Using Tree Adjoining Grammars in the 
Systemic Framework* 
Kathleen F. McCoy, K. Vijay-Shanker, Gijoo Yang 
University of Delaware 
Newark, DE 19716 
Abstract 
In this paper we investigate the incorporation of 
Tree Adjoining Grammars (TAG) into the systemic 
framework. We show that while systemic grammars 
have many desirable characteristics as a generation 
paradigm, they appear to have problems in generating 
certain kinds of sentences (e.g., those containing discon- 
tinuity or long-distance dependencies). We argue that 
these problems can be overcome with an appropriate 
choice of structural units of realization. 
We show that TAG provides appropriate units of 
structural realization because they localize all depen- 
dencies and allow the realization of two independent 
subpieces to be interspersed with each other. We go on 
to show how TAG can be incorporated without affect- 
ing the basic tenants of systemic grammar. Finally, we 
indicate how the incorporation of TAG yields several 
benefits to the systemic framework. 
Introduction 
As pointed out by many researchers (e.g., \[Davey 1978; 
Mann 1983; Matthiessen & Kasper 1985; Patten 1988; 
Bateman & Paris 1989\]), systemic linguistics offers 
many advantages to a sentence generation component 
of a text generation system. Perhaps the strongest as- 
set of systemics is its view of the generation process 
as a goal directed enterprise. Its emphasis is on func- 
tion rather than form \[Halliday 1985; Fawcett 1980; 
Hudson 1971\], where the functional distinctions that 
are required in the grammar manifest themselves in the 
eventual output form. 
While systemic linguists have remained agnostic with 
respect to certain processing decisions, a computer im- 
plementation of a systemic grammar requires that ex- 
plicit decisions be made concerning realization opera- 
tors and the structures available for manipulation at 
each point in the processing. The explicit decisions 
that were made in previous implementations of systemic 
*This work is supported in part by Grant #H133ES0015 
from the National Institute on Disability and Rehabilita- 
tion Research. Support was also provided by the Nemours 
Foundation. 
grammar (e.g., \[Mann 1983; Mann & Matthiessen 1985; 
Matthiessen & Kasper 1985\]) have proven to be prob- 
lematic in some respects. In particular, the current 
implementations have difficulty in generating certain 
sentences which exhibit discontinuities or long distance 
dependencies. To date, these can only be handled in a 
limited fashion, and the solutions provided are not very 
satisfying. 
We argue that Tree Adjoining Grammar (TAG) pro- 
vides a structural unit that is precisely appropriate for 
the implementation of a systemic grammar for the gen- 
eration task. Moreover, we believe our use of TAG for 
this purpose is completely consistent with the systemic 
paradigm and helps to overcome the above difficulties. 
In this paper we first introduce the notion of a sys- 
temic grammar and the processing paradigm it es- 
pouses. We indicate problems with current implemen- 
tations of this paradigm. Next, we introduce the notion 
of lexicalized Tree Adjoining Grammars, emphasizing 
their structural domains of locality, and justify that the 
basic structures of TAG are appropriate structures to 
be used in an implementation of a systemic grammar. 
Following this we indicate how a tree adjoining gram- 
mar can be used as the basis for an implementation 
of systemic grammar indicating the differences between 
the approach of current implementations of systemic 
grammar and that which would result from the incor- 
poration of TAG. Finally, we indicate potential gains 
resulting from the incorporation of TAGs and the scope 
of current work. 
Generating in Systemic Paradigm: 
Problems? 
Systemic linguistics deals with the meaning and func- 
tion of an utterance, it is a semantics driven approach 
rather than a syntax driven approach. In systemics, 
form follows function. The grammar itself is factored 
into three metafunctional domains (each of which affect 
the final text): ideational (concerning the characteris- 
tics of the conceptual situation to be presented), inter- 
personal (concerning the interaction between speaker 
and hearer) and textual (concerning the coherence of 
1 
the text as a whole). 
A systemic functional grammar consists of networks 
of grammatical choice alternatives, where individual 
networks are concerned with one of the metafunctional 
domains. In generation, these networks are traversed 
and fragments of grammatical structure are built up 
and manipulated using a set of realization operators 
which are associated with the various choice alterna- 
tives. The correct choice is made by consulting the 
information concerning the planned utterances. As the 
choices are made, the associated realization statements 
in the network are evaluated in order to realize the final 
structure. 
For instance, figure 1 contains a small fragment of a 
systemic grammar for clauses. The network indicates 
that a clause may either be simple or complex. If it is 
simple and full, then the grammatical function process 
is inserted (indicated by +process). The realization op- 
eration '% subject process" indicates that the subject 
should be ordered before the process in the final real- 
ization. 
Systemics deals with communicative function and its 
eventual surface manifestation at many levels. The 
basic processing starts with a semantically meaning- 
ful piece of representation which is decomposed into its 
component pieces via network traversal. The compo- 
nent pieces may then be further specified by re-entering 
the network. Given this rank-based decomposition (or 
stepwise decomposition), it is not unreasonable to as- 
sume that 1) decisions at a higher rank are made prior 
to decisions at a lower rank - that is, the decompo- 
sition of a particular semantic unit may not be influ- 
enced by the eventual decomposition of its component 
pieces, and 2) the structural realizations of the com- 
ponent pieces of a decomposed unit must be handled 
independently. These criteria, which we call the inde- 
pendence criterion, are implicitly followed by current 
computer implementations of systemic grammar. 
Implementing a systemic grammar on computer has 
forced researchers to be very explicit about certain rep- 
resentational issues. Such explicitness (coupled with an 
implicit following of the independence criterion), has 
enabled the uncovering of certain constructions which 
appear to be problematic for the systemic framework. 
For instance, Matthiessen points out that: "There are 
various structural relationships (e.g., discontinuity...) 
that do not pose a problem for the informal box diagram 
representation \[used by systemic linguists\] but prove to 
be a problem for explicit realization statements \[nec- 
essary for computer implementations\]."\[Matthiessen & 
Kasper 1985, p. 6\]. We believe that the same applies 
to (so called) long distance dependencies. 
It can be argued that the problems uncovered by 
Matthiessen are not due to explicit realization state- 
ments, but rather result from using explicit realization 
statements coupled with the implicit assumption that 
grammatical functions are realized as atomic strings. 
We further argue that if the independence criterion is 
to be followed, the choice of realization operators, the 
scope of the realization operators, and the choice of 
appropriate units of realization must be considered to- 
gether. 
Consider, for example, the problems which arise in 
generating a sentence containing a long distance de- 
pendency when atomic strings are taken as the struc- 
tural unit of realization. Consider the sentence: "It was 
Mary that John thought Bill liked". A natural decom- 
position would result in the semantic subpieces which 
correspond to "john thought" and the dependent clause 
"bill liked mary". The extraction of "mary" will be 
done subsequent to this decomposition; this extraction 
should influence the realization of the structural unit for 
the dependent clause (and should not affect the func- 
tional decomposition). But notice, if we assume the 
structural units of realization are atomic strings, we can 
get "John thought it was Mary that Bill liked" but not 
our desired utterance "It was Mary that John thought 
Bill liked" because to do so would necessitate inserting 
one atomic string ("John thought") within another ("It 
was Mary that Bill liked"). 
Two possible ways to get around this problem are: 
(1) to say that the intended sentence is made up of sev- 
eral independent but smaller functions (e.g., realized 
as "john thought", "bill", "liked", "mary"). But this 
solution goes against the step-wise decomposition into 
semantically meaningful units. Moreover, this method 
is not sufficient because we can always consider a sen- 
tence that requires one more level of embedding which 
would necessitate a revision of units. (2) not to break up 
the sentence into the two functional constituents men- 
tioned, but use some other decomposition. But this is 
not a correct solution because then one functional de- 
cision at a lower level (extraction of "mary") influences 
another which logically precedes the first (decomposi- 
tion in the clause complex network). 
In order to follow the independence criterion (i.e., 
realization of different independent functions do not in- 
fluence each other), the structural units being built as 
the realization of a grammatical function must be ca- 
pable of localizing all structural dependencies because 
they must embody all constraints specified with that 
function. In addition, the chosen structural units must 
be composable in such a way as to allow the surface 
string of one unit to be interspersed with the surface 
string for another (as was required in our example). If 
this is the case, then it makes sense that these struc- 
tural units should be taken as the bounding domain 
for the realization operators. The structures used by 
TAGs have precisely these qualities and are thus an ap- 
propriate choice for a structural unit of realization in 
an implementation of systemic grammar. The struc- 
tures used in a TAG have an enlarged domain of lo- 
cality that factor all dependencies (e.g., agreement con- 
straints, predicate-argument structure, filler-gap depen- 
2 

Transitive Tree Group 
S S 
NPO ~\[+wh\] S 
NPO~ VP 
NP0 VP 
V NP1 ~ \[ 
V NP1 ~¢ 
S 
NP1 ~ S 
NPO ~ VP 
V NP1 
I 
means that at that node substitution must occur. 
Figure 3: A Tree Group Selected by Like 
from one another by syntactic variations (which we shall 
call transformations). For example, the verb like, which 
takes a nominal subject and a nominal object, selects 
the transitive tree group. Some of the members of this 
tree group axe shown in Figure 3. The figure contains 
three initial trees, the first corresponds to a declarative 
sentence, the second to a wh-question on the subject, 
and the third to an it-cleft construction on the object. 
S-TAGs 
The processing within a Systemic Tree Adjoining 
Grammar (S-TAG) is similar to that in systemic gram- 
mar (e.g., the networks axe traversed and the realization 
operators associated with the choices taken are evalu- 
ated). In S-TAG we have already stated that the indi- 
vidual grammatical functions are realized structurally 
by elementary trees and that elementary trees provide 
the bounding scope for application of realization opera- 
tors. Thus, the "functions" which are inserted into the 
systemic structure will be associated with elementary 
trees in the TAG formalism. While the types of realiza- 
tion operators required by S-TAG will be the same as 
for general systemic grammars, the individual operators 
will be tailored to the TAG formalism. 
Regions in S-TAG 
The basic processing within a systemic grammar must 
take into account two dimensions of processing deci- 
sions: 
1. Metafunctional domains. Structures are built in the 
three metafunctional domains (ideational, interper- 
sonal, and textual) simultaneously. Certain realiza- 
tion operators are used to "conflate" the indepen- 
dently built structures. 
2. Processing from one "rank" to another. It is through 
changes in rank that semantic structures are eventu- 
ally realized as surface form. The general method- 
ology is to insert functional units into a structure. 
Following this, these functional units are refined by 
re-entering the network at a lower rank. This process 
continues until a surface structure has been fleshed 
out. 
While the processing in the S-TAG grammar follows 
the same principles, we differ in some implementation 
issues to accommodate TAG. One of the major contri- 
butions of this work is in the processing from one rank 
to another. In particular, this work makes explicit the 
bounding domains for the realization operators which 
are responsible for realizing a given grammatical func- 
tion. Thus it becomes clear what is available for manip- 
ulation when a network is entered (and re-entered for 
specifying a function inserted during the initial network 
traversals). We employ the notion of a region for this 
purpose. 
In general a region is created to expand a grammat- 
ical function. Since we have said that elementary trees 
are appropriate structural units for realizing the func- 
tions and for the bounding domains for the realization 
operators, we state that an elementary tree will eventu- 
ally be associated with every region. The appropriate 
elementary tree will be chosen after a decision has been 
made to insert a lexical item. Informally, this lexical 
item will be the lexical anchor of the elementary tree 
that will be chosen in the region. For this reason we 
will call this lexical item the lexical anchor of the re- 
gion also) The region serves as the bounding domain 
on the realization operations. All realization operations 
1This approach has interesting consequences, such as 
adopting a head driven generation strategy within the sys- 
4 
used within a region are applicable on the features of the 
lexical anchor of the region or the tree selected by this 
anchor. In the section on "Lexicon and Tree Groups" 
we will discuss how the features of the anchor, the tree 
groups selected, and trees selected will be maintained 
m a region. 
Once a lexical anchor is picked, the tree groups asso- 
ciated with that anchor will be considered. The choices 
in the network will cause realization operators to be 
evaluated which will narrow this set of trees to one. 
This single tree is then said to be associated with the 
region and will be the structural.realization of the gram- 
matical function being expanded (whose expansion was 
the reason for the creation of the region). This filtering 
will be done by using realization operations that select 
between tree groups and those that select tree members 
within tree groups. Such realization operations will be 
discussed in the section on "Realization Operators". 
Based on the characteristics of the tree associated 
with a region and directives from the networks be- 
ing traversed, decisions to expand certain grammati- 
cal functions (previously inserted in the region) will be 
made. Sub-regions will be created to expand these func- 
tions and the network will be re-entered to determine 
the expansion. Notice that this will cause an elementary 
tree to be associated with each sub-region. These sub- 
regions must eventually be combined with the super- 
regions which spawned them using realization opera- 
tions for adjoining and substitution. 
This view of the process is potentially complicated 
since some of the realization operations to be evaluated 
in the region may not be applied before the set of trees 
being considered is narrowed down sufficiently. For ex- 
ample, an operator which selects a particular tree need 
not be applied until the tree groups have been narrowed 
to one. If at the point an operator which selects a tree 
is called for, if the number of tree groups has not been 
reduced to one, it serves no purpose to apply the oper- 
ation on each tree group. Hence, in such cases, within 
a region we will maintain a record of realization opera- 
tions that have to be completed later. These operations 
will be applied at the appropriate time. 
Lexicon and Tree Groups 
In the lexicon, a lexical item will be associated with 
a set of tree groups, each of which contains a set of 
elementary trees. We choose to represent a tree group 
in the lexicon as a feature-structure/tree-group-name 
pair. The feature-structure includes all of the common 
features of the trees within the tree groups, the features 
of the lexical item itself, and its lexical idiosyncrasies, 
if any. The tree group name can be thought of as a 
pointer to the actual tree group kept in a separate area 
(allowing for sharing of tree groups by lexical items). 
temics framework. It will also have implications on the de- 
sign of the network. 
For example, with the lexical item, walk, we will 
associate the pairs (ft,.a,~s,trans), (fiotra,~,,intrans), 
(f, ou,~, noun), trans, for instance, is the name of the 
tree group for transitive verbs. All trees in this group 
share the information that the lexical anchor is a tran- 
sitive verb. Thus, this information is stored in ftrans 
along with any other features that are common to the 
trees in the group. The other two pairs represent the 
fact that walk can be used as a intransitive verb as well 
as a noun. 
The trees that constitute a tree group are kept to- 
gether. Some realization operators which will be evalu- 
ated in a region will refer to certain grammatical func- 
tions that are represented as nodes in the tree associ- 
ated with that region. Hence we will use a mapping 
table that maps abstract positions (grammatical func- 
tions) to actual nodes in an elementary tree. 
Realization Operators 
Having set up the notion of a region (and its asso- 
ciated elementary tree) as the bounding domain over 
which the realization operators can function, we are 
now in a position to discuss some of the realization 
operators that will be used in S-TAG. These opera- 
tors parallel those found in other systemic grammar 
implementations, although they are particular to the 
use of TAG. According to \[Matthiessen & Kasper 1985\] 
the realization operators used in a systemic network 
can be viewed along three dimensions: (1) Structuring 
(which defines the structure and its organization within 
one rank and within one functional domain), (2) Rank 
(which "organizes the grammar into a scale of units: 
clause- group/phrase - word - morpheme" \[Matthiessen 
& Kasper 1985, p. 25\]), and (3) Metafunctional lay- 
ering (which integrates the structures developed within 
the various metafunctional domains (e.g., interpersonal, 
ideational, and textual)). 
We concentrate on the rank and structuring opera- 
tors because they appear to be most affected by the 
addition of TAG. Aside from the nature of the actual 
structural units, a major difference between S-TAG and 
previous implementations of systemics is that previous 
implementations have built up structures of minimal 
import: upon proper evidence a functional unit is added 
to the current structure, ordered with respect to the 
other elements, accumulates features, and is then ex- 
panded so as to satisfy those features. There appears 
to be no automatic mechanism for carrying out syn- 
tactic implications of decisions that have been made. 
In S-TAG we take an opposite approach. In the TAG 
we have precompiled minimally complete packages of 
syntactic structure. Rather than building up structure 
only when we have enough evidence to know that it 
is correct (as has previously been done), our operation 
can be characterized as deciding between the syntac- 
tic possibilities that are consistent with what is known 
at the given point in the network. 2 As a result many 
of the structuring operators we introduce are designed 
to narrow down the trees that could possibly realize a 
particular function. 
Introducing the Lexical Anchor: Insert(Lex- 
item) When the lexical anchor is identified in the 
network, this operation will be used. The purpose of 
this operation is not only to introduce the lexical an- 
chor into the region but also to bring the associated set 
of feature-structure/tree-group-name pairs. Thus the 
tree group itself is not brought in but is indirectly ac- 
cessible. A tree is brought into the region only after 
the narrowing process is completed. The anchor is then 
inserted into the tree. 
Filtering Tree Groups in a Region We will use 
one realization operation to choose among the the tree 
groups being considered in a region. This choice is made 
on the basis of some features that become known during 
the traversal of the network and is basically a decision 
about the functional units the realization must repre- 
sent. Thus, in some sense it is analogous to the "in- 
sert" operator in Nigel. For example, the insertion of a 
particular lexical item, say walk, will bring into consid- 
eration all possible tree groups it can participate in. If 
it becomes known (in the transitivity network) that the 
recipient function will have to be realized, then among 
the various tree groups of the lexical anchor of the re- 
gion, only the appropriate tree groups (such as those 
corresponding to transitive verb form) will have to be 
considered. 
For current purposes, the realization operation that 
filters the tree groups will be called Select-Group 
which takes a feature as an argument. In the 
above example, the network may cause the opera- 
tion: Select-Group(transitive) to be evaluated. Re- 
call that the three tree-groups referenced for this lexi- 
cal item are represented by the pairs: (ftrans,trans), 
(fintran,,intrans), and (fnoun,noun). Since the 
feature-structures ftran,, fintrans, fnoun are kept in the 
lexicon itself rather than with the tree group, these tu- 
pies will be brought into the region on lexical insertion. 
If the realization operation Select-Group(transitive) 
(which is analogous to insert process and recipient in 
• the Nigel grammar) is evaluated in the region, the 
feature transitive is unified with the three feature- 
structures ftrans, fintrans, f.o~,n. Since this feature is 
only consistent with the features in ftrans, only the pair 
(ftrans, trans) will remain in the region. 
Selecting Trees from a Tree Group The realiza- 
tion operation used to narrow down the choice of ele- 
mentary trees within a tree group considered in a region 
2Note it is not necessary to bring all of the syntactic 
structures into the region, rather much of this processing 
can be done based on the features stored with the lexical 
anchor. 
is called Select-Tree. We had described a tree group 
to correspond to a specific semantic entity with all of 
its relevant semantic features inserted. The group it- 
self represents all syntactic realizations of this entity. 
Therefore the purpose of this operation is to choose 
among different syntactic forms possible. Its effect is 
somewhat analogous to that of the "order" operators 
in Nigel. For example, if during the traversal of the 
network it is realized that the object is to be topical- 
ized then the Select-Tree operation will be evaluated. 
Among the various syntactic variations possible, the 
tree(s) which realize this thematization will thus be 
identified. 
Composing Trees Recall that sub-regions are cre- 
ated to expand grammatical functions. The elementary 
trees associated with the sub-regions are to be com- 
posed with the tree associated with the super-region 
either by substitution or by adjunction. Expansion of 
a grammatical function is done, in the Nigel grammar, 
when a function is preselected with a set of features. 
The preselected features determine where to re-enter 
the network in order to expand the given function. The 
resulting realization will replace the original function 
in the eventual realization of the input. In S-TAG this 
is accomplished by using the realization operation Ex- 
pand(function, features). This will cause the creation 
of a sub-region (which is named by the function). The 
realization of the function will occur in this sub-region 
by re-entering the network at a point determined by the 
preselected feature (as in Nigel). 
The tree which eventually realizes the function must 
be composed (by substitution or adjoining) with the 
tree in the super-region at the node corresponding to 
the function (as given by the mapping table). The 
decision to adjoin or substitute is made based on the 
types of the trees that are picked in the sub- and super- 
regions. 
Discussion 
The strongest asset of systemic grammar is its view of 
generation as a goal-directed enterprise with emphasis 
laid on function rather than form. While our work in- 
volves the incorporation of a syntactic formalism into 
systemic grammar, we have not departed from the gen- 
eral approach of systemics view of generation. Sys- 
temic linguists, however, have not been interested in 
the details of the mapping between functional choices 
and the resulting form. In particular, they are not con- 
cerned with the details of the structural units that are 
realized. In a computer implementation, a programmer 
needs to be concerned about the details of the structural 
units, how they are realized, and how the constraints of 
systemic grammars are translated as principles of im- 
plementation. It is in this context that we propose the 
use of TAG trees as appropriate structural units and 
examine the processing paradigm (and its logical con- 
6 
sequences) that follows from such a choice. Thus, the 
incorporation of TAG is more than just a simple addi- 
tion of a syntactic formalism to the systemic framework. 
We argue that the incorporation of TAGs enriches a sys- 
temic grammar implementation for the following rea- 
sons: 
First, systemic linguists have stressed the notion of 
stepwise semantic decomposition as a constraint on any 
implementation of a systemic grammar. Hence it is not 
unreasonable to expect the realization of the form to 
conform to the decomposition of the semantic units. We 
have called this the independence criterion, indicating 
that independent decomposed functional units be real- 
ized independently in any implementation of systemic 
grammar. We argued that in order to be consistent with 
this paradigm, we have to choose appropriate struc- 
tural units as realizations of semantic/functional pieces. 
These units must capture all necessary structural rela- 
tionships, and should be the bounding domains for the 
realization operators that build them. Under these con- 
ditions, we argued that the structural units should have 
a "large" enough notion of locality to be able to factor 
out all structural dependencies. Since the elementary 
structures of TAGs are "minimally complete" to allow 
for the factoring of dependencies, we have argued that 
they are appropriate structures that can be built and 
manipulated in an implementation of systemic gram- 
mar. They also form appropriate bounding domains 
for the realization operators. Our preliminary work on 
incorporating TAGs in the systemic framework gives us 
encouragement to believe that this is indeed the case. 
Second, in addition to justifying the use of TAG 
structures for systemics, we can show that we can han- 
dle the discontinuity (which we did not explicitly dis- 
cuss for lack of space) and long distance dependency 
problem which plague other implementations of sys- 
temic grammars. The key point to make here is that 
not only are these handled but that generation of utter- 
ances with discontinuity or long distance dependencies 
is conceptually no different than generation of utter- 
ances without any form of discontinuity. 
Third, systemic grammar places emphasis on func- 
tion over form and makes clear that functional distinc- 
tions in the input manifests themselves in the different 
available forms. It is clear that our approach brings 
this aspect of systemics to the fore front. Note that the 
different trees in a tree group yield various syntactic re- 
alizations of a single predicate-argument structure. As 
we step through the network, various choices are made 
on the basis of the functional content of the planned 
utterance. These choices will result in choosing one 
syntactic realization over another. 
Finally, what we have suggested calls for putting to- 
gether two formalisms so that a mainly semantics driven 
processor (systemics) is able to reap some of the advan- 
tages of a syntax driven (TAG) approach. Currently a 
systemic network employs limited mechanism for carry- 
ing through the syntactic consequences of the decisions 
that it makes. Thus one of two things has to happen: 
1. the network designer must anticipate all syntactic 
consequences and explicitly state each of them at the 
time a decision is made. This is not an ideal solution, 
especially when the network becomes very large. 
2. the system must depend on the environment being 
consistent in order to carry out the desired conse- 
quences. In this case the syntactic consequences are 
strung throughout the network (perhaps prompted 
by different questions that are asked). The envi- 
ronment must be counted on to answer those ques- 
tions in a consistent fashion. Even if the informa- 
tion is straightforwardly captured in the environment 
(which is unclear), due to lack of the ability to carry 
out syntactic consequences, it becomes necessary to 
ask questions of the environment (to make choices) 
that are redundant. In addition, this arrangement 
goes against the systemic enterprise in which the en- 
vironment keeps track of semantic content. 
The addition of TAG allows an independent mech- 
anism (e.g., the TAG processing) to maintain consis- 
tent syntactic consequences of decisions made. For ex- 
ample, agreement constraints are precompiled into the 
tree. Also, for example, given the choice of a particular 
lexical item - once the trees for that item have been 
narrowed to one, the tree itself will contain the infor- 
mation about what functions must be expanded. Thus 
this information need not be included in the network as 
well. 
We believe that network design will be simpler be- 
cause the incorporation of TAG makes possible clear 
demarcation of semantic choices from syntactic conse- 
quences. Also it allow for the separation of lexical id- 
iosyncrasies into the lexicon rather than the network. 
Our work so far has been concerned with identifying 
the TAG structures as appropriate structural units in a 
computer implementation of a systemic grammar. The 
implementation decisions that have been discussed are 
given to indicate the logical consequences of incorpo- 
rating the TAG formalism in the systemic paradigm. 
These consequences would necessarily be handled in 
any actual implementation (e.g., breaking processing 
into regions, associating elementary trees with regions, 
the nature of realization operators). Considerable work 
remains to be done. We need to investigate the con- 
sequences of using the TAG formalism on the design 
of the systemic network especially in terms of uncov- 
ering redundancy and separation of syntax, semantics, 
and lexicon design. While currently used networks will 
be helpful in this task, we anticipate considerable revi- 
sions in the network design due to the incorporation of 
TAG. Furthermore, we have only examined some func- 
tional domains and a subset of realization operations 
that will be required. These topics are the focus of our 
current research. 
7 

References 
\[Bateman ~ Paris 1989\] Bateman, J. and Paris, C. 
1989. Phrasing a text in terms the user can under- 
stand. In: Proceedings of the 11th International Joint 
Conference on Arlificial Intelligence, IJCAI-89, De- 
troit, Michigan: 1511-1517. 
\[Davey 1978\] Davey, A. 1978. Discourse Production. 
Edinburgh University Press, Edinburgh. 
\[Fawcett 1980\] Fawcett, R.P. 1980. Cognitive linguis- 
tics and social interaction. Julius Groos Verlag Hei- 
delburg and Exter University. 
\[Halliday 1985\] Halliday, M. A. K. 1985. An introduc- 
tion to functional grammar. Edward Arnold, London 
England. 
\[Hudson 1971\] Hudson, R.A. 1971. English Complex 
Sentences: An Introduction to Systemic Grammar. 
North Holland. 
\[Joshi 1985\] Joshi, Aravind K. 1985. How Much 
Context-Sensitivity is Necessary for Chracterizing 
Structural Descriptions : Tree Adjoining Grammar. 
In: D. Dowry, L. Karttunen, and A. Zwicky, Eds., 
Natural Language Processing : Theoretical, Compu- 
tational and Psychological Perspectives. Cambridge 
University Press, New York. 
\[Mann 1983\] Mann, William C. 1983. A Linguistic 
Overview of the Nigel: Text Generation Grammar. 
Technical Report ISI/RS-83-9, ISI/USC. 
\[Mann & Matthiessen 1985\] Mann, W. and 
Matthiessen, C. 1985. Nigel: A systemic grammar 
for text generation. In: O. Freedle, Ed., Systemic 
Perspectives on Discourse. Norwood, NJ. 
\[Matthiessen & Kasper 1985\] Matthiessen, Chris- 
tian and Kasper, Robert. 1985. Representational Is- 
sues in Systemic Functional Grammar -and- Systemic 
Grammar and Functional Unification Grammar. In: 
12th International Systemic Workshop, Ann Arbor, 
Michigan, Also appears as : ISI/USC Technical Note 
RS-87-179, May 1987. 
\[Patten 1988\] Patten, T. 1988. Systemic Text Gen- 
eration as Problem Solving. Cambridge University 
Press, Cambridge. 
\[Schabes et al. 1988\] Schabes, Y., Abille, A., and 
Joshi, A. 1988. Parsing Strategies with 'Lexicalized' 
Grammars: Application to Tree Adjoining Gram- 
mars. In: Proceedings of the 12th International 
Conference on Computational Linguistics (COLING' 
88), Budapest, Hungary. 
\[Vijay-shanker & Joshi 1988\] Vijay-shanker, K. and 
Joshi, Aravind K. 1988. Feature Structure Based 
Tree Adjoining Grammar. In: Proceedings of the 12th 
International Conference on Computational Linguis- 
tics (COLING' 88), Budapest, Hungary. 
