The elementary trees in our grammar are the pro-
jection of a lexical item as usual in Lexicalized
TAGs. However, root nodes do not correspond to
syntactic categories, but to grammatical functions.
The node for the function then dominates syntactic
category nodes, according to the way the function
is realized syntactically. Figure 1 shows trees for
an intransitive main clause and an NP subject.1
Main
S
a8a8 a72a72Subj ↓ Pred
V P
V diamondmath
Subj
NP
Ndiamondmath
Figure 1: Elementary trees for Intransitive Main
Clause and NP Subject.
NP
a8a8a8 a72a72a72AdnAdjLeft
NP
Ndiamondmath
NP∗
NP
a8a8a8 a72a72a72AdnAdjLeft
ADJP
Adiamondmath
NP∗
Figure 2: Elementary trees for Left Adnominal
Adjuncts.
Figure 2 has trees for NP left modi ers (adnom-
inal adjunct) realized either as an NP or an ADJP.
Finally, in Figure 3 we can see the trees for
coordination of left adnominal adjuncts. Notice
that they adjoin at the function node (AdnAdjLeft)
therefore allowing for the coordination of anything
that can ful ll that role, be them equal categories
as in (2) or the UCP case in (4). In Figure 4
we show an additional example with a PP right
NP modi er. It should be straightforward to see
how to build trees for AdnAdjRight coordination of
constituents realized by a PP or a relative clause.
In Figure 5 we  nally get to subcategorization.
In any approach to grammar development we have
to make decisions between explicitly modeling
certain restrictions in the tree structure or through
features (of a feature based TAG). That can be
seen ubiquitously in the XTAG grammar (XTAG
Research Group, 2001). We can use the tree of the
 gure with verbs such like eat and know, having
trees to realize the direct object as either an NP
or a sentence. Features in the lexical items would
prevent the derivation of eat with a sentential com-
plement. Another approach would be to further
detail the tree into one with a built in NP object
1Figures generally show templates where a diamond indi-
cates where the lexical item would be substituted in, though
occasionally we insert the lexical item itself.
138
AdnAdjLeft
a8a8
a8a8
a8a8
a72a72
a72a72
a72a72
NP ↓ CCdiamondmath AdnAdjLeft∗
AdnAdjLeft
a8a8
a8a8
a8a8
a72a72
a72a72
a72a72
ADJP ↓ CCdiamondmath AdnAdjLeft∗
Figure 3: Elementary trees for Coordination of
Left Adnominal Adjuncts.
NP
a8a8
a8
a72a72
a72
NP∗ AdnAdjRight
PP
a8a8 a72a72Pdiamondmath NP ↓
Figure 4: Elementary trees for a PP as Right Ad-
nominal Adjunct.
and another with a sentential complement. How-
ever, realization constraints would still have to be
present to allow for the coordination of only the
constituents that are allowed for the speci c verb.
For the reader unfamiliar with grammar modeling
we notice this is not a drawback of the approach.
Constraints beyond those represented in the struc-
ture are constantly made as a way to avoid irra-
tional growth of a grammar.
In Figure 6 we show still another interesting
case: the predicative clauses.2 We include it for
2Again this is one approach to modeling predicatives,
Main
S
a8a8
a8
a72a72
a72
Subj ↓ Pred
V P
a8a8 a72a72V diamondmath DirObj ↓
Figure 5: Elementary tree for a Verb that has a
Direct Object
Main
S
a8a8
a8a8
a72a72
a72a72a72
Subj ↓ NomPred
V P
a8a8
a8a8
a72a72
a72a72
V aux[be] Predicative ↓
Figure 6: Elementary tree for Predicative Clauses
this is a rich context for unlike coordination. One
can easily see how to generate trees for coordi-
nating NPs, PPs and ADJPs, as predicative con-
stituents so as to allow for (7).
(7) John was [ a gentlemen, always happy, and
never in bad mood ].
3 Conclusions
We showed in this paper how to build a Tree Ad-
joining Grammar of grammatical functions and
syntactic categories, mixed together in a princi-
pled way of function and possible realizations. It
brings the bene ts of allowing handling language
phenomena which are generative at each of the
two sides.
In particular, we showed how it solves the prob-
lem of coordination of constituents of distinct syn-
tactic categories.
Elementary trees are not clumsy. On the con-
trary they bring additional information to the
structure with minimal addition of nodes. This in-
formation could otherwise be hidden in node fea-
tures, which are generally used to represent infor-
mation that would be costly to maintain explicit in
the tree structure.
Finally we can see that this structure can be eas-
ily incorporated in a supervised grammar infer-
ence algorithm such as that of (Xia, 2001), pro-
vided the annotated corpus has grammatical func-
tion information. In fact this is the case in the Penn
Treebank, and Xia’s algorithm allows it to be used
3. Inferring the different kinds of verbs, with re-
spect to the functions they subcategorize for and
with the auxiliary verb be anchoring the tree and the predica-
tive as a substitution node. The alternative used in the XTAG
grammar of having the predicative head as anchor would be
possible as well.
3The same is true of other algorithms such as (Chen and
Vijay-Shanker, 2000)’s.
139
their realizations is an important issue here, and is
also feasible (see (Kinyon and Prolo, 2002)).

References
Anne Abeill·e and Marie-Helene Candito. 2000. Ftag:
A lexicalized Tree Adjoining Grammar for French.
In Abeill·e and Rambow (Abeill·e and Rambow,
2000), pages 305 329.
Anne Abeill·e and Owen Rambow, editors. 2000.
Tree Adjoining Grammars: formalisms, linguistic
analysis and processing. CSLI, Stanford, CA, USA.
Anne Abeill·e. 2004. A lexicalist and construction-
based approach to coordinations. In Stefan
Mller, editor, Proceedings of the 10th International
Conference on HPSG (HPSG’03), Michigan State
University, Michigan, USA. Available at:
http://cslipublications.stanford.edu/HPSG/4.
Tilman Becker. 1993. HyTAG: A new Type of
Tree Adjoining Grammars for Hybrid Syntactic
Representation of Free Word Order Lang uages.
Ph.D. thesis, Universitcurrency1at des Saarlandes.
Ann Bies, Mark Ferguson, Karen Katz, and Robert
MacIntyre. 1995. Bracketing guidelines for the
Penn Treebank II style Penn Treebank Project.
Marie-Helene Candito. 1998. Building parallel
LTAG for french and italian. In Proceedings of
the 36th Annual Meeting of the Association for
Computational Linguistics and 16th International
Conference on Computational Linguistics, pages
211 217, Montreal, Canada.
John Chen and K. Vijay-Shanker. 2000. Automated
extraction of TAGs from the Penn Treebank. In
Proceedings of the 6th International Workshop on
Parsing Technologies, Trento, Italy.
David Chiang. 2000. Statistical parsing with an
automatically-extracted Tree Adjoining Grammar.
In Proceedings of the 38th Annual Meeting of the
Association for Computational Linguistics, Hong
Kong, China.
N. Chomsky. 1957. Syntactic Structures. Mouton,
The Hague.
L. Cl·ement and A. Kinyon. 2003. Generating paral-
lel multilingual LFG-TAG grammars using a Meta-
Grammar. In Proceedings of the 41st Annual
Meeting of the Association for Computational
Linguistics, Sapporo, Japan.
Robert Frank. 2002. Phrase Structure Composition
and Syntactic Dependencies. MIT Press, Cam-
bridge, MA, USA.
Rebecca Hwa. 1999. Supervised Grammar Induction
Using Training Data with Limited Constituent Infor-
mation. In Proceedings of 37th Annual Meeting of
the Association for Computational Linguistics (ACL
’99), pages 20 26, College Park, MD, USA.
Aravind K. Joshi and Yves Schabes. 1997. Tree-
Adjoining Grammars. In A. Salomaa and G. Rozen-
berg, editors, Handbook of Formal Languages, vol-
ume 3, pages 69 123. Springer-Verlag, Berlin.
Aravind K. Joshi. 2001. The XTAG project at Penn.
In Proceedings of the 7th International Workshop on
Parsing Technologies (IWPT-2001), Beijing, China.
Invited speaker.
Alexandra Kinyon and Carlos A. Prolo. 2002. Iden-
tifying verb arguments and their syntactic function
in the Penn Treebank. In Proceedings of the Third
International Conference on Language Resources
and Evaluation (LREC), pages 1982 87, Las Pal-
mas, Spain.
Alexandra Kinyon. 2000. Hypertags. In
Proceedings of the 18th International Conference
on Computational Linguistics (COLING’2000),
Saarbrcurrency1ucken, Germany.
Anthony S. Kroch and Aravind K. Joshi. 1985. The
linguistic relevance Tree Adjoining Grammar. Tech-
nical Report MS-CIS-85-16, University of Pennsyl-
vania.
Sususmu Kuno. 1987. Functional Grammar. Univer-
sity of Chicago Press, Chicago, Il, USA.
Mitchell Marcus, Grace Kim, Mary Ann
Marcinkiewicz, Robert MacIntyre, Ann Bies,
Mark Ferguson, Karen Katz, and Britta Schas-
berger. 1994. The Penn Treebank: Annotating
predicate argument structure. In Proceedings of the
1994 Human Language Technology Workshop.
Carlos A. Prolo. 2002. Coping with problems
in grammars automatically extracted from tree-
banks. In Proceedings of the Workshop on Grammar
Engineering and Evaluation, pages 36 42, Taipei,
Taiwan.
Ivan Sag. 2003. Coordination and underspeci -
cation. In Jongbok Kim and Stephen Wechsler,
editors, Proceedings of the 9th International
Conference on HPSG (HPSG’02), Kyung-
Hee University, Seoul, Korea. Available at:
http://cslipublications.stanford.edu/HPSG/3/hpsg02.htm.
Fei Xia, Chung-Hye Han, Martha Palmer, and Aravind
Joshi. 2001. Automatically Extracting and Compar-
ing Lexicalized Grammars for Different Languages.
In Proc. of the Seventeenth International Joint
Conference on Arti cial Intelligence (IJCAI-2001),
Seattle, Washington.
Fei Xia. 2001. Investigating the Relationship between
Grammars and Treebanks for Natural Languages.
Ph.D. thesis, Department of Computer and Informa-
tion Science, University of Pennsylvania.
The XTAG Research Group. 2001. A Lexicalized Tree
Adjoining Grammar for English. Technical Report
IRCS 01-03, University of Pennsylvania.
