P O P E L - H O W : A Distributed Parallel Model for Incremental Natural Language Production w i t h Feedback Giinter Neumann Wolfgang Finkler Lehrstuhl fur Computerlinguistik Deutsches Forschungszentrum f u r U n i v e r s i t a t des Saarlandes Kiinstliche Intelligenz G m b H I m S t a d t w a l d 15, B a u 1 7 Stuhlsatzenhausweg 3 6600 Saarbriicken 1 1 , F R G 6600 Saarbriicken 1 1 , F R G
[email protected] saarland.dbp.de
Abstract
ceding level.
T h i s allows the c o n t i n u a t i o n o f t h e p r o d u c -
t i o n , set i n m o t i o n b y the h o w - t o - s a y c o m p o n e n t .
1
This paper presents a new model for the production of natural language. The novel idea is to combine incremental and bidirectional generation w i t h parallelism. The operational basis of our model is a distributed parallel system at every level of representation. Starting point of the production are segments of the conceptual level. These segments are related to active objects which have to map themselves across the linguistic levels (i.e. functionalsemantic, syntactic and morphologic level) as fast and as independently as possible. Therefore the incremental behavior caused by a successive input is propagated on all levels. Linguistic requirements detected by objects which are related to already produced segments at any level influence and restrict the decision of what to say next.
As a
consequence, l i n g u i s t i c r e s t r i c t i o n s detected d u r i n g t h e p r o d u c t i o n p r o v i d e feedback for the p l a n n i n g of w h a t to say n e x t [ H o v y 87], [ R e i t h i n g e r 88]. Because t h e single p a r t i a l s t r u c t u r e s shall be verbalized i n d e p e n d e n t l y , we decided to consider t h e m as active a n d a u t o n o m o u s o b j e c t s w h i c h have t o verbalize themselves.
T h e n the whole utterance is pro-
duced in a d i s t r i b u t e d system of such o b j e c t s . T h e r e f o r e o u r m o d e l can be characterized as a c o m b i n a t i o n of m e t h o d s of
Distributed
Artificial
Intelligence
and
natural
language
generation. U p t i l l n o w systems w i t h i n t e r a c t i o n between t h e w h a t to-say and how-to-say component don't treat incremental f o r m u l a t i o n e x p l i c i t l y (e.g. [ H o v y 87], [ A p p e l t 85]) a n d i n c r e m e n t a l g e n e r a t i o n systems h a v e n ' t considered a bidirect i o n a l f l o w o f c o n t r o l (e.g. [DeSmedt and K e m p e n 87]). W e
1
present P O P E L - H O W , a h o w - t o - s a y c o m p o n e n t w h i c h for
Introduction
t h e first t i m e realizes a n i n c r e m e n t a l a n d b i d i r e c t i o n a l p r o -
In an incremental generation system t h e c o m p o n e n t deter-
d u c t i o n w h i c h is based on a d i s t r i b u t e d p a r a l l e l m o d e l at
m i n i n g t h e contents of t h e desired utterances (what-to-say
several l i n g u i s t i c levels, i.e.
c o m p o n e n t ) can s u b m i t p r e m a t u r e l y c o m p u t e d p a r t i a l re-
t h e utterances.
sults to t h e how-to-say c o m p o n e n t in order to a c t i v a t e t h e p r o d u c t i o n process before the p l a n n i n g is c o m p l e t e .
Such
systems s i m u l a t e a n i m p o r t a n t p r o p e r t y o f language prod u c t i o n processes i n h u m a n speakers:
h u m a n s often s t a r t
s p e a k i n g before t h e y k n o w e x a c t l y w h a t the w h o l e contents of t h e u t t e r a n c e w i l l be [Schriefers and P e c h m a n n 88]. T h e h o w - t o - s a y component must t r y to map its input structures across a l l representation levels as fast as possible to m a i n t a i n t h e i n c r e m e n t a l behavior i n v o l v e d by t h e successive i n p u t . T h i s results i n speeding u p t h e generator's o u t p u t . A q u i c k v e r b a l i z a t i o n of p a r t s of t h e contents is necessary if a user in m a n - m a c h i n e c o m m u n i c a t i o n m u s t i m m e d i a t e l y react on t h i s i n f o r m a t i o n (e.g. in systems for air traffic cont r o l ) . In t h e case t h a t t h e m a p p i n g of a p a r t i a l s t r u c t u r e is n o t possible because of unspecified b u t r e q u i r e d elements, i t i s necessary t o request missing i n f o r m a t i o n f r o m t h e pre-
The which
component is
the
is
f r o m c o n c e p t u a l s t r u c t u r e s to
part
generator
of
of the
POPEL XTRA
[Reithinger 88 dialog
system
[Allgayer et a l . 89]. T h e i n t e g r a t i o n in a d i a l o g system suggests t h e c o m m o n usage of t h e k n o w l e d g e sources of t h e ent i r e s y s t e m . W i t h the e x c e p t i o n o f t h e g r a m m a r , t h e same k n o w l e d g e sources c o u l d be used for b o t h analysis a n d gene r a t i o n . Because o f t h e i n c r e m e n t a l a n d p a r a l l e l processing in P O P E L - H O W , a declarative grammar t r e a t i n g t h e specific r e q u i r e m e n t s h a d to be developed.
2
Overview of the A r c h i t e c t u r e of P O P E L - H O W
In P O P E L - H O W there are f o u r processing levels (see figure 1) a c c o r d i n g to the different k n o w l e d g e sources w h i c h are used d u r i n g t h e v e r b a l i z a t i o n .
T h e k n o w l e d g e sources
for t h e first t w o levels - t h e C o n c e p t u a l K n o w l e d g e Base 1
The work presented here is being supported by the Gernan Science Foundation (DFG) in its Special Collaborative Program on AI and Knowledge-Based Systems (SFB 314), project Nl (XTRA). Thanks to Barbara Fairlight and Norbert Reithinger for their helpful comments on previous versions of this paper.
1518
Speech and Natural Language
( C K B ) a n d t h e F u n c t i o n a l Semantic S t r u c t u r e (FSS)
are
defined using S B - O N E , a r e p r e s e n t a t i o n language s i m i l a r t o K L - O N E [ K o b s a 89]. T h e n e x t t w o levels are r e l a t e d t o s y n t a c t i c s t r u c t u r e s . T h e c e n t r a l k n o w l e d g e source for
these levels is P O P E L - G R A M , a unification grammar based on P A T R - I I [Shieber et al. 83]. T h e rules are divided i n t o a n I D - and L P - p a r t i n order t o separate the construct i o n of syntactic structures at the D B S (Dependency Based Structures) level and the linearization of these structures at the I L S (Inflected Linearized Structures) level 2 . T h e inflect i o n of the stems is performed at t h i s level by means of the package M O R P H I X [Finkler and N e u m a n n 88].
next level by creating a new process which represents the goal structure of the mapping. If the m a p p i n g fails, the process can communicate w i t h processes of the level above in order to request the missing i n f o r m a t i o n . It is possible t h a t a process which receives such a request cannot handle it at its level. In order to fullfil the request such processes themselves send requests. In this way P O P E L - W H A T , the w h a t - t o - s a y component of P O P E L , can get feedback f r o m P O P E L - H O W which influences the planning of w h a t to say next. T h e d i s t r i b u t e d processing permits the incremental specification of the i n p u t for P O P E L - H O W . Parts of the i n p u t can be mapped i n t o the next level before new i n p u t structures arrive. Because the result of the mapping is also a process which tries to map its structure i n t o the next level, the stepwise i n p u t of segments is propagated over all levels. I n p u t which arrives later has to be integrated into the existing structures in order to allow an incremental generation at each level of description.
3
The Underlying Parallel M o d e l
O u r model follows the design methodology of objectoriented concurrent programming [Yonezawa and Tokoro 87]. Characteristic for such models is t h a t there exist a collect i o n of concurrently w o r k i n g self-contained objects which must interact in order to solve a problem cooperatively. An essential design issue is the question of how to decompose the overall problem i n t o components which can be represented and handled by the objects. W h e n an object in P O P E L - H O W has solved its part of the problem at a given moment, it does not yet terminate. T h e reasons are: first, it can send information to other objects in order to help t h e m and second, because of the incremental processing in P O P E L - H O W , new information can alter the object's task 3 . The relevant steps of the life cycle of an object are:
Figure 1 : T h e A r c h i t e c t u r e o f P O P E L - H O W T h e first three processing levels have been conceived w i t h the same u n d e r l y i n g distributed parallel model The p r o d u c t i o n of structures at these levels is performed by a set of cooperating processes. Therefore the knowledge source of each level is decomposed i n t o segments and related to the processes. A l l processes together represent the s t r u c t u r e of the utterance constructed thus far. T h e main task of each process is the m a p p i n g of its segment i n t o the 2
T h e i n t e r p r e t a t i o n of the P A T R - I I f o r m a l i s m doesn't allow the defi n i t i o n of grammars which express b o t h relations explicitly. Therefore the i n t e r p r e t a t i o n of rules has been m o d i f i e d ; the context free p a r t of rules at the D B S level is interpreted as the dominance p a r t , the LP rules constrain the linear order of phrases.
1. become acquainted w i t h newer objects, 2. eventually establish communication links to these objects, 3. t r y to m a p the represented component into the next level, 4. send or answer requests, 5. start w i t h the first step. Every new object enters its address into an attendance list which is managed by a particular object. Objects already present in the system can then become acquainted w i t h newer ones. An object tests whether a neighborhood relation to a found new object can be set up by a directed communication link. This decision must be performed synchronously by the objects and is made w i t h regard to their 3
The t e r m i n a t i o n of all objects happens only when no more i n p u t for P O P E L H O W is specified. Of course, this criterion is determined only by P O P E L - W H A T , which sends a t e r m i n a t i o n message in this case.
Finkler and Neumann
1519
u n d e r l y i n g represented components. T h e set of the comm u n i c a t i o n links forms the context of an object. D u r i n g the m a p p i n g which is performed by local rules, an object compares the context w i t h the c o n d i t i o n part of a rule. A c c o r d i n g to the c o m p l e x i t y of t h e rule, this step requires c o m m u n i c a t i o n w i t h linked objects. T h e result of a successful m a p p i n g is the creation of a new object at the next level. T h e new object is provided w i t h patterns for the c o m m u n i c a t i o n links. These patterns describe the relation between the objects of the level above and the represented components of their m a p p i n g results (see figure 2). T h e
a modified P A T R - I I - f o r m a l i s m . I D - r u l e s are denned as follows: There exists a p a r t i c u l a r constituent - the so-called head element - w h i c h determines the phrase by i n d i c a t i n g the essential g r a m m a t i c a l features and w h i c h restricts the number and q u a l i t y i.e. syntactic f u n c t i o n of the remaining constituents according to its valence. Each I D - r u l e is regarded as an atomic syntactic u n i t and can be assigned to objects in the parallel m o d e l . T h i s is done by associating a phrase or its head-element w i t h an object and the dependent constituents w i t h the communicat i o n links. Because of the explicit description of g r a m m a t ical relations between constituents, the object representing the phrase has to m a i n t a i n the feature specification as well. So the whole dag 4 of a phrase is assigned to the object. T h e omission of the linear order at this level and the dependency relation defined between the constituents of a phrase allow an o b j e c t - o r i e n t e d view of syntactic structures and hence s u p p o r t the parallel and incremental p r o d u c t i o n .
5
The D i s t r i b u t e d M a p p i n g between the Levels
Figure 2: Objects on two Successive Levels bidirectional l i n k connecting father and son objects and the patterns of the son objects support the h a n d l i n g of requests between objects of successive levels. T h e answering of requests takes place asynchronously so t h a t the object w h i c h receives requests is not i n t e r r u p t e d needlessly often. Because the presented p r o g r a m is the same for all objects, they have equal weights. Therefore the desired p r o cess model should be fair.
The Decomposition and Distribution of t h e K n o w l e d g e Sources In order to use the declaratively f o r m u l a t e d knowledge sources of each level efficiently in the d i s t r i b u t e d parallel model described above, it is necessary to decompose t h e m i n t o atomic components. T h e definition of the atomic u n i t s depends on the u n d e r l y i n g f o r m a l i s m of the knowledge source. Basic u n i t s at the epistemological level of S B - O N E are concepts and their a t t r i b u t e descriptions for the general and i n d i v i d u a l level of representation [Kobsa 89]. Therefore we defined a concept plus its a t t r i b u t e descriptions as atomic component for the S B - O N E - b a s e d levels C K B and FSS. W h e n a component is related to objects, the roles of the a t t r i b u t e descriptions correspond to the communicat i o n links of the object and the value restrictions are used as patterns to reduce the number of possible candidates for communication. T h e I D - r u l e s o f P O P E L G R A M c o n s t i t u t e the k n o w l edge
source
linguistic
o f the
basis
of
DBS-level the
in
grammar
POPEL HOW. is
dependency
The the-
ory, a t r a d i t i o n a l approach of continental linguists (c.f. [Tesniere 59],[Engel 82]), which is expressed and applied in
1520
Speech and Natural Language
T h e different representation levels must be related to each other in order to allow a m a p p i n g between conceptual u n i t s and the utterance. In our model the m a p p i n g has to be performed in a d i s t r i b u t e d way so t h a t the components of the knowledge sources can be verbalized independently. For the knowledge sources based on S B - O N E , we define a set of local t r a n s i t i o n rules for every general concept. T h e precond i t i o n p a r t of simple rules specifies a t t r i b u t e descriptions for the concept. T h e action parts of the m a p p i n g rules f o r m u l a t e relations between the detected elements of the left hand side and the structures to be created at the next level. In the case where a compositional and independent m a p p i n g of several structures is not possible for example if some conceptual knowledge units are represented in greater detail t h a n the corresponding linguistics! knowledge - complex rules can be defined allowing the simultaneous m a p p i n g of structures consisting of several concepts which are connected by roles i n t o structures at the next level. S t a r t i n g p o i n t for the d i s t r i b u t e d m a p p i n g of structures at the D B S - l e v e l are the dags of the objects. An object has to test whether its dag fulfills a c o n d i t i o n of local completeness. T h i s is t r u e if the subdag of the head element contains no unspecified features. In this case the dag is sent to the I L S - l e v e l , where the last step of the m a p p i n g process takes place: the linearization, inflection and incremental o u t p u t of the surface strings. D u r i n g the m a p p i n g the choice of words is performed in t w o steps. In the first step the content words (e.g. nouns) are selected in the course of the t r a n s i t i o n f r o m C K B - t o FSS-structures. T h e precondition o f a n applica4
Dags (directed acyclic graphs) are the basic means of representation
in PATR-II.
ble rule serves as a d i s c r i m i n a t i o n net to select the appropriate w o r d . T h e choice of descriptors - the d e t e r m i n a t i o n whether a n o m i n a l concept should be realized as a definite or indefinite n o m i n a l phrase, as an anaphora or should be suppressed in the case of elliptical verbalization - is done by a c o m m u n i c a t i o n w i t h P O P E L - W H A T , which knows about previously mentioned individualized concepts. In the second step there is an a d d i t i o n a l choice of function words, for instance, prepositions. It is performed d u r i n g the m a p p i n g i n t o the D B S - l e v e l .
P O P E L - W H A T decides which structures of the i n d i v i d ual C K B have to be included or excluded in a dialog cont r i b u t i o n and the order in which they shall be uttered. T h e order is computed according to focus values w h i c h are ext r a c t e d f r o m the linguistic dialog m e m o r y [Reithinger 88]. T h e sequence of selected structures is passed to P O P E L H O W as i n p u t step by step. In our example the first concept to be verbalized is object-transfer-2. At this moment this is the only known concept i n the h o w - t o - s a y component. P O P E L H O W creates an object c k b - O 1 which represents this concept. T h e patterns of the c o m m u n i c a t i o n links of this object (see nextpicture) refer to objects which w i l l represent the indicated i n d i v i d u a l concepts if these structures are passed as i n p u t later on. Because c k b - O i is as yet the sole object, it tries to map itself i n t o the next level i m m e d i a t e l y 5 . Therefore it checks the a p p l i c a b i l i t y of its rules.
Because none of the rules is applicable, b u t on the other hand a m a p p i n g must be performed as quickly as possible to m a i n t a i n the incremental behavior, the rule which specifies the fewest preconditions (rule 2) is chosen to request the missing i n f o r m a t i o n f r o m P O P E L - W H A T . I n our scenario the request is satisfied. T h e objects c k b - 0 2 representing f o r m - 4 and ckb-03 for person-1 are created and c o m m u n i cation links to t h e m are established by c k b - O 1 . N o w c k b - O i successfully applies rule 2. T h e result is the new object fss-O 1 which represents only a p o t e n t i a l i n d i v i d u a l concept 6 at the FSS-level.
In the meantime another object fss-0 2 was created and linked w i t h fss-Oi. T h e only rule for the m a p p i n g of a predicate in the scheme above requires one more role t h a n is described in the context of fss-Oi. As a consequence, the father of fss-Oi is asked for the missing communication l i n k . c k b - O i searches for a rule w h i c h enables the creation of an object for predicate w i t h a sufficient context. T h e application of the appropriate rule causes a request for an object representing person-2. I f P O P E L - W H A T d i d n ' t want to select this concept up t i l l now, then the satisfaction of this request means t h a t linguistic restrictions detected by P O P E L - H O W affect the planning 7 . A f t e r the creation of the requested object c k b - 0 4 and after its connection w i t h c k b - O i , fss-O 1 is the result of the application of rule 1 by ckb-O 1 . T h i s new object is in comp e t i t i o n w i t h fss-O 1 because of the same father. T h e next figure illustrates the net of the connected objects at the FSS-level after the objects c k b - 0 2 and c k b - 0 4 successfully t r a n s m i t t e d themselves to this level.
Finkler and Neumann
1521
lexical i n f o r m a t i o n of the head element "geben" ( t o give). T h i s means t h a t the selection mechanism is lexically driven. W h e n objects at the D B S - l e v e l install c o m m u n i c a t i o n links t h e y f o r m a dependency-based connection of I D - s t r u c t u r e s . T h e explicit feature description of this connection has to be constructed by means of unification of the dags associated w i t h the objects. T h i s causes feature exchange and determ i n a t i o n of the syntactic f u n c t i o n of the dependent object. Assume t h a t in the meantime f s s - 0 3 has created object d b s - 0 2 representing an np w i t h an unspecified feature 'case'. D u r i n g the s e t - u p of the c o m m u n i c a t i o n l i n k el between dbs-O 1 and d b s - 0 2 the syntactic f u n c t i o n of the n o m i n a l phrase represented by d b s - 0 2 is determined as " n p in the accusative". Because all relevant features of the dag of d b s - 0 2 have values, the dag is now local complete and m a p p e d to the I L S - l e v e l . T h e inflection and linearization of the lexemes in t h i s dag produces the first part of the utterance: d a s f o r m u l a r (the f o r m ) . T h e object dbs-O 1 is not yet complete because an object representing the subject (the constituent named eO) is missing. A request has been made but is not propagated f u r t h e r because at this moment f s s - 0 2 creates its successor object d b s - 0 3 , w h i c h represents an np t h a t becomes the subject of the sentence. A f t e r the feature exchange, the dags of b o t h objects are complete. If d b s - 0 3 ' s dag arrives at the next level first, it cannot be uttered because no comm o n linear order w i t h the existing constituent can be given. So the new s t r i n g : "der steuerberater" (the t a x advisor) is retained. W h e n the dag of d b s - O i arrives, the lexeme for its head-element, the verb ''geben" ( t o give) is inflected. Because the head -element dominates the existing phrases subject and direct object (eO and e l ) , L P - r u l e s are applicable t h a t describe the linear order of the constituents so far. Therefore the next o u t p u t is: g i b t d e r S t e u e r b e r a t e r (gives the t a x advisor). F i n a l l y the n p d e m K o l l e g e n ( t o the colleague) is u t t e r e d if the object c k b - 0 3 succeeds in m a p p i n g itself across all representation levels.
7
The General Behavior of the N a t u r a l Language P r o d u c t i o n i n o u r M o d e l
T h e example in the previous section was presented in detail i n order t o demonstrate the p r o d u c t i o n i n P O P E L - H O W . T h e produced utterance is influenced by the order of the stepwise i n p u t . If the i n p u t of the example is given in a different sequence, then the w o r d order in the utterance can differ, t o o ; e.g. the following sentence can be produced: "Dem Kollegen gibt der Steuerberater das Formular." ("*To the colleague gives the tax advisor the form.") Changes in the word order of the utterance are not the only consequence of another i n p u t sequence. T h e choice of words can be influenced because the application of alternat i v e m a p p i n g rules results in the construction of different structures on the next levels. If the following i n p u t is passed
1522
Speech and Natural Language
to P O P E L - H O W in the indicated sequence: concepts den o t i n g the destination " D e t r o i t " , a " t r a v e l l i n g " a c t i o n , a traveller " P e t e r " and finally the vehicle " a i r p l a n e " , then the utterance: "Nach Detroit
fahrt
Peter"
( " * T o D e t r o i t goes Peter")
and its c o n t i n u a t i o n : "mit dem Flugzeug." ("by airplane.") is produced step by step. If the i n p u t sequence is: the vehicle, the person, the action and the destination, then the utterance w i l l be: "Peter flight nach Detroit." ("Peter flies to Detroit.") Because the i n s t r u m e n t of the action is k n o w n in due t i m e , a verb w h i c h is more special t h a n " f a h r e n " ( t o go) and which incorporates the vehicle is chosen. Therefore the p r o d u c t i o n of the prepositional phrase " m i t dem F l u g z e u g " (by airplane) is suppressed. In a d d i t i o n , it is possible t h a t a different utterance is produced f r o m the same i n p u t sequence. T h i s happens if i n f o r m a t i o n gained by c o m m u n i c a t i o n between objects i n P O P E L - H O W and the c o n t e x t - h a n d l e r o f P O P E L W H A T direct the application o f different m a p p i n g rules and the choice of descriptions. For instance it is determined whether a noun phrase or a p r o - w o r d w i l l be generated or whether a predicate shall be expressed as a verb phrase or a noun phrase. T h e course of the generation process in the example of the previous section can be characterized as a synthesis free of conflicts. New structures are integrated i n t o the already existing structure of the utterance. T h i s means t h a t the representation of the utterance at each level grows monotonicly. W e call this special behavior m o n o t o n i c i n c r e m e n t a l generation. U p w a r d and d o w n w a r d expansion 8 as characterized in [DeSmedt and K e m p e n 88] for the syntactic incrementation take place at each level in P O P E L HOW. T h e incremental augmentation of the structure related to an object must be processed by this object itself in order to t r a n s m i t its new quality i n t o the next level. If the augmented object already possesses a successor, then a mapp i n g creates an object which w i l l become a c o m p e t i t o r to the existing successor. If the c o m p e t i t o r represents a more detailed segment than the previously existing object, in the sense t h a t the c o m p e t i t o r is an augmentation in comparision w i t h the other object, the incremental behavior is still monotonic. It is possible t h a t objects in c o m p e t i t i o n w i t h each other d o n ' t represent more detailed segments but describe reformulations. T h e integration of these structures i n t o the already existing one changes the s t r u c t u r e in a non 8
Insertion cannot occur in our model because two connected objects represent immediate constituents. Of course, objects representing n o n immediate constituents can be p u t together by a combined upward and downward expansion.
monotonic manner. As a consequence, in a d d i t i o n to the already described operations, e.g., establishing communication links or u n i f i c a t i o n , non monotonic operations are defined at each level in order to be able to handle reformulations. For instance an operation for the replacement of subdags in a dag in question is defined for D B S - o b j e c t s . We denote this k i n d of generation as n o n — m o n o t o n i c i n crementation. M a n a g i n g the o u t p u t in such an incremental system is not a t r i v i a l task. T h e order in the utterance is det e r m i n e d by i n d i c a t i n g a desired order of specified concepts at the i n p u t level and is constrained by linearizat i o n rules at the I L S - l e v e l . However, in the parallel model the i n p u t sequence is generally lost. An immediate outp u t of the phrases which first arrive at the I L S -level causes some problems because it can happen t h a t new parts would have to be positioned in front of already uttered parts, or t h a t the new parts have to augment or replace previously uttered fragments. If a filter is used in order to control the o u t p u t of parts of the utterance then there is only a p a r t i a l l y incremental p r o d u c t i o n on the last step. However, if a phrase shall be uttered immediately then devices for 'self-corrections' are necessary as stated in DeSmedt and K e m p e n 87]. A t present i n P O P E L H O W the parts of the utterance in question are repaired by repeating the utterance of the corresponding constituents.
8
Conclusions and F u t u r e Research
T h i s paper has demonstrated how n a t u r a l language product i o n is performed incrementally based on a bidirectional and d i s t r i b u t e d parallel model.
We showed t h a t the i n -
cremental generation process across several representation levels can be influenced by linguistic restrictions.
Such a
model has not yet been investigated in research about natu r a l language generation systems. Therefore we focused on describing the parallel generation process which requires a decomposition of all used knowledge sources and their dist r i b u t i o n to active objects.
For this reason we could not
present the details regarding the linguistic background and the knowledge sources which would have demonstrated the linguistic ability. T h e model has been f u l l y implemented on a Symbolics 3640 lisp-machine under Genera. T h e simulation of the parallel machine was done by using the process facilities and the scheduler of the l i s p - m a c h i n e .
Because of the formu-
l a t i o n of a declarative generation g r a m m a r in P A T R - I I its a u g m e n t a t i o n can be performed easily. However the gramm a r w r i t e r has to pay a t t e n t i o n to the application of the g r a m m a r in a d i s t r i b u t e d system. O u r next research concerns the elaboration of the dist r i b u t e d w o r d choice, which we feel cannot be performed in such an independent way.
References [Allgayer et al. 89] J. A l l g a y e r , K. H a r b u s c h , A. K o b s a , C . R e d d i g , N . R e i t h i n g e r and D . S c h m a u k s . X T R A : A natural-language access system to expert systems. International Journal of ManMachine Studies, 1989. To appear. [Appelt 85] D. A p p e l t . Planning English Sentences. Cambridge: Cambridge University Press, 1985. [DeSmedt and Kempen 87] K . D e S m e d t and G . K e m pen. Incremental Sentence P r o d u c t i o n , SelfCorrection and C o o r d i n a t i o n . I n : G. K e m p e n (ed.), Natural Language Generation, pp. 365-376, Dordrecht: M a r t i n u s Nijhoff, 1987. [DeSmedt and Kempen 88] K . D e S m e d t and G . K e m pen. The Representation of Grammatical K n o w l edge in a Model for Incremental Sentence Generation. Paper presented at the 4 t h I W G , Santa Catalina Island, 7 1988. [Engel 82] U. E n g e l . wartssprache. 1982.
Syntax Berlin:
der deutschen Erich Schmidt
GegenVerlag,
[Finkler and xNeumann 88] W. F i n k l e r and G. N e u mann. M O R P H I X : A Fast Realization of a Classification-Based Approach to Morphology. I n : Proceedings of the ''Wiener Workshop Wissensbasierte Sprachverarbeitung'' , B e r l i n , 1988. [Hovy 87]
E. H o v y . Generating Natural Language Under Pragmatic Constraints. P h D thesis, Yale University, 1987.
[Kobsa 89] A. K o b s a . The S B - O N E Knowledge Represent a t i o n Workbench. I n : J. Sowa (ed.), Formal Aspects of Semantic Networks, Los A l t o s , C A : Morgan K a u f m a n n , 1989. To appear. [Reithinger 88] N. R e i t h i n g e r . POPEL: A Parallel and Incremental Natural Language Generation System. Paper presented at the 4th I W G , Santa C a t a l i n a Island, 7 1988. [Schriefers and Pechmann 88] H. S c h r i e f e r s and T. P e c h mann. Incremental p r o d u c t i o n o f referential noun-phrases by human speakers. I n : M. Zock and G. Sabah (eds.), Advances in Natural Language Generation, V o l . 1, pp. 172 179, L o n d o n : Pinter Publishers, 1988. [Shieber et al. 83] S.M. Shieber, H. Uszkoreit, F.C.N P e r e i r a , J . J . R o b i n s o n and M . T y s o n . T h e Formalism and I m p l e m e n t a t i o n o f P A T R - I I . I n : Research on Interactive Acquisition and Use of Knowledge, Chapter 4, pp. 39 79, Menlo P a r k , California: A r t i f i c i a l Intelligence Center, S R I I n t e r n a t i o n a l , 1983. [Tesniere 59] L. T e s n i e r e . Elements de Syntaxe Structural. Paris: Klincksieck, 1959. [Yonezawa and Tokoro 87] A. Y o n e z a w a und M. T o k o r o (eds.). Object-Oriented Concurrent Programming. Massachusetts I n s t i t u t e of Technology: T h e M I T Press, 1987.
Finkler and Neumann
1523
