De-Constraining Text Generation 
Stephen Beale, Sergei Nirenburg , Evelyne Viegast and Leo Wanner§ 
tComputing Research Laboratory 
Box 30001, Dept. 3CRL 
New Mexico Sate University 
Las Cruces, NM 88003-0001 USA 
sb,sergei,viegas@crl.nmsu.edu 
§Computer Science Department, IS 
University of Stuttgart 
Breitwiesenstr. 20-22 
D-70565 Stuttgart, .Germany 
wannerlo~informatik.uni-stuttgart.de 
Abstract 
We argue that the current, predominantly task-oriented, approaz~hes to modularizing text 
• generation, while plausible and useful conceptually, set up spurious conceptual and operational 
constraints. We propose a data-driven approach to modularization and illustrate how it elimi- 
nates • •the previously ubiquitous constraints on combination of evidence across modules and on 
• control. We also briefly overview the constraint-based control architecture that enables such an 
approach and facilitates near linear-time processing with realistic texts. 
-1 Introduction 
This paper addresses the area of text generation known as microplanning \[Levelt1989, Panaget1994, 
Huang and Fiedler1996\], or sentence planning \[Rambow and Korelsky1992\]; \[Wanner and Hovy1996\]. 
Microplanning involves low-level discourse structuring and marking, sentence boundary planning, 
clause•internal structuring and all of the varied subtasks involved in lexical choice, These complex 
tasks are often modularized and treated separately. The general argument is that since sentence 
planning tasks are not single-step operations, since they do not have to be performed in strict se- 
quence, and since the planner's operation is non-deterministic, each sentence planning task should 
be implemented by a separate module or by several modules (see, e.g., \[Wanner and Hovy1996\]). 
Such an argument is natural if generation is viewed as a set of coarse-grained tasks. Indeed, with the 
exception of a few researchers (\[Elhadad et a1.1997\] and the incrementalists listed below), the task- 
oriented view is standard in the generation community. Unfortunately, task-oriented generation 
sets up barriers among the components of the generation process, primarily because, in a realistic 
scenario, the tasks are intertwined to a high degree. Overcoming these barriers has become a central 
topic in generation research (see below). In our approach the basis of modularization is sought in 
the nature of the input data to the generation process, in our case, a text meaning representation, 
formulated largely in terms of an ontology. This data-oriented approach is similar to that taken by 
many incremental generators \[De Smedt1990, Reithinger1992\], although these tend to concentrate 
on syntactic processing. But see \[Kilger1997\], who explicitly addresses microplanning. We feel that 
our work provides an optimal path between task-oriented generators (which face problems due to 
the interrelationships between the tasks) and traditional incremental generation (which does not 
take advantage of problem decomposition as discussed below). 
In what follows we describe our ontology-based modularization, the kind of constraints which can 
be automatically set up within such a paradigm, and the control mechanism we employ to process it. 
We focus on the task of lexicalization, but other microplanning tasks have been handled similarly. 
We conclude with a discussion of the avoidable barriers inherent in most current approaches, along 
• 48 
Slock-40 + . 
" (~ " -- coreference wlth previous text a • + 
"Lml~rdere indlc~ted tiiII lille IPlew ¢onlpllrly, which wlU continue to be Ilst~cl on IJhe illm~k market, will be divided Into nlrl~ dlvisiomul 
which correspond to the nine working divisions of Mslrm ~.d itlachet~:" 
Figure 1: Input Semantic Representation to the Mikrokosmos Generator 
with their attempts at circumventing them, and how our approach eliminates many of the problems. 
We also point out differences between our approach and that of the incremental generators. 
2 Ontology-Based Modularization 
In contrast to modularization by tasks such as discourse structuring, clause structuring and lex- 
ical choice, the Mikrokosmos project (http://crl.nmsu.edu/Research/Projects/mikro/index.html) 
attempts to modularize on the ontological and linguistic data that serves as inputs to the text 
generation process, that is, based on the types of inputs we expect, not on the types of processing 
we need to perform. A typical semantic representation that serves as the input to the generation 
process is shown in Figure 1. This semantic input was produced by the Mikrokosmos analyzer from 
an input Spanish text. 
The generation lexicon in our approach is essentially the same as the analysis lexicon, but with 
a different indexing scheme: on ontological concepts instead of NL lexical units 1. (\[Stede1996\] is 
an example of another generator with a comparable lexicon structure, although our work is richer, 
including collocational constraints, for example). The generation lexicon contains information (such 
as, for instance, semantics-to-syntax dependency mappings) that drives the generation process, 
with the help of several dedicated microtheories that deal with issues such as focus and reference 
(values of which are among the elements of our input representations). The Mikrokosmos Spanish 
core lexicon is complete with 7000 word senses defined; the English core lexicon is still under 
development with a projected size of over 10,000 word senses. Both of these core lexicons can be 
expanded with lexical rules to contain around 30,000 entries (\[Viegas et al.1996\]). 
Lexicon entries in both analysis and generation can be thought of as "objects" or "modules" 
corresponding to each unit in the input. Such a module has the task of realizing the associated 
unit, while communicating with other objects around it, if necessary (similar to \[De Smedt1990\]). 
1In semantic analysis, the input is a set of words, thus the lexicon is indexed on Words. In generation, the input 
is concepts, So it is indexed on concepts. 
49 
i 
STOCK-MARKET 
##o~anti¢s \] f ~.~ tl . STOCK-MARKET s~ n ~s ~l~'.'ctcd STOCK-MAIFt KET - #x.oG¢'~d ~ \[.,~'AI~\[OI ~" L 
\[ .= ..... \] 
Figure •2: OBJECT Lexicon Entries - A Simplified View 
Each module can be involved in carrying out several of the •tasks like those listed by Wanner 
and Hovy. For • instance, modules for specific •events or properties are used in setting up clause and 
sentence structures as well as lexical choice, as will be shown below. Interactions and constraints • 
flow freely, with the control mechanism dynamically tracking the connections 2. One outcome of 
this division of labor between declarative data and the control architecture is that the bulk of 
knowledge processing resides in the lexicon, indexed for both analysis and generation. This has 
greatly simplified knowledge acquisition in general \[Nirenburg et a1.1996\] and made it easier to 
• adapt analysis knowledge sources to generation as well as to convert knowledge sources acquired 
for one language to use with texts in another. 
Below we sketch out how this organization works. We begin by describing the • main types of 
lexicon entries with the goal of demonstrating how each performs various generation •tasks. We 
then take a look at the different types of constraints associated with each kind of entry. 
I 
I 
I 
I 
I 
I 
I 
I 
2.1 Types of Lexicon Entries 
The main types of lexicon entries correspond to the ontological categories of OBJECTS, EVENTS 
and PROPERTIES (for simplicity, we will avoid discussion of synonyms and stylistic variations): 
• Objects. In English, Objects are typically realized by nouns, although the actual mapping might be 
rather complex \[Beale and Viegas1996\]. In general, object generation lexicon entries can have one-to- 
one mappings between concept and lexical unit, or can contain additional semantic restrictions, both 
of which are illustrated in Figure 2. The use of collocational information is described below. • 
• , Events. • Events, as shown pictorially in Figure 3, can be realized as verbs ("divided") or nouns 
("division") in English.• Furthermore, the lexicon entries for events typically determine the structure 
• of the nascent clause by mapping the expected case roles into elements of the verb subcategorization 
frame (or other structures). Rules for planning passives and relative clauses, for instance, are also 
available. These rules can be used to fulfill different grammatical and clause combination requirements 
as described below. Conceptually, all the entries produced by these rules can be thought of as being 
resident in the lexicon. Practically speaking, many of them can be produced on the fly automatically, 
reducing the strain on knowledge acquisition. 
• Properties. Properties 3 are perhaps the most interesting of the input types discussed here because 
• •• they are so flexible. They can be realized as adjectives , relative clauses, complex noun phrases and 
• complete sentences. Often a property is included in the definition of another object or event, such as 
in Figure 2, where the LOCATION is included in the object entry. CASE-ROLE-RELATIONS 
2Alth0ugh our constraint-based planner supports truth maintenance operations, in the "fuzzy" domain of natural 
language semantics it is often more appropriate to speak of "preferences" 
SRELATION and ATTRIBUTE are the subtypes of PROPERTY. 
50 
DIVIDE 
\[ $emnglc$ 
expected 
p~ssIvo 
nomlnalization 
rel~tlveB 
.ubj VAt-1 
• dlvld ol~ VAR2 IBM divided Apple into 
pp4c~u~• sg~ conapanh~s 
root ¢'\[~m o" 
°bJ~t VAR3 
#u~ VAR2 
dW'llde (ix~s|) 
pp4,d,I u "tr~m'" ,Applm w,w.$ dlvkled Lnto ~ix 
°kilt VAtZa comi~si~ bJt IBM. 
PP-~eUam~t (opo, omm~ . 
soot "by" 
• °bJ~t VAILI 
diwLcdon of,Apple Into ... 
IBM, who dlvdded Apple gngo ... 1 J 
sg~ companies, which IBM divided --. \] J 
Mkppl¢. whkh was divided Into ... \[ J 
Figure 3: EVENT Lexicon Entries •A Simplified View 
ORGANIZATION-INVOLVED-IN 
i gemaant~s ~ ~ng~t i~\] VARI expected generated root "him" The co•pony has stoc~ org~l~Vedo~ 
obj VAI~ 
~¢ VAR2 Tke company's s~ck 
.... p...v, .h~,h has ,~.k \] 
the coml~tny , whose •took I 
• Figure 4: PROPERTY Lexicon Entries - A Simplified View 
typically are consumed by the event entry, except in the case of some nominalizations. DISCOUI~SE- 
• RELATIONS contribute to setting up sentence boundaries, sentence ordering and pronominalization. 
Figure 4 is an example of RELATION. 
• 2.2 Constraints • in Sentence Planning 
The above generation lexicon entries are the primary knowledge sources used in the generation 
process 4. Five different types of constraints are automatically generated which constrain the combi: 
nations of entries allowed to globally realize a semantic input. The Mikrokosmos control mechanism 
efficiently processes constraints to produce optimal global answers. 
Binding Constraints. One of the primary advantages of input-based modularization is that the 
individual knowledge sources (lexicon entries) can be grounded in the input they expect to be 
matched against. For instance, in Figure 3, the semantic• input expected shows three variables, 
corresponding•to the three case roles normally associated with a DIVIDE event. The process of 
linking these variables to the actual semantic structures for a particular input is known as binding. 
4Due to space limitations, we are glossing over important generation microtheories such as sentence boundary 
determination and corderence implementation. 
51 
ASSERTIVE ACT 
root SAY 
~mp 
roo¢ "rl~A'r 
obJ woog 
VAIRL2 (~l~u~) 
CONSTRAINT PRODUCED: 
VAI~2 : clause ~ 
r binding 
DIVIDE-31 : clause 
Figure 5: Grammatical Constraints 
STOCK-MARKET \] 
I • I/ I 
• • PREFER 
LOCATION 
I i.-pp (,o~,o I 
~- on-pp Our.face) \] 
I at-pe a,l,~e~ I 
Figure 6: Collocational Constraints in Lexicon Entries • 
For the input shown in Figure 1, VAR2 will be bound to CONGLOMERATE-32 and VAR3 
will be bound to CORPORATION-34.• 
• Notice that, for this example, no AGENT exists for the DIVIDE-31 event, so that VAR1 
will be left unbound. Binding constraints will simply eliminate any syntactic choices that contain 
non-optional unbound variables. In this case, it will rule out the first syntactic realization for 
DIVIDE shown in Figure 3. 
The grounding of the input afforded by the binding process also allows us to simplify the other 
types of constraints described below. Each of these types of constraints ~ automatically processed 
in our system, in task-based systems typically require complex rules to be acquired manually. 
Grammatical Constraints. An example of a grammatical constraint is shown in Figure 5. A 
lexicon entry can specify grammatical constraints on the realization of any of the variables in it. 
One •possible syntactic realization for ASSERTIVE-ACT is shown. It requires its VAR2 to be 
realized as a clause. This particular entry allows the system to produce "John said that Bill went 
to the store" but not "John said that Bill." A comparison with Figure 1 shows that the binding 
process will link VAR2 of the ASSERTIVE-ACT entry to DIVIDE-31. In effect, the•resulting 
constraint will eliminate any realization for DIVIDE-31 (in Figure 3) that does not produce a full 
clause at the top-level, through nominalization and relativization. It should be stressed that this 
filtering occurs only in conjunction with the given •realization of ASSERTIVE-ACT; there may 
be other realizations that would go fine with, for example, a nominalized realization of DIVIDE. 
CoUocational Constraints. Figure 6 illustrates the familiar notion of collocational constraints. 
Again, the fact that the lexicon entryis grounded in the input allows a simple representation 
of collocations. In this case, the different realizations of LOCATION usually correspond to the 
semantic type Of the object. Collocations :can be used to override the default. The co-occurrence 
zone of the STOCK-MARKET entry simply states that if it is used as the range of a LOCATION 
52 
I 
I 
I 
I 
ii 
ii 
I 
i 
*| • I ! 
. ~ case ro/e 
Situation 1: Link through 
discourse relation 
Situation 2: Link through 
• case role relation 
I 
Situation 3: Implicit link, 
Figure 7: Inputs that May Lead to Clause Combinations 
relation, then the LOCATION relation should be introduced with "on." This produces an English 
collocation such as "the stock is sold on the stock market" as opposed to the less natural "... sold at 
the stock market." Notice that no additional work on collocations needs to be performed beyond 
the declarative knowledge encoding. The constraint-based control architecture will identify and 
assign preferences to collocations. 
Clause Combination Constraints. Various kinds of constraints arise when clauses are com- 
bined :to form complex sentences. The strategies for clause combination come from three sources: • 
• Directly from a lexicon entry associated with an input. For example, a discourse relation such as 
CONCESSION might directly set up the syntax to produce a sentence structure such as "Although 
Jim admired her reasoning, he rejected her thesis." 
• Verbs which take complement clauses as arguments also set up complex sentence structures and impose 
grammatical constraints (if present) on the individual clause realizations: "John said that he went to 
the store" or "John likes to play baseball." 
• Indirectly, from a language-specific source of clause combination techniques (such as relative clause 
formation or coordination in English). 
These three sources correspond to the three input situations depicted in Figure 7. The first two 
have explicit relations linking two EVENTs. The first (the non-case-role relation) will have a 
• corresponding lexicon entry which directly sets up the sentence structure, along with specific con- 
straints on the individual clauses. The second possibility typically occurs with EVENTs that 
take complement clauses as case-role arguments. The lexicon entries for these usually will specify 
the complex clause structure needed. The third situation has no explicit connection in the input; 
therefore, some sort of language-specific combination strategy must be used to fill the same task. 
Even though the latter case appears to be a situation that requires a task-oriented procedure, 
in reality it is as easy to use general purpose structure constraints along with a declarative repre- 
sentation of possible transformations available. Assuming, for the sake of illustration, that due to 
some external reason a single sentence realization of two clauses is preferred 5, a general purpose 
structural constraint prevents two clauses from embedding a single referent into distinct syntactic 
structures. For instance, 1 and 2 below are grammatical, but 3 is not, because both the clauses try 
to use "conglomerate" as their subject. 
1. The conglomerate, whose stock is sold on the stock market, was divided into nine corporations. 
SConstraints which might produce such a preference can come from a variety of Sources; a common one is the 
realizations of discourse relations. 
53 
2. The conglomerate , which was divided into nine corporations, is sold on the stock market. 
3. *The conglomerate was divided into nine corporations is sold on the stock market. 
The general purpose constraint will automatically prevent such a realization and trigger the con- 
sideration of subordinate clause transformations. 
In addition, the examples of clause combination given above and in Figure 7 all contain e~amples 
of coreference across clause boundaries. Although coreference realization has its own microtheory 
that is triggered by instances of coreference in the •text, clause combination techniques may interact 
with it. For instance, the lexicon entry for a RELATION might specify that a pronoun be used 
in the second clause. 
The important thing to note for this presentation is that these types of constraint are either 
directly found in the lexicon or are produced automatically by the planner. Special situations such 
as coreference can be easily identified because the lexicon entries ar e grounded in their inputs. This 
method appears to be much simpler than those needed by task-based generators. 
Semantic Matching Constraints. Matching constraints take into account the fact that, first 
of all, certain lexicon entries may match multiple elements of the input structure and, secondly, 
that the matches that do occur may be imperfect or incomplete. 
In general, the semantic matcher keepstrack of which lexicon entries cover which parts of the 
input, which require other plans to be used with it, and which have some sort of semantic mismatch 
with the input. The following sums up the types of mismatches that might be present, each of which 
receives a different penalty (penalties are tracked by the control mechanism and help determine 
which combination of realizations is optimal): 
• slots present in input that are missing in lexicon entry -> undergeneration penalty, plan separately 
• extra slots in lexicon entry ~ overgeneration penalty 
• slot filler discrepancies (different, or more or less specific) 
- constant filler values 
HUMAN (age 13-19) - i.e. "teenager" from English input vs. 
HUMAN (age 12-16) i.e. "age b~te" in French lexicon 
- concept fillers 
HUMAN (origin FRANCE) vs. 
HUMAN (origin EUROPE) 
• A more detailed explanation of these issues is presented in \[Beale and Viegas1996\]. The impor- 
tant thing to note here is that input-based modularization in our knowledge sources enables this 
type of constraint to be tracked automatically. In combination with the other constraints described 
above, we Can avoid the complex mechanisms needed by task-based generators for interacting real- 
izations of input semantics. : 
3 Efficient Constraint-based Processing • 
The Mikrokosmo s project utilizes an efficient, constraint:directed control architecture called Hunter- 
Gatherer (HG). \[Beale et a1.1996\] overviews how it enables semantic analysis to be performed in 
near linear-time. Its use in generation is quite similar. \[Beale1997\] describes HG in detail. 
Consider Figure 8, a representation of the constraint interactions present in a section of Figure 
1. Each label, such as DIVIDE, is realizable by the set of choices specified in the lexicon. Each 
54 
. 
I - ! 
Figure 8: Problem Decomposition 
* a 
J 
Figure 9: Sub-problem i 
! 
i 
l 
! 
i 
! 
I 
I 
I 
solid line represents an instance of one of the above constraint types. For example, DIVIDE and 
ORG-INVOLVED-IN are connected because of the structural constraint described above (they 
both cannot set up a structure which nests the realization of CONGLOMERATE=32 into different 
subject positions). 
The key to the efficient constraint-based planner Hunter-Gatherer is its ability to identify con- 
straints and partition the overall problem into relatively independent subproblems. These subprob- 
lems are tackled independently and the results are combined using solution synthesis techniques. 
This "divide-and-conquer" methodology substantially reduces the •number of combinations that 
have to be tested, while • guaranteeing an Optimal answer. For example, in Figure 8, if we assume 
that each node had 5 possible choices (a conservative assumption), there would be 51°, or almost 
10 million combinations of choices to examine. Using the partitions shown in dotted lines, however, 
HG only examines 1200 combinations, In general, HG is able to process semantic analysis and 
generation problems for natural language in near linear-time \[Beale et a1.1996\]. = 
While a detailed explanation of Hunter-Gatherer is beyond the scope of this paper, • it is fairly 
easy to explain the source of its power. Consider Figure 9, a single subproblem from Figure 8. 
The key thing to note is •that, of the three nodes, BUY, LOCATION and STOCK-MARKET, 
only BUY is connected by constraints to entities outside the subproblem. This tells us that by 
looking only at this subproblem we will not be able to determine•the optimal global choice for 
BUY, •since there are constraints we cannot take into account. What we can do, howeve r , is, for 
each possible choice for BUY, pick the choices for LOCATION and STOCK-MARKET that 
optimize it. Later, when we combine the results of this subproblem with other subproblems and 
thus determine which choice for BUY is optimal, we will already have determined the choices for 
LOCATION and STOCK-MARKET that go best with it. 
The following sums up the advantages Hunter-Gatherer has for text generation: 
55 
• Its knowledge i s fully declarative. Note that this is allowed by unification processors \[Elhadad et a1.1997\], 
but HG gives the added benefits of speed and capability of "fuzzy" constraint processing. 
• It allows "exhaustive" enumeration of local combinations. 
• It eliminates the need to make early decisions. 
• It facilitates interacting constraints, and accepts constraints from any source, .while still utilizing 
modular, declarative knowledge. - " 
• It guarantees optimal answers (as measured by preferences). 
• It is very fast (near linear-time). 
4 Comparison to Other Generation Systems 
Related work • exists in two areas: (i) the processing strategy of microplanning tasks, and (ii) the 
nature and organization of resources used by the microplanner. 
There is a strong tendency in generation to deal with microplanning tasks in a small •number 
of modules, •which are either structurally or functionally motivated. However, it is recognized that 
many of the tasks are highly intertwined, so that, in principle, the modules should run in parallel 
and nearly Constantly exchange information. We consider this as a clear hint that a coarse-grained, 
task-oriented division of microp!anning sets up artificial barriers. Repeated efforts of researchers 
to try and breach those barriers confirm our view. 
\[Elhadad et a1.1997\] recognizes that constraints on lexical choice come from a wide variety of 
sources and are multidirectional, making it difficult to determine a systematic ordering in which 
they should be taken into account. They propose a backtracking mechanism within a unification 
framework to •overcome the problem. \[Rubinoff1992\] is perhaps the most strongly focused on this 
issue. He argues that the accepted division into components "ultimately interferes with some of the 
decisions necessary in the generation process." He utilizes annotations as a feedback mechanism to 
provide the planning stages with linguistically relevant knowledge. 
• Another area of research that belies the unnatural task-based division widely accepted by.text 
generation researchers today is the attempts to control sentence planning tasks. \[Nirenburg et al. 1989\] 
and more recently, \[Wanner and Hovy1996\] advocate a blackboard control mechanism, arguing that 
the order of sentence planning tasks cannot be pre-determined. Behind this difficulty is the real- 
• ity that different linguistic phenomena have different, unpredictable requirements. Grammatical, 
stylistic and collocati0nal constraints combine at unexpected times during the various tasks of sen- 
tence planning.• Blackboard architectures, theoretically, can be used to allow a certain thread of 
operation to suspend operation until a needed bit of information is available. Unfortunately, in 
the best case, such an architecture is inefficient and difficult to control. In practice, such systems, 
as is admitted in both papers above, resort to a •"default (processing) sequence for the modules" 
along with a simplistic truth-maintenance system which ultimately becomes a fail-and-backtrack 
type of control, completely negating the spirit of the blackboard• system. While these shortcomings 
might eventually be •overcome, the fact remains that it was the unnatural division into tasks that 
necessitated •the blackboard processing in the first place. 
• In this paper, we propose an input data-oriented division of the microplanning task--similar 
to the way many incremental generators \[De Smedt1990, Reithinger1992, Kilger and Finkler1995\] 
divide the task of surface processing. However, the processing of input units as done by the Hunter- 
Gathere r ---our microplanning enginc differs significantly from the processing in the incremental 
generators cited.• Thus, an important feature of HC is that it possesses a strategy for dividing 
the problem of verbalizing a semantic structure into relatively independent subproblems. The 
55 
I 
!I 
I 
I 
il 
I 
ii 
I 
I 
I 
subproblems can be of different size. Into which subproblems the problem is divided depends on 
constraints that hold between units in the input structure. This strategy •greatly contributes to the 
efficiency of HG. In traditional incremental generators, a unit in the input structure is considered to 
be a subproblem. Furthermore, HG is bidirectional, i.e., it is usable for both analysis and generation. 

References 
\[Be.ale and Viegas1996\] S. Beale and E. Viegas. 1996. Intelligent planning meets intelligent planners. Pro- 
eeedings of the Workshop on Gaps and Bridges: New Directions in Planning and Natural Language Gen- 
eration, ECAI'96, Budapest, pages 59--64. 
\[Beale et a1.1996\] S. Beale, S. Nirenburg, and K. Mahesh. 1996. Hunter-gatherer: Three search techniques 
integrated for natural language semantics. Proc. Thirteenth National Conference on Artificial Intelligence 
(AAAI96), Portland, Oregon. " : 
\[Beale1997\] S. Beale. 1997. Hunter-gatherer: Applying constraint satisfaction, branch-and-bound and solu- 
tion synthesis to computational semantics. Ph.D. Diss., Program in Language and Information Technolo- 
gies, School of Computer Science, Carnegie Mellon University. 
\[De Smedt1990\] K: De Smedt. 1990. IPF: An Incremental Parallel Formulator. In R. Dale, C.S. Mellishl 
and M. Zock, editors, Current Research in Natural Language Generation. Academic Press. 
\[Elhadad et a1.1997\] M. Elhadadl J. Robin, and K. McKeown. 1997. Floating constraints in lexical choice. 
Computational Linguistics (2), 23:195-239. 
\[Huang and Fiedler1996\] X. Huang and A. Fiedler. 1996. Paraphrasing and aggregating argumentative text 
using text structure. In Proc. of the 8th INLG, Herstmonceux. 
\[Kilger and Finkler1995\] A. Kilger and W. Finkler. 1995. Incremental Generation for Real-Time Applica- 
tions. Technical Report RR-95-11, DFKI. 
\[Kilger1997\] A. Kilger. 1997. Microplanning in Verbmobil as a Constraint-Satisfaction Problem. In DFKI 
Workshop on Generation, pages 47-53, Saarbriicken. 
\[Levelti989 \] Willem J.M. Levelt. 1989. Speaking. The MIT Press, Cambridge, MA. 
\[Nirenburg et a1.1989\] S. Nirenburg, V. Lesser, and N. Nyberg. 1989. Controlling a language generation 
planner. Proc. of IJCAI-89, pages 1524-1530. 
\[Nirenburg et al.1996\] S. Nirenburg, S. Beale, S. Helmreich, K. Mahesh, E. Viegas, and R. Zajac. 1996. Two 
principles and six techniques for rapid mt development. Proc. of AMTA-96. 
\[Panaget1994\] F. Panaget. 1994. Using a Textual •Representation Level Component in the Context of 
Discourse or Dialogue Generation. In Proceedings of the 7th INLG, Kennebunkport. 
\[Rambow and Korelsky1992\] Oi Rambow and T. Korelsky. 1992. Applied text generation. Applied Confer- 
ence on Natural Language Processing, Trento, Italy. 
\[Reithinger1992\] N. Reithinger. 1992. Eine parallele Architektur zur inkrementellen Generierung multi- 
modaler DialogbeitrSge. Infix Verlag, St. Augustin. 
\[Rubinoff1992\] R. Rubinoff. 1992. Integrating text planning and linguistic choice by annotating linguistic 
structures. Proc. 6th international Workshop On Natural Language Generation, 7¥ento, Italy. 
\[Stede1996\] M. Stede. 1996. Lexieal Semantics and Knowledge Representation in Multilingual •Sentence 
Generation. Ph.D. thesis, University of Toronto. 
\[Viegas et a1.1996\] E. Viegas, B. Onyshkevych, V. Raskin, and S. Nirenburg. 1996. From submit tO submitted 
via submission: on lexical rules in large-scale lexicon acquisition. In Proceedings of the 3gth Annual meeting 
of the Association for Computational Linguistics, CA. 
\[Wanner and Hovy1996\] L. Wanner and E. Hovy. 1996. The healthdoc sentence planner. Proc. Eighth 
International Natural Language Generation Workshop (INLG-96). 
