Using a textual representational level component in the context of 
discourse or dialogue generation 
Franck Panaget* 
Centre National d'Etudes des T@l@communications (CNET) 
LAA/TSS/RCP 
Route de Tr@gastel -- 22300 Lannion - France 
email: panaget@lannion.cnet.fr / tel: +33 96.05.28.52 / fax: +33 96.05.35.30 
Abstract 
A natural language generation system is typically con- 
stituted by two main components: a content planning 
component (e.g., text planner or dialogue act planner) 
and a linguistic realization component. But, this is not 
sufficient since, on the one hand, the message built by the 
content pldnning component is generally not adequately 
detailed in order to control the many possibilities for its 
expression and, on the other hand, the content planner 
cannot influence the way in which the message will be 
verbalized. 
Generation systems require a third component, called 
the micro-planning (or sentence planning or phrasing) 
component, which acts as an intermediary between the 
pragmatico-semantic level and the purely syntactic level. 
The micro-planner is responsible for transforming the 
message into a textual structure. For this transformation 
to be achieved, grammatical and lexical resources must 
be selected. 
1 Introduction 
Traditionally, the architecture of a generation system is 
either pipelined (e.g., TEXT \[McKeown 85\]), interleaved 
(e.g., PAULINE \[Hovy 90\]) or integrate.d (e.g., KAMP 
\[Appelt 85\]). Whatever the approach, the system is di- 
vided into two main components: a message planning 
component and a linguistic realization component. The 
former is responsible for the selection and organization 
of the information to be conveyed; it builds a structure 
often called message. The latter involves the formula- 
tion of the message in grammatically correct sentences. 
In a context of discourse, the message planning compo- 
nent is a text planner and the message is a text plan 
*This work was realized at the lstituto per la Ricerca Scientifica 
e Tecnolo#ica (IRST) in Trento (Italy). 
(e.g., see \[Maier & Not 93\]). In a context of dialogue, the 
message planner is a 'fictitious' component consisting of a 
dialogue act planner, which precedes the generation sys- 
tem, and a planner for surface linguistic acts, which is a 
component of the generation system. The dialogue act 
planner builds plans of dialogue acts. A plan of dialogue 
acts is a tree-like structure in which non-leaf nodes are. 
macro-acts (e.g., "disjunction" or "sequence" of dialogue 
acts) and leaf nodes are dialogue acts (see \[Sadek 91\]). 
A dialogue act conveys information about its type (e.g., 
"inform", "yn-question" or "confirm"), its propositional 
content (i.e., a semantico-conceptual representation of the 
information to be communicated), and the interlocutors. 
The surface linguistic act planner transforms a plan of 
dialogue acts into a plan of surface linguistic acts 
(i.e., the message). In particular, the dialogue act type 
is transformed into a surface linguistic act type such as 
"assertion", "answer" or "question", and the semantico- 
conceptual content into a semantic propositional content 
(cf. figure 1 and section 3). 
One of the problems which appear in traditional gen- 
eration systems concerns the message itself: it is under- 
specified with respect to the language. On the one hand, 
the message planner cannot influence the way in which 
the message will be verbalized, and, on the other hand, 
the message is not detailed enough for the many possibil- 
ities for its expression to be controlled. Figure 1 shows 
several messages (a text plan and three surface linguistic 
acts corresponding to the same dialogue act) and their 
different verbalizations. 
A second problem is related to the fact that most gen- 
eration systems produce texts successfully since they im- 
plicitly ensure that the messages are expressible. For 
instance, in the case where the message is a text plan, 
each atomic unit in the message must be a proposition, 
and thus can always be realized as a clause. Each unit 
can be independently translated into the language us- 
ing the linguistic realization component, since there are 
127 
7th International Generation Workshop • Kcnncbunkport, Maine * June 21-24, 1994 
Text plan: (rl / result 
nucleus (el / shoot 
agent john 
patient mary) 
satellite (e2 / ldll 
agent john 
patient mary)) 
(Aa) John shot Mary. He killed her. 
(A2) Mary was shot by John. He killed her. 
(A3) Mary was shot by John. She was killed. 
(A4) John shot Mary killing her. 
(As) John shot Mary who was killed. 
Dialogue act: <S, Inform(U, ,x arrival-airport(x, ,y flight(y) ^ departure-time(y, 10:00)) = JFK)> 
This dialogue act can be transformed into three distinct "assertions" or "answers". 
• Content 1 = a destination-event whose theme is "the 10:00 flight" and destination is "3FK": 
(BA1) The lO:OO flight goes to JFK. 
(BA2) It goes to JFK. 
.(BA3) To JFK. 
• Content 2 -- an identity between "the arrival airport of the 10:00 flight" and "JFK": 
(Bin) The arrival airport of the 10:00 flight is JFK. 
(BB2) The arrival airport is JFK. 
(BB3) JFK. 
• Content 3 = a destination-event whose theme is "the hearer" and destination is "JFK": 
(Bca) You will arrive at JFK. 
(Be2) At JFh'. 
Figure 1: Messages and their possible verbalizations 
few restrictions on how to connect clauses. In particular, 
clauses can be connected with coordinate or subordinate 
conjunctions or can be simply realized as separate sen- 
tences. However, this kind of approach does not use the 
full expressive power of the language in which units may 
be much more tightly composed. Moreover, as argued 
in \[Hovy 92\], a two-component generation system is not 
satisfactory in producing 'good quality' texts since too 
many phenomena are ignored or processed in a simplis- 
tic way, for example, anaphora planning, choice of lexical 
items, aggregation of clauses and noun phrases, selection 
of grammatical constructions (e.g., the choice of the verb 
voice), and segmentation into sentences (i.e., single sen- 
tence vs. separate sentences, hypotaxis vs. parataxis, 
and determination of clause order). The main difficulty 
in taking these phenomena into account is due to the fact 
that they are 
"supra-grammatical but not really related to con- 
tent selection and organization" \[Hovy 92, p. 2\]. 
In fact, Hovy argues that these phenomena should be 
treated after the message planning and before the lin- 
guistic realization. 
A third problem is related to the fact that in a two- 
component generation system the conceptual decisions 
are taken by the message planner and the the linguis- 
tic decisions by the realization component. Since concep- 
tual and linguistic decisions are dependent on each other 
(see \[Danlos 87\]), the two components should be able to 
communicate with each other. On the one hand, the defi- 
nition of such a communication is a complex task, due to 
the gap between the two activities (see \[Meteer 92\]). On 
the other hand, all the grammatical details that can be 
provided by a linguistic realizer may not be useful to take 
these decisions. We think that, in a first step, it is use- 
128 
! 
7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 
sem-cat 
  named-object unique-object --object instance-of-a-kind . sample-of-a-kind 
definite-set 
r---result-object r----object -~ 
l process-object 
/ V state 
event ~time-anchored-event --~process-event 
/ P-'event-activ~':: ansiti°n-even~ 
t---activity --~process-activity 
manner 
_~property 
__quality ~fra~e-adverbial 
t---durative-adverbial 
Figure 2: Partial hierarchy of semantic categories 
ful to know that such linguistic phenomena exist in the 
target language without being able to realize them. For 
example, it is useful to know that a present participial 
clause can verbalize the rhetorical relation "result", but 
it is not useful to know (in a first step) that a present 
participial clause is formed by adding "ing" at the end of 
the stem of a verb. 
2 Micro-planning component 
In order to solve the above mentioned problems, a compo- 
nent can be added between the message planner and the 
linguistic realizer (e.g., see \[Levelt 89\] or \[Hovy 92\]). The 
role of this third component, called the micro-planner 
in \[Levelt 89\], is to plan the best informational perspec- 
tive that can be given to the message. The component 
often uses an intermediate representation level. For ex- 
ample, in the system Epicure \[Dale 88\], a recoverable se- 
mantic structure, which specifies the underlying content 
of the utterance to be generated, and an abstract syntac- 
tic structure, which is much closer to the surface syntax 
of the linguistic form which will be produced, are used. 
Dale also shows how the two levels of representation play 
a role in the generation of elisions and one-anaphoric ex- 
pressions. However, the abstract syntactic level of most 
generation systems is often hazy and dependent on the 
syntactic formalism used by the linguistic realization com- 
ponent. 
Our intermediate structure, which we call textual 1 
structure, is expressed in terms of abstract linguis- 
tic resources \[Meteer 92\] 2. These resources are abstrac- 
tions over concrete resources of language (e.g., syntactic 
structures, words, and grammatical features); they reflect 
linguistic constraints while abstracting away from syntac- 
tic details. They are useful for generation systems since 
allowing a generation system to select the con- 
crete resources.., makes available many more 
degrees of freedom than the language actually 
permits \[Meteer 92, p. 67\]. 
Abstract linguistic resources become the terms in which 
the micro-planning component takes its decisions and 
that way they constrain the system to only do what is ex- 
pressible in the language since it is not allowed to choose 
them independently. We made some modifications on 
two kinds of abstract resources defined by Meteer: we 
improved her semantic categories by using perspectives 
from systemic functional linguistics (cf. section 2.1) and 
we extended the notion of resource trees (cf. section 2.2). 
l"Textual" is used in the sense of the textual metafunc- 
tion in the systemic functional theory (see below). For pre- 
sentations of systemic functional linguistics, see \[Halliday 85\] or 
\[Matthiessen & Bateman 91\]. 
2Meteer used the abstract linguistic resources to build a struc- 
ture which she called text structure. We do not use the same name 
because Meteer's structure could exist at the same level of rhetor- 
ical structure theory (i.e., tile message planning level) whereas our 
structure is complementary to this level. 
129 
7th International Generation Workshop ° Kennebunkport, Maine • June 21-24, 1994 
Moreover, we explicitly defined lezical resources (cf. sec- 
tion 2.3). 
2.1 Semantic hierarchies 
Semantic categories (cf. figure 2) reflect pertinent combi- 
nations of linguistic features (e.g., syntactic features and 
lexical items). For example, the five sub-categories of 
object s are specified from definiteness (i.e., definite or 
indefinite) and number (i.e., singular or plural). These 
categories constrain the composition of the utterance by 
defining what kinds of constituents can be extended and 
how they can be composed. For example, an adjunct 
constituent of type manner can be added to an event 
(e.g., Peter suddenly decided to go to the beach) whereas 
an adjunct of type property cannot (e.g., *Peter impor- 
tantly decided to go to the beach4); a process-event can 
be composed with a durative-adverbial (e.g., Karen 
watched TV for an hour) whereas a transition-event 
may be composed of a frame-adverbial (e.g., Karen 
solved the problem in one hour). 
On the one hand, we think that according to sys- 
temic linguistic theory s, the hierarchy of semantic cat- 
egories mikes ideational constraints and textual con- 
straints; both are semantic, but semantic of rather dif- 
ferent kinds. Thus, while it is correct that the categories 
event, manner, and property are semantic (ideational) 
categories, taking an event perspective on something is a 
textual decision, e.g., 
"The ship sank." 
is an ideational event, and it is textually presented within 
an EVENT-PERSPECTIVE. 
"The sinking of the ship" 
is still an ideational event, but it is now presented within 
an OBJECT-PERSPECTIVE. Even clearer are cases such 
as definite and indefinite noun phrases. Thus the dif- 
ferences between "a man" and "the man" can be char- 
acterized as follows. They are both ideational of cate- 
gory PERSON. The noun phrase '% man" presents a dis- 
course entity of that ideational semantic category which 
is textually UNIDENTIFIABLE. This is a category drawn 
from textual semantics. Textual semantic classifications 
of discourse entities change as a text develops. The noun 
phrase "the man" is classified textually as IDENTIFI- 
ABLE. The former is realized grammatically (largely on 
3The semantic categories are noted in typewriter text. 
4FoUowing the general convention in linguistics, we use "*" to 
mark magrarrmaatical sentences. 
sin this theory, language is organized functionally according to 
three metafunctions, concerned with the representation of experi- 
ence (ideational), symbolic interaction between interlocutors (in- 
terpersonal), and presentation of information as text in context 
(textual). 
the basis of the textual semantic classification) as an in- 
definite noun phrase, the latter as a definite noun phrase. 
In this way, ideational semantics and textual semantics 
are complementary to each other and both help constrain 
grammatical realization. 
On the other hand, we think that the textual part of 
Meteer's hierarchy is too abstract to treat phenomena 
that need syntactic information such as anaphora plan- 
ning and selection of grammatical constructions (see also 
the other examples of linguistic phenomena given in sec- 
tion 1); in particular, syntactic information such as gen- 
der or verb voice is missing. 
In order to solve these problems, we propose to re-- 
place the hierarchy with two constituent hierarchies: an 
ideational semantic hierarchy and a textual semantic hi- 
erarchy. The former is based on the Upper Model de- 
fined in \[Bateman et al. 90\]. The latter was developed 
by ourselves; we call it the hierarchy of textual semantic 
categories. These two hierarchies are not incompatible 
since 
\[a\] language.., embodies a multiplicity of con- 
stituent hierarchies, coexisting in different parts 
of the system. They are not unrelated: on the 
contrary, lhey are all different facets of the same 
phenomenon. \[...\] At the same time, each of 
these hierarchies is independent of the others, 
and lhere are infinite possibilities of matching 
them up in a meaningful way \[Halliday 85, p. 
18\]. 
The Upper Model provides a domain-independent, 
general and reusable conceptual organization which is 
used to classify the world knowledge when linguistic pro- 
cessing is to be performed. This general semantic taxon- 
omy can also be seen as an inheritance hierarchy that 
organizes concepts according to how they can be ex- 
pressed (cf. figure 3). The Upper Model also'expresses 
(ideational) semantic restrictions on arguments required 
by entities it involves. For example, the realization-event 6 
requires two participants: an agent that must be an an- 
imate and a patient that must be a product. The Upper 
Model also provides the type of qualities which may be 
ascribed to particular concepts. For example, the Up- 
per Model expresses that the destination relation relates 
a place object to an event. 
The hierarchy of textual semantic categories 
is a property inheritance network that represents tex- 
tual knowledge of the target language. Figure 4 shows 
a part of our hierarchy; the whole hierarchy contains 
about eighty categories. The definition of the textual 
semantic categories is based on an analysis of French 
\[Bescherelle 90\] and of English \[nalliday 85, Shinghal 92\]. 
6'rhe ideational semantic categories are noted in italic text. 
130 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
urn-thing 
V-- realization-event 
\[ ~gent < ~nim~te> ~ process ,,, 
event --1 patient <product> 
t_.- destination-event 
theme < phyal¢~l> 
-- COnCept destination <pl~ce> 
\[--..- location time 
thing ~ physical animate 
L_. ideal product   agent 
domain <event> 
re, ride <~nim~te> 
participant patient 
domain <process> 
range <thing> 
m relation theme 
domain <process> 
r~nse < physi¢~l> 
circumstance destination 
domain <event> 
rsnse < pl&ce> 
Figure 3: Partial Upper Model 
tsc - 
text 
complex-clause 
object-perspective 
-- clause ~ 
event-perspective 
clause-nucleus 
-- time-anchored-clause 
\[ infinitive-activity i--"- activity 
participial-activity 
intransitive-nucleus 
~ irect-transitive-nucleus 
bitransitive-nucleus _  
subject-argument 
direct-argument 
indlrect-argument 
other-argument 
clause-nucleus-argument 
C present-activity 
past-activity 
direct-transitive-passive-nucleus 
C direct-transitive~active-nucleus 
unidentifia ble-object f--- textual-object 
-L._. identifiable-object named-object 
plural-neuter-nucleus textual-object-nucleus 
singular-feminine-nucleus 
complex-clause-modifier 
h modifier ~ clause-modifier 
L_ object-modifier C object-descriptor object-restrictor 
Figure 4: Partial textual semantic hierarchy 
131 
7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 
We will explain the meaning of each textual category 
by describing the way in which they are realized from 
a grammatical viewpoint. Complex-clauses 7 are realized 
as sentences. The sub-hierarchy of clause represents dif- 
ferent linguistic forms of clause; in particular, the time- 
anchored-event is realized as a finite clause, the activity 
as a non finite clause and the object-perspective as clause 
nominalization. The clause-nucleus sub-hierarchy classi- 
fies different kinds of verbs (e.g., transitive and intransi- 
tive verbs) and verb voice. The other-argument is realized 
as verb argument that has no specific grammatical func- 
tion, whereas the three other clause-nucleus-arguments 
are realized as constituents whose grammatical function 
is subject, direct object or indirect object. For exam- 
ple, the verb "to get to" requires two arguments: one 
subject-argument (realized as the subject) and one other- 
argument which represents the destination. The textual- 
object is realized as a noun phrase, and, in particular, 
the unidentifiable-object as an indefinite noun phrase and 
the identifiable-object as a definite one. The hierarchy of 
textual-object-nucleus corresponds to nouns; it represents 
various linguistic forms of nouns with respect to their gen- 
der and number. The modifiers are realized as adjectives, 
adverbs, prepositional phrases, and subordinate clauses. 
2.2 Resource trees 
Resource trees are primitive trees from which textual 
structures are built. In particular, they indicate the way 
in which a constituent (i.e., a textual structure node) can 
be expanded, that is, what are its sub--constituents, or ex- 
gended, that is, what kind of additional sub-constituents 
can be added. They are defined from elementary trees. 
Meteer distinguished two basic forms of elementary trees: 
the kernel tree and the composite tree. These two trees 
represent Halliday's distinction between the multivariate 
and the univariate structure, respectively \[Halliday 85\]. 
The kernel tree (i.e., the HEAD-ARGUMENT structure) is 
an atomic structure, that means, it cannot be modified in- 
dependently from the whole structure. Essentially, it rep- 
resents a predicate/argument structure. The composite 
tree is an incremental structure, that is, new child nodes 
can be added at any time of the expansion of the mother 
node. Meteer distinguished the COMPOSITE-MATRIX- 
ADJUNCT structure, which corresponds to the hypotac- 
tic relation (i.e., the relation between dependent ele- 
ments and their dominating node), and the COMPOSITE- 
COORDINATE structure, which represents the paratactic 
relation (i.e., the relation between elements of equal sta- 
tus). Meteer defined resource trees as follows: 
(define-composite-tree complex-event-tree 
(matrix &rest adjunct) 
7The textual semantic categories are noted in sans serif text. 
:constructor-function build-complex-event-tree 
:matrix-restrictions (state event) 
:adjunct-restrictions (manner location)) 
This definition expresses that the complex-event re- 
source tree is a COMPOSITE tree whose MATRIX has 
to be of type state or event and ADJUNCTs of type 
raarmer or location. It also specifies a constructor func- 
tion which builds the tree. 
We think that the language for defining resource trees 
is not powerful enough. For instance, it must be explic- 
itly specified that a composite tree only has one matrix 
node or that an activ±~y does not have an explicit sub- 
ject. In order to solve this problem, a more powerful lan- 
guage had to be defined. In particular, the restriction on 
daughter nodes is more precise. In addition, we introduce 
the notions of relevance and constraints. Restrictions are 
expressed by using a specific language which expresses 
quantities (e.g., "one" or "several") and description of 
the daughter constituents. In particular, each daughter 
constituent is described by a structure, which we call ab- 
stract constituent, which specifies the ideational and 
textual semantic categories of the constituent, its struc- 
tural relationship with the mother node, and its rele- 
vance. The relevance is the link between textual struc- 
ture nodes and parts of the message; it indicates the part 
of the message verbalized by the mother node and its child 
nodes. Constraints are conditions which must be satis- 
fied by the textual structure. For example, constraints 
can express that the subject-argument of an infinitive- 
activity will not explicitly appear in the final text. Figure 
5 shows the definition of the "infinitive-activity" resource 
tree s, which is a composite tree whose mother node has 
the textual category infinitive-activity. The tree has one 
and only one matrix node of the ideational category pro- 
cess and of the textual category clause-nucleus, and ad- 
ditional adjunct nodes of the ideational category circum- 
stance and of the textual category clause-modifier. It spec- 
ifies a constraint which states that the subject-argument 
must be implicit. It also indicates that the mother node 
and its daughter nodes correspond to the same part of 
the message, that is, they have the same relevance (cf. 
the variable X in figure 5) 9. 
2.3 Lexical resources 
Lexical items, i.e., words, are lexical resources of the lan- 
guage. We represent their meaning by associating each 
of them with entities of the Upper Model and of the tex- 
8Meteer's resource trees axe expressed in terms of semantic cat- 
egories whereas our resource trees axe defined in terms of textual 
categories. 
9 The variable of the field "relevance" is unified with the relevance 
of a textual structure's node during the expansion of the node with 
the resource tree. 
132 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
(define-resource-tree 
:id rt-10 
:tel-to-daughter composite 
:mother-textual-cat infinitive-activity 
:relevance X (\[= 
clause-nucleus :restriction 
one process 
x 
:constraints 
\]) (\[ \]) several clause-modifier ' circumstance 
X 
((implicit-constituent subject-argument X))) 
Figure 5: The "infinitive activity" resource tree 
tual semantic hierarchy. For example, the lexical resource 
which corresponds to the direct transitive verb "to write" 
is defined as follows: 
\["to-write" \] 
(direct-transltive-nucleus) 
(write-action) 
A lexical resource can be linked to more than one textual 
category; for example, the following resources 
"article" \] 
(slngular-neuter-nucleus plural-neuter-nucleus) (a icle) 
"quick" \] 
(object-modifier clause-modifier) 
(quick) 
correspond to the nouns "article" and "articles" and to 
the adjective "quick" and the adverb "quickly", respec- 
tively. A lexical resource can also be associated with 
several ideational categories; for example, the following 
resource \[ \] 
(clause-modifier) 
(purpose enablemen~} 
can be used to verbalize a rhetorical relation "purpose" 
or "enablement". 
2.4 Textual structure 
The textual structure is a tree-like structure built from 
resource trees. It represents the constituency of the 
utterance (each node is a complete constituent of the 
utterance), the structural relations among constituents 
(head-argument, composite-matrix-adjunct, composite- 
coordinate), the ideational and textual categories, the rel- 
evance and the restriction of each node, and the lexemes. 
In addition, a marker which describes the lexical cohe- 
sion of the text (see \[Halliday & Hasan 76\]) is associated 
to each lexeme. The textual structure provides linguistic 
information which is required for treating specific linguis- 
tic phenomena such as anaphora planning (see section 1). 
For example, in figure 6 (only the main information con- 
tained in each node is depicted), the textual structure 
indicates that the domain instance "Stock" corresponds 
to a male and will be verbalized as a singular masculine 
proper name (named-object + singular-masculine-name). 
Several textual structures can be built to verbalize one 
message. The structure which is selected is the one which 
has the higher cohesive score. This score is obtained by 
summing the scores of lexical cohesion of all lexical re- 
sources used in the textual structure. The score of each 
resource is obtained as follows: 
• if the lexical resource appears in the local context (eft 
history of dialogue acts) then the lexical cohesion is 
scored 3; 
• if the lexical resource appears in the active set of the 
history of linguistic resources then the score is 2; 
• if the lexical resource appears in the open set of the 
history of linguistic resources then the score is 1; 
* otherwise, the score is 0. 
3 Utterance generation in the 
context of dialogue 
A plan of dialogue acts is verbalized by transforming it 
into two successive representational levels: surface lin- 
guistic acts and textual representations. These two main 
transformations are achieved with respect to the current 
dialogue context. We represent the context through four 
histories. The history of dialogue acts stores, in par- 
ticular, the requests of the interlocutor that have not been 
133 
7th International Generation Workshop • Kennebunkport, Maine * June 21-24, 1994 
message = (rl /sequence 
(el /write-action 
:actor stock 
:actee (ol/article 
:subject-matter alfresco)) 
(e2/write-action 
:actor pianesi 
:actee (o2/article 
:subject-matter maia))) 
I CO(RR.DINATE \[ wRn~-ACrK~N 
DEO..ARATIVE~LEX~.,AUSE 
I COM~osrr~ I 
WRn.~.ACrION 
'r~A/~CH~D-CL~LSE 
! co~osrr~ \[ 
WRII~-ACIII~I ~.~,~,~,~.~ 
AI~UMI~t'r 
ACTOR 
$L, BL.~r-ARGUI~ El 
I 
I 
SINGULAR-M.A~CULINE-NAME SINGULAR-NEUTER-NUCLEUS s~o~ I 
Ol Olivi~ro $tod¢ ~ FIRST-DIRECT ankle / FIRST-DIRECT 
COGRDI~IAT-. 
wRrrE-ACTION 
DE~.ARATIVE-C~EX-(X.AUSE 
COMPOSrI~ I 
I ~'~ \[ WR.rfF.-ACTION 
TIME-ANC~ORED-O.AUSE Su'uctaral relation with moth,r-....._ \[ m \] 
Ideational catcgo~.~~ COMPOSrm I 
ARGUM~T 
ACTIVE 
DIP.ECr.OBJECT-ARGUM~T 
El 
HEAD t 
INDENNnT~I~2T 
Ol 
~Ml~rrE ........----- ----...... 
ADIUNCT 
$UBTECT~,'Ud~ 
DI~C~Jt'roR-2 
Ol 
IboQU FIRST-DIRECT 
H-AD 
ARGUMENT 
NON.~NS~OUS-THING 
~!~rrE 
MATRIX 
N ON-CONSaOUS-'rH~G 
51NGULAR .NELrI'IER.NAME 
ALFRESCO 
Affr~co I FIRST-DIP.ECr 
ARGUM~rr 
ACTOR 
SUB F~CT-ARGUM\]~rr 
E2 
HEAD I 
I NAMED-CI~L.~2T 
I 
MA'I~.IX I I ~'DI.IX 
SINGULAR-MASCULINE-NAM... SINGULAR-NEUTER-NU(Z~JS 
PI~ESI I 02 
Pebio PiRncsi ( F'~.ST-D~R.ECT m't ick ~ I,uc,i'T~ rl i ~ON 
I ARGU~'r 
DIRECT-O GUM\]~rr 
I I ARGU~T 
INDiCia.rEfit 
COMi~ITE 
ADJUNCT 
SUET-MATTER 
DE~CPJFI'OR-2 O2 
al~t IREPETITiON 
I ARGUMI~rr 
NO N-CCRNSCIOUS-'r ~-IING 
NAMED.G~nECT 
COMPOSlrrE I 
NON-CONSCIOUS-THING 
$1NGULAR.N~.rrER.NAME 
MAIA Malu 
/ FIRST-DIRECT 
"Oliviero Stock wrote an article about Alfresco. Fabio Pianesi made one about MAIA." 
Figure 6: An example of textual structure 
134 
7th International Generation Workshop • Kennebunkport, Maine ° June 21-24, 1994 
answered yet. Each item of this list provides information 
on the dialogue act itself, the way in which conceptual 
entities that appear in its propositional content were re- 
alized with respect to the lexical semantics and the list 
of linguistic resources used. This information constitutes 
a local context during the verbalization of the answer. 
The history of conceptual entities contains informa- 
tion on the conceptual entities (e.g., "arrival-airport") al- 
ready verbalized in the dialogue and the way in which 
they were realized with respect to lexical semantics. This 
history is used for the verbalization of all surface linguis- 
tic acts, except for "answers" (the semantic realization 
of conceptual entities that appear in an "answer" is con- 
strained by the local context retrieved from the history 
of dialogue acts). The history of objects involves in- 
formation on the instances of conceptual entities (e.g., 
"JFK") mentioned in the dialogue. It is used to retrieve 
objects that are required from the lexical semantics point 
of view during the surface linguistic act planning (cf. the 
ideational semantic restrictions on arguments of entities 
of the Upper Model). It also provides information that is 
useful during anaphora planning, such as the information 
status (given or new) of objects. Finally, the history of 
linguistic-resources indicates which abstract linguistic 
resources have been used in the dialogue. It is divided into 
two sets: an active set and an open set. The active set 
contains the linguistic resources used in the last sentence 
of the dialogue, and the open set the resources used in 
the preceding sentences. In particular, this history is used 
to evaluate the cohesion of textual structures. 
We will take an example to show how the generation 
system deals with these histories in order to verbalize a 
plan of dialogue acts. We will assume that the interlocu- 
tor (i.e., the user) uttered the following question during 
the dialogue: 
The 10:00 flight goes to which airport? 
The dialogue act planner builds a dialogue act of inform- 
ing that the airport is "JFK "1° (this act is shown in fig- 
ure 1). Since there is a request in the history of dia- 
logue acts that corresponds to the dialogue act, the sur- 
face linguistic act planner decides to build an "answer" 
surface linguistic act. Then, it uses the local context pro- 
vided by this history in order to transform the semantico- 
conceptual content into a semantic content. In particular, 
it builds a propositional content expressing a destination- 
event whose theme is "the 10:00 flight" and whose desti- 
nation is "JFK" ~cf. content 1 in figure 1). Finally, the 
constructed linguistic act is passed to the micro-planning 
component. This component builds at least two textual 
structures representing the cases where the des~ination- 
even~ is verbalized by "to go" or "to arrive". Since "to go" 
1°The way in which dialogue acts and plans of dialogue acts are built is explained in \[Sadek 91\]. 
appears in the local context, the selected textual struc- 
ture is the one corresponding to the following utterance: 
The 10:00 .flight goes to JFK. 
According to the content of the history of objects, the 
micro-planner can introduce anaphoric expressions. In 
that case, the built textual structure corresponds to: 
It goes to JFK. 
Finally, using the history of linguistic resources, the plan- 
ner can produce ellipsis, in which case, the textual struc- 
ture built corresponds to: 
To JFK. 
4 Concluding remarks 
In this paper we described a component of a natural lan- 
guage generation, called the micro-planner, which is situ- 
ated between the message planner and the linguistic real- 
ization component. Its role is to bridge the gap between 
the two traditional components in order to avoid some 
of the problems which arise in two-component genera- 
tion systems. In order to achieve this, the micro-planner 
transforms the message built by the message planner into 
an intermediate structure, which we call textual struc- 
ture. This structure represents an abstract grammatical 
organization of the final text. In particular, it provides 
semantic and syntactic information, which is useful to 
handle linguistic phenomena such as anaphora planning. 
This new representational level is based on the notion of 
abstract linguistic resources \[Meteer 92\]. We made two 
main modifications. Firstly, Meteer's semantic category 
hierarchy was substituted by two distinct semantic hier- 
archies: an ideational semantic hierarchy (i.e., the Upper 
Model \[Bateman et al. 90\]) and a textual semantic hier- 
archy. Secondly, we increased the expressibility of the 
resource tree definition language. Moreover, we defined 
a simple mechanism to represent lexical resources of the 
language. 
In a three-component generation system, the micro- 
planner makes the linguistic decisions (e.g., the selection 
of syntactic structures and the choice of lexical items) 
that are usually taken by the linguistic realizer in a 
two-component generation system. The advantages are, 
firstly, taking a decision is simpler since the micro-planner 
de~ls with abstract linguistic knowledge, secondly, the 
micro-planner is closer to the message planner than the 
linguistic realizer, and, thirdly~ the linguistic realization 
component is simplified since it no longer has to take lin- 
guistic decisions. Such a component appears in the gen- 
eration system of \[Danlos 87\]: the strategic component 
takes conceptual and linguistic decisions, and the tacti- 
cal component only applies grammar rules. In fact, the 
135 
7th International Generation Workshop • Kennebunkport, Maine ° June 21-24, 1994 
linguistic realization component only deals with syntactic 
details such as agreement and complete surface order and 
morphology. 
Our micro-planning component is currently being in- 
tegrated in a text generation system developed at IRST 
(see \[Maier & Not 93, Panaget 94\]). A previous version 
(\[Panaget 92\]) was implemented in order to be integrated 
in a human-computer oral dialogue systefn developed at 
CNET \[Sadek et al. 94\]. For the former case, the linguis- 
tic realization component \[Pianesi 91\] is based on the gov- 
ernment and binding theory and a head-driven bottom- 
up generation algorithm. The component consists of an 
integrated system for editing, modifying and maintain- 
ing unification-based grammars. For the latter case, the 
linguistic realizer is based on Tree Adjoining Grammar 
\[Joshi 87\]. In particular, the textual structure is trans- 
formed into a derivation tree (i.e., a tree which specifies 
a derived tree through a bottom-up traversal), which is 
then 'executed' by the linguistic component in order to 
build a derived tree (i.e., a syntactic tree). 
Acknowledgements 
I am extremely grateful to John Bateman, Elisabeth 
Maier, Elena Not, David Sadek and Oliviero Stock for 
their helpful comments on early versions of this paper. 

References 
\[Appelt 85\] Appelt D.E., Planning english sentences; 
Cambridge University Press, 1985. 
\[Bateman et al. 90\] Bateman J.A., Kasper R.T., Moore 
J.D. & Whitney R.A., A general organization of ~he 
knowledge for natural language processing: the 1lehman 
Upper Model, ISI research report, California, 1990. 
\[Bescherelle 90\] Bescherelle 3 - La grammaire pour tous, 
Hatier, 1990. 
\[Dale 88\] Dale R., Generating referring expressions in a 
domain of objects and processes, PhD, University of 
Edinburgh, 1988. 
\[Danlos 87\] Danlos L., The linguistics basis of text gen- 
eration, Cambridge University Press, 1987. 
\[Halliday 85\] Halliday M.A.K., Introduction to functional 
grammar, Edward Arnold, 1985. 
\[Halliday & Hasan 76\] Halliday M.A.K. & Hasan R., Co- 
hesion in english, Longman inc., New York, 1976. 
\[Hovy 90\] Hovy E., Pragmatics and natural language 
generation, Artificial Intelligence 43(2), pp 153-197, 
1990. 
\[Hovy 92\] Hovy E., Sentence planning requirements for 
automated explanation generation, DIAMOND Bericht 
n ° 23, 1992. 
\[Joshi 87\] Joshi A.K., The relevance of Tree Adjoining 
Grammar to generation, In Kempen G. ed., Natu- 
ral language generation, Dordrecht, Martinus Nijhoff, 
1987. 
\[Lever 89\] Levelt W.J.M., Speaking -from intention to 
articulation, The MIT Press, 1989. 
\[McKeown 85\] Me Keown K.R., Discourse strategies for 
generating natural language text, A rtificial Intelligence 
~7(i), pp 1-41, 1985. 
\[Maier & Not 93\] Maier E. & Not E., Increasing cohesion 
in automatically generated natural language texts, In 
Torasso P. ed., Advances in AI- Third congress of the 
Italian Association for AI, Lectures notes in AI 728, 
Springer-Vexing, 1993. 
\[Matthiessen & Bateman 91\] Matthiessen C. & Bateman 
J.A., Text generation and systemic-functional linguis- 
tics - Experiences from English and Japanese, Pinter 
Publishers, London, 1991. 
\[Meteer 92\] Meteer M., Expressibility and the problem of 
efficient text planning, Pinter publishers, 1992. 
\[Panaget 92\] Panaget F., De l'acte de dialogue au tan- 
gage natural : un ggndrateur d'dnoncds en con- 
texte de dialogue oral, CNET Technical Report 
NT/LAA/TSS/440, Lannion, France, 1992. 
\[Panaget 94\] Panaget F., The micro-planning component 
of a generation system, IRST Technical Report 9402- 
05, Trento, Italy, 1994. 
\[Pianesi 91\] Pianesi F., Head driven bottom up genera- 
tion and government and binding: a unified perspec- 
tive, In proceedings of the 2 "a European Workshop on 
Natural Language Generation, Austria, 1991. 
\[Sadek 91\] Sadek M.D., Dialogue acts are rational plans, 
In proceedings of the ESCA/ETRW workshop on "the 
structure of multimodal dialogue", Italy, September 16- 
20, 1991. 
\[Sadek et al. 94\] Sadek M.D., Ferrieux A & Cozannet A., 
Towards an artificial agent as the kernel of a spoken di- 
alogue system: a progress report, To appear in the pro- 
ceedings of the AAA1 workshop on integration of natu- 
ral language and speech processing, 1994. 
\[Shinghal 92\] Shinghal R., Natural language processing: 
a prescriptive grammar, In Formal conceptions in Arti- 
ficial Intelligence, pp 131-232, Chapman & Hall, 1992. 
