Knowledge Structures for Natural Language Generatlon ~ 
Paul S. Jacobs 
Knowledge-Based Systems Branch 
General Electric Corporate Research and Development 
Schenectady, NY 12301 USA 
Abstract 
The development of natural language interfaces to Artifi- 
cial Intelligence systems is dependent on the representation of 
knowledge. A major impediment to building such systems has been the difficulty in adding sufficient linguistic and concep- 
tual knowledge to extend and adapt their capabilities. This 
difficulty has been apparent in systems which perform the 
task of language production, i. e. the generation of natural 
language output to satisfy the communicative requirements 
of a system. The Ace framework applies knowledge representation fun- 
damentals to the task of encoding knowledge about language. 
Within this framework, linguistic and conceptual knowledge 
are organized into hierarchies, and structured associations are 
used to join knowledge structures that are metaphorically 
or referentially related. These structured associations per- 
mit specialized linguistic knowledge to derive partially from 
more abstract knowledge, facilitating the use of abstractions 
in generating specialized phrases. Thls organization, used by 
a generator called KING (Knowledge INtensive Generator), 
promotes the extenslbility and adaptability of the generation system. 
1 Introduction 
The task of natural language generation is that of producing 
linguistic output to satisfy the communicative requirements 
of a computer system. The principal limitation of existing 
programs which perform this function is that they fail to 
realize a sufficiently broad range of requirements to demon- 
strate a convincing linguistic capability. This drawback seems 
founded in aspects of the systems which hinder the develop- 
ment of a large base of knowledge about language. A great 
deal of knowledge is required to produce any given utterance, 
yet much of this knowledge cannot easily be exploited across 
a range of utterances. 
Partial success in generation systems is often achieved by 
applying linguistic knowledge to particular domains. Exem- 
plary of this success are text generation programs such as 
PROTEUS \[6\], and Ann \[15,14\]; as well as generation com- 
ponents of on-line systems; for example, in HAM-ANS \[5\], 
UC \[9,11\[, and VIE-LANG \[4 I. These systems, while em- 
bodying a variety of generation techniques, serve to illustrate 
the importance of the command of specialized constructs and 
the ability to utilize specialized knowledge in generation. A 
close examination of the knowledge used in such programs, 
however, reveals that a great deal of linguistic information 
seems to be encoded redundantly, thus impeding the use of 
generalizations in "scaling up ~ the systems. 
The UNIX 2 Consultant system \[22\] is a program which 
answers questions from naive users about the UNIX operating 
system. Scaling up the user interface required a generator 
which could produce responses such as the following: 
1. 'Chmod' can be used to give you write permission. 
2. You don't have write permission on the directory. 
3. You can't get write permission on the directory. 
4. You need ethernet access. 
'This paper is based on the thesis research conducted while the au- thor was at the University of California, Berkeley. The research 
was supported in part by the Office of Naval Research under con- tract N00014-80-C-0732, the National Science Foundation under 
grants IST-8007045 and IST-8208602, and the Defense Advanced 
Research Projects Agency (DO D), ARPA Order No. 3041, Moni- 
tored by the Naval Electronic Systcras Command under contract 
N00039-82-C-0235. 
5. You don't have ethernet access. 
The PHRED generator initially used by UC \[11\] produced 
output such as the above by treating each verb use as an in- 
dependent specialized construct. This allowed no benefit to 
the system of abstract knowledge about the use of the verbs, nor of applying its knowledge about one specialized construct 
to another. This difficulty proves to be a major handicap 
in building large-scale generation systems: A key element 
is to facilitate the exploitation of generalizations while still 
providing for specialized uses. In order for the UC system 
to have this capacity, the linguistic knowledge representation 
used had to be redesigned. 
This knowledge-based approach has led to the design and 
implementation of the Ace knowledge representation frame- 
work \[12\]. Ace is a uniform, hierarchical representation sys- 
tem, which facilitates the use of abstractions in the encoding 
of specialized knowledge as well as the representation of ref- 
erential and metaphorical relationships among concepts. A 
general-purpose natural language generator, KING (Knowl- 
edge INtensive Generator)f10\], has been implemented to ap- 
ply knowledge in the Ace form. The generator works by ap- 
plying structured associations, or mappings, from conceptual 
to linguistic structures, and combining these structures into 
grammatical utterances. This has proven to be a simple but 
powerful mechanism, easy to adapt and extend, and has pro- 
vided strong support for the use of Ace knowledge structures 
in generation. 
While this presentation describes the Ace knowledge struc- tures from the point of view of language production, the rep- 
resentation framework is designed to be unbiased with respect to language analysis or generation. 
The discussion which follows focuses on the representa- 
tion of linguistic and conceptual knowledge in Ace, using as 
an example knowledge about the verbs "give ", "take ", "buy" 
and "sell ". The examples show briefly how information is en- 
coded in Ace which enables the generator to produce dative 
constructs such as "John sold Mary a book ~ and specialized 
forms such as "John gave Mary a kiss", making use of abstract 
knowledge about events such as giving. These verbs provide 
a good testing ground for a representational framework, as 
they may be characterized by certain linguistic generaliza- 
tions while appearing in a variety of specialized constructs. 
For further examples and a description of the generation al- 
gorithm used by KING, the reader is referred to \[10\]. 
2 Ace Fundamentals 
I have suggested that the development of extensible and adapt- 
able natural language systems depends on a knowledge rep- 
resentation framework within which generalizatlons are effec- 
tively exploited. This is the primary goal of the Ace frame- 
work. The starting point of Ace was an implementation of 
a knowledge representation called KODIAK \[21\], which was 
extended to include explicit structured relationships between 
language and meaning. This section presents the basic knowl- 
edge representation principles behind Ace, and provides an 
example of how conceptual knowledge is used to relate knowl- 
edge about selling to knowledge about gi~ng. 
2.1 Basic Principles 
Many knowledge representation systems, however different 
they appear superficially, may be shown to have the same 
SUNIX is a trademark of AT & T Bell Laboratories 
554 
formal expressive or inferential power. This discussion centers 
not on the question of formal power but on the nature of 
the knowledge which must be expressed. The Ace knowledge 
representation provides a framework for expressing essential 
linguistic knowledge in a form suitable for encoding within a 
representational formalism. 
The following principles guide the encoding of knowledge 
important in the generation task: 
o Principle 1. Inheritance of Conceptual Relations. 
Concepts in memory are organized into a hierarchy of 
categories, in which more specific concepts inherit "fea- 
tures" from more general concepts. This inheritance is a 
representational tool which has been employed through- 
out the history of Artificial Intelligence (e.f.\[18,19,2,3\]). 
The question of what exactly is inherited, however, can 
be answered in a variety of ways. Ace takes advantage 
of structured inheritance, (of. \[3\]), in which concepts 
linked to a particular structure may inherit from super- 
categories of that structure. For example, knowledge 
about the seller of a sellin 9 action may be inherited 
from knowledge about the giver of a giving action. 
*, Principle ~. Proliferation of Conceptual Categories. 
Individual concepts are themselves categories, and any 
concept about which there is particular knowledge is 
considered to form a category. Thus categories prolif- 
erate: Probably, there are far more conceptual cate- 
gories than there are lexical items in the system. For 
example, it will be shown later in this section that it 
is reasonable to postulate a concept specifically for the 
action of paying money in exchange for merchandise, 
although there is no lexical item corresponding to this 
concept. The lexical term "pay" is associated with a 
more general concept, that of providing money in ex- 
change for virtually anything. The lexical term "give" 
may be associated with a general giving concept, but 
giving to charity, giving an idea, and giving a chance 
are distinct concepts with distinct linguistic manifesta- 
tions. For example, the use of the verb "give" without 
object or indirect object as in "Bill gave" and "I gave 
at the office" is a linguistic phenomenon which appears 
almost exclusively when referring to charitable giving. 
e Principle 3. Explicit Referential Relationships. 
There are a range of conceptual relationships impor- 
tant in language use which are not easily described as 
factual or ontological relationships. The one which is 
considered here is the view relationship, which helps 
to determine how concepts may be used in express- 
i.ng other concepts. The concepts of giving and tak- 
ing may be related to the concept of a transfer-event, 
but the instantiation of the abstract giving and tak- 
ing concepts cannot be factually inferred from the in- 
stantiation of transfer-event. For example, "John gave 
five dollars to charity" does not imply that a charita- 
ble organization took the five dollars from John. "Mary 
took the money from John" does not imply that John 
gave Mary the money. In many circumstances, how- 
ever, the same event may be described using "give" or 
"take". For example, (1) "John gave Mary five dollars 
for the book" may imply (2) "Mary took five dollars 
from John for the book". Representing giving and tak- ing as views 
of transfer-event permits the encoding of 
knowledge about describing transfer-events without re- 
~luiring a given event to be classified as giving or tak- 
ing. These views may thus represent the knowledge 
that "John took <x> from Mary" and "Mary gave John 
<x>" might be used to describe the same event. Such 
views will be shown to be useful in determining how 
linguistic structures are used to refer to events. 
The next section describes the basic elements of the hier- archical framework of Ace. 
2.2 Structured Asuociatlons in Ace 
Ace makes use of a notation in which there are two types of 
entities: objects and structured associations2 A structured 
association is a relation among two or more objects which 
also relates corresponding objects associated with the related 
objects. 
The most common structured associations in Ace, taken 
from the KODIAK representation, 4 are the DOMINATE, or 
"D", relation, which associates a subcategory with its parent 
category, and the MANIFEST or "m" relation, which asso- 
ciates a category with an aspectual or role. This hierarchical 
system is analogous to isa-links and slots in other similar rep- 
resentation systems; the motivation behind KODIAK was to 
preserve the ideas behind frame-based representations while 
clarifying the semantics of a "slot". For a comparison of KO- 
DIAK with other research, see Wilensky (forthcoming). 
The term ROLE-PLAY, also taken from KODIAK, is used 
to indicate corresponding concepts across structured associa- 
tions. For example, the assertion, 
(DOMINKI'E action selling 
with (ROLE-PLAY actor seller)) 
indicates that selling is a subcategory of action, with seller 
playing the role of actor. A graphical representation of this 
relation, with the ROLE-PLAY implicit, is illustrated in fig- 
ure 1. 
Figure 1: The selling action 
Structured associations with RObE-PLAYs are the basic mechanism for organizing knowledge m Ace. The next section 
describes how these associations are used to represent basic 
knowledge about buying and selling. 
2.3 The Commercial Transaction Example 
I have proposed that generalizing about linguistic constructs 
such as the dative form in "John gave Mary a dollar", "John 
told Mary a story", and "John sold Mary a book", seems 
to depend on the representation of concepts such as giving, 
telling, and selling. This section presents the foundation for 
the encoding of selling in Ace. 
Consider the concept of the commercial transaction \[7\]. 
The commercial-transaction represents an event in which a 
merchandise object is exchanged for legal tender. The essen- 
tial knowledge about this event may be represented by clas- 
sifying the commercial-transaction as a complex-event, com- 
posed of at least two simpler events, ct-merchandise..transfer 
and or-tender-transfer. Each of these two sub-events is a 
kind of transfer-event, and is thus used to associate roles of 
the commercial-transaction with roles of transfer-event. This 
knowledge is captured in figure 2. 
Figure 2 illustrates the important knowledge that the mer- 
chant receives the tender fl'om the customer, and the cus- tomer 
receives the merchandise from the merchant. Con- cepts such as merchant, cu,.itomer, merchandise, and tender 
are aspectuals of the commercial-transaction; that is, they 
"This term, and the idea of using general structured associations as a 
language processing tool. are due to Wile.sky. 
4These associations, as well as m~my of the ideas here, have evolved 
during a series of seminars among the Berkeley Artificial Intelligence 
Research group, led by Rohart Wileasky. Other participants in these 
discussions were: Richard AIterma~b Margaret Butler, David Chin, 
Charley Cox, Marc Laria, Anthony Maida, James hl~ tin, James May- 
field, Peter Norvig, Lisa Rau. au'd Nigel Ward. 
555 
are specific concepts whose meaning is undetachable from the commercial-transaction 
event. However, much of the knowl- 
edge about these concepts, such as the recipient and source 
roles, is inherited from other concepts. As in other frame-like 
systems \[19,2\], this organization allows roles of a concept to 
be inherited in this manner. The ROLE-PLAY relationship 
in Ace, however, permits more than this simple form of inher- 
itance: It allows for the semantics of aspectuals to be defined 
in terms of other aspectuals. For example, the meaning of 
the merchant aspectual of the commercial-transaction here is 
represented in part by the ROLE-PLAY relation which links 
this aspectual to the source of the d-merchandise-transfer and 
that which links it to the recipient of the st-tender-transfer. 
The assertions above form an im:port ant core of knowledge 
about commercial-transactions. This knowledge is important 
Figure 2: The commercial-transaction event 
in the way language is used to describe such events. For ex- 
ample, it will be shown in section 4 that the knowledge that 
merchandise and tender play object roles is linked to knowl- 
edge about transitive verb forms, so that phrases such as 
"bought a book" and "paid five dollars" conform to a general 
rule. The next section discusses how concepts such as buying 
and selling, used to refer to the commercial-transaction con- 
cept, are represented in Ace. 
2.4 Actions as VIEWs of Events 
I have proposed that verbs such as "give" and "take" refer 
to the actions giving and taking, and thus refer indirectly to 
the general transfer-event concept. The motivation for this 
analysis is to facilitate the representation of knowledge about 
the roles which giver and taker play, thereby enabling "John 
gave Mary a book" and "Mary took a book from John" to 
describe indirectly the same event, as "All gave Frazier a 
punch", and "Frazier took a punch fl'om Ali" may indirectly 
describe the same event. The commercial-transaction event is generally described 
using the verbs "buy", "sell", and "pay". "Sell" and '~pay ~ 
behave similarly to the verb "give"; "buy", behaves more llke 
"take". For example, "John sold Mary a book", and "Mary 
paid five dollars for the book" both use the dative form, and 
"John bought the book from Mary" exhibits a structure iden- 
tical to "John took the book from Mary". The representation 
of the concepts buying and selling in Ace relates these con- 
cepts to giving and taking so that knowledge about expressing giving 
and taking may be used also for buying and selling. 
556 
The concepts giving and taking in Ace are related to the transfer-event 
concept by a structured association called a VIEW. 
Inspired by the notion of a view in earlier knowledge 
representations \[17,2\], the Ace VIEW is applied to metaphor- 
ical and analogical relationships, similar to those described 
in \[16,8\]. VIEWs are used to represent knowledge about 
concepts which may be used in expressing other concepts. 
The Ace network in figure 3 represents the basic knowledge 
about giving, taking, buying, and selling. In this hierarchy, the 
structured association view1 between transfer-event and giv- ing 
DOMINATEs the structured association view3 between c t-merchandise-transfer 
and selling, and view2between transfer event 
and taking DOMINATES view$ between st-merchandises 
transfer and buying. 
The representation in figure 3 demonstrates on a small 
scale how the hierarchical arrangement of VIEWs is used in 
the encoding of structured associations. Structured associa-. 
tions such as view1 between transfer-event and giving DOM~ 
INATE other more specific relations, such as view3. Note 
that this makes the explicit representation of ROLE-PLAY 
relations for view3 unnecessary, as the relationship between merchant 
and seller in view3 is specified by the relationship 
between source and giver in view1. 
The representation of the selling concept is a simple ex- 
ample of how Ace encodes abstractions which may be used 
in language processing. The abstraction here is the relation- 
ship between a general category giving and a general category 
transfer-event. There are two ways in which this abstraction 
may be used: (1) A more specific association may be rep- 
resented as a subcategory of the abstract association. This 
is the case in the selling example presented here. In this 
case, knowledge about the abstract association may be used 
in applying the specific association, thus knowledge about 
expressing an abstract concept may be used in expressing a 
more specific concept. This allows much of the same knowl- 
edge to be used for phrases involving "giving" and "selling". 
(2) A concept which is associated by another VIEW with the 
abstract concept may then also be expressed using the ab- 
stract VIEW. This is the case with expressions such as "give 
a punch" (cf. section 4.3), which takes advantage of the ab- 
stract action as transfer-event view in combination with the transfer-event as giving 
VIEW. 
The next two sections discuss the application of the Ace 
framework to the representation of linguistic knowledge and 
to the representation of the knowledge which associates lin- 
guistic and conceptual structures. 
Figure 3: The hierarchy of VIEWs 
3 Basic Grammatical Knowledge 
The principles outlined in section 2.1 motivate a representa- 
tion in which knowledge is dispersed throughout a hierarchy, 
with a greater number of structures, each containing more 
limited information. Linguistic knowledge in Ace is organized 
into a hierarchy which incorporates this type of organization. 
In fact, the same knowledge representation language is used 
to encode both linguistic and conceptual knowledge in Ace. 
The linguistic hierarchy provides for a simple set of basic lin- 
guistic templates, with little redundancy and relative ease of 
extension. The following principles of linguistic representa- 
tion are suggested: 
Princ@h, $. Inheritance of Linguistic Features. 
Sets of features which are common to a certain class of templates need not be specified independently for each 
template in the class. Thus, if there is a set of fea- 
tures shared among passive sentences~ or among prepo- 
sitional phrases, these features belong by default to any 
template in the class. This eliminates the need for fully 
specifying the structure of phrases which have special- 
ized properties or meaning• 
o Principle 5. Proliferation of Linguistic Categories. 
In order to take advantage of the inheritance of fea- 
tures, there must be a wide range of classes of iinguls- 
tic templates which share sets oi features. Often these 
categories depart from the traditional syntactic clas- 
sifications. Requiring that a template bca member 
of a unique category in the case of a gerund or nom- 
inalization can prove difficult, as these may inherit cer° 
rain attributes fl'om verbs and certain attributes from nouns Thus any template may inherit features, includ- 
ing structural descnptmns, from multiple categorms. 
Categories are arranged hierarchically, so each category 
inherits from all its ancestors in the hierarchy. 
o Principle 6. Distinguishing Grammatical Relations from Grammatical Patterns. 
A great deal of linguistic information seems associated 
with structural relationships between linguistic constit- 
uents which are dependent neither on their order in a 
surface structure nor on the precise nature of the struc- 
ture in which they appear. For example, the relation be- 
tween subject and verb retains its linguistic features re- 
gardless of how the subject and verb appear in any sur- 
face structure: The agreement between subject and verb 
in ~John was given the book by Mary" is the same as in 
aWas John given the book by Mary?" as is the concep- 
tual recipient role which John plays. Such information 
does not pertain to a particular surface structure, but 
to any surface structure in which a noun phrase and 
verb are in the subject-verb relation. In "John kissed 
Mary on the cheek" and "The kiss on the cheek pleased 
Mary", the role of %n the cheek" as it relates to "kiss" 
is independent of whether the prepositional phrase is 
part of a verb phrase or noun phrase. In general, struc- 
tural linguistic relationships are not limited to those 
which are directly linked to constituent order, and thus 
a more general facility than a syntactic pattern is re- 
quired to represent these relationships. 
o Principle 7. Uniformity of Representation. 
Linguistic knowledge is knowledge, and thus cart be en- 
coded using the same representational framework as 
conceptual knowledge. The same structured associa- 
tion~ used in the conceptual hierarchy can be used in 
the linguistic hierarchy, u Having such uniformity of 
representation has the practical vahm of facilitating the 
interaction of conceptual and linguistic structures. 
SThe use of a knowledge representation language to encode syntactic knowledge for file purpose of semantic interpretation ha.~ been praco 
riced with KL-ONE \[20\] and its successors, also favoring uniformity of' representatiom Such t;ystcms have not been used, to my knowledge, to 
encode associations between conceptual and linguistic knowledge. 
The discussion which follows demonstrates how linguistic 
knowledge may be encoded using the framework described 
in the previous section, and discusses the effect of property 
inheritance on linguistic knowledge representation. 
3.1 Multiple Inheritance in the Ace Lin- 
guistic Hierarchy 
This section shows how the structured associations of Ace, 
and particularly the capacity for multiple inheritance, are 
used to encode some of the linguistic knowledge used in the 
construction of simple sentences. 
~$.1.1 Verb phrase~ i~ Ace 
Verb phrases provide a good example of the use of a linguistic 
hierarchy because they exhibit a variety of surface forms while 
obeying certain regularities. One type of verb phrase is the dative-vp, a verb-phrase 
made up of a constituent dvp-indir, 
which includes the verb and noun phrase corresponding to 
the indirect object, followed by the nonn phrase correspond- 
ing to the direct object. The treatment of the dvpoindir as 
a separate constituent is done to facilitate the handling of 
other dative forms• Figure 4 shows how knowledge about the 
dative verb phrase pattern is encoded in Ace, as well as how 
this dative-vp is positioned in the verb phrase hierarchy. The 
pattern dvp-pattern represents the ordering of the constitu- 
ents of the verb phrasej while t'he relation dvp-indir is used to 
represent the relationship between the dative verb and indi- 
rect object. Section 4 will show how this relation is associated 
with conceptual knowledge° By allowing aspectuals which represent patterns and pat° 
tern constituents to play multiple roles, the representation of 
linguistic knowledge as shown in figure 4 shows how knowl.- 
edge about infinitives, gerunds, and verb phrases is distributed 
in the Ace hierarchy. The various verb phrase patterns fall 
beneath the verb-phrase template, as do the gerund phrase, 
infinitive phrase, and finite, verb phrase nodes, in order to 
dative~ 
,st cans~ ¢c ol~j 
Figure 4: The verb phrase hierarchy 
produce a finite verb ~hrasel a node lower ill the hierarchy 
must be instantiated, this hmrarclfical organization permits 
the gerund phrase and infinitive phrase to have the same lin- 
guiqtic structure as the verb phrase, modulo the fore1 of the 
verb part. An instantiated verb-phrase in Ace thns inherii, s 
most of it,~ internal structure from one category, for example dative-vp, 
and its external behavior from another, for exam° 
pie, finite-verb-phrase. The verb~phrase category !tself plays 
no external syntactic role---there is no pattern in which a con° 
stituenlJ belongs to the verb-.ph,'ase category and to no lower 
category~ although theoretically there could be. The effectiyc 
organization of information ~,bout verb phrases stems directly 
from the application of the basic knowledge x~presentatioa 
principles of Ace ~o linguistic knowledge. 
557 
This section has shown how a hierarchical representation 
is used to encode linguistic relations. The next section con- 
centrates on how the representation of explicit referential rela- 
tionships links this linguistic knowledge to conceptual knowl- 
edge. 
4 Associating Language and Mean- ing 
Section 2.1 presented the foundations of the Ace framework 
and its application to meaning representation, proposing that 
concepts be hierarchically organlzed, linked together via struc- 
tured associations. Section 3 described how linguistic knowl- 
edge can also be hierarchically organized within this frame- 
work, and how features can thus be inherited through linguis- 
tic categories. The idea of a linguistic relation was proposed, 
to distinguish structural relationships from ordering relation- 
ships in grammatical constructs. Naturally, the knowledge 
required to produce correct and appropriate utterances in- 
c\],udes both of these classes in addition to the links which 
DIUO l;ne ClasSes ~oge~ner. 
4.1 Basic Principles 
The goal of taking advantage of explicit referential knowledge, 
discussed in section 2.1, along with the framework presented 
in the previous sections, suggests two principles directed to- 
wards the association of linguistic and conceptual knowledge: 
• Representation Principle 8, Correspondence of Linguis- tic and Conceptual Structures. 
Linguistic structures, such as lexical terms, linguistic 
categories, and grammatical structures, are directly as- 
soclated with conceptual categories. General linguis- 
tic categories in Ace, such as verb-indir-relation, corre- 
spond to general conceptual categories, such as transfer- event. 
Specific lexical terms, such as "buy" and "sell', are linked to more specific concepts, such as buying and 
selling. Grammatical structures, such as the modifier- noun 
relation, are associated with conceptual relations, 
such as MANIFEST. 
• Representation Principle 9. Association of Linguistic Features tuith Conceptual Attributes. 
The structured association permits the relation of lin- 
istlc aspectuals with conceptual aspectuals via ROLE- 
AY. Linguistic features, such as dvp-pattern and ivp- verb as 
described in section 3 are represented as aspec- 
tunis in Ace, as their status depends on the template 
of which they are a part. The association of these fea- 
tures with conceptual attributes goes along with the 
association of the template with a concept. For exam- 
ple, the indirect object feature iobj is an aspectual of verb-indir-relation. 
The linguistic feature iobj is linked 
to the conceptual attribute recipient, an aspectual of 
the transfer-event concept. The direct-object feature obj, 
an aspectual of verb-obj-relation~ is linked to the 
conceptualobject concept, an aspeetual of the simple- event 
concept. Thus linguistic features have conceptual 
correlates. 
The next section describes how these principles are real- 
ized in the association of language and meaning in Ace. 
4.2 Linking Linguistic and Conceptual Struc- 
tures Using REF 
The main tool for representing relationships between linguis- 
tic structures and conceptual structures in Ace is a structured 
association called REF. The REF association, designating 
"referential" relationships, is similar to VIEW, except that 
it joins language to concepts instead of concepts to concepts. 
The linguistic knowledge presented earlier has included 
information about how verbs and their objects or indirect 
objects may appear in surface structure. The knowledge us- 
sential to build these structures, however, is contained in the 
correspondence between linguistic relations and conceptual 
entities. This information is presented in figure 5. 
The diagrams in figure 5 illustrate how the syntactic struc- 
ture of the dative verb phrase is associated with its conceptual 
structure. The verb part and noun phrase in the ivp-pattern 
in the top diagram belong to the verb-indir-relation, which 
associates with the indirect object the concept of recipient. 
This diagram presents a slightly abbreviated version of the 
Ace representation given in section 3, as ivp-verb and ivp-np 
are associated only indirectly with the verb-indir-relation via 
the ivp-relation aspectual. 
The etructuredassoeiation between the verb-obj-relation 
and the simple-event concept in the lower diagram of Figure 
5 links the object of the verb in the dative verb phrase to 
the object of the simple event. In this association, as in the 
association of the verb-indir-relation with the transfer-event, 
the verb part dvp-verb is associated via ROLE-PLAY with 
the event itself, rather than with any aspectual of the event. 
Like the pattern-concept pair \[1 l\]or the unification gram- 
mar template \[13,1\], REF is a means of associating linguistic 
structure with conceptual structure. The template, however, 
is replaced with an explicit structural link in the knowledge 
Figure 5: Knowledge linking verb phrases to events 
network. This makes it easier to perform the knowledl~e- 
driven aspects of generation because no querying or complex 
matching is necessary. The use of the REF association also fa- 
cilitates incremental generation by encoding knowledge about 
referential relationships as structured associations at various 
levels in the Ace hierarchy: The construction of a complete 
sentence involves the combination of structures derived from a number of REF associations, each of which refines those 
structures already active, e 
4.3 Knowledge About Specialized Construe 
The examples given in the previous section are simple illustra- 
tions of the use of associations between language and meaning 
at a variety of levels. These levels become especially appar- 
ent in the representation of knowledge about specialized con- 
structs, whose meaning can be only partially represented as 
associations from more general linguistic structures. 
~Those familiar with KL-ONE and other similar representation lan- guag~ will observe that the Ace representation tends to avoid matching 
by assuming that the concept being generated from is assigned to a spe- cific category. Thus some of the work done by template matching in 
systems such as PHRED is done during the classification of a concept in Ace. This is consistent with the Proliferation of Conceptual Categories 
.principle: the generator need not obtain conceptual information that is more specific than that contained in the category in which a concept 
has been classified. 
558 
The means in which specialized knowledge is encoded varies 
according to the particular construct. The "give a hug" ex- 
pression is interesting because there is no single syntactic 
structure which can be identified with the specialized inter- 
pretation, yet intuitively the specialized meaning seems tied 
to the use of the verb "give" in conjunction with the object 
"hug". While the metaphorical connection between the hug- ging 
action and the giving action is dependent on the "act- 
ing upon is giving" metaphor, this "hugging is hug-giving" metaphor must also be associated with the particular lexical 
items "give" and "hug". Figure 6 shows how both objec- 
tives may be accomplished: The link between hug-giving and hug-transfer 
is labeled view1" to indicate that it is DOM- 
INATEd by the structured association between giving and transfer-event. 
This association, view1, relates the source of 
the transfer-event to the giver or actor of the giving. By in- 
heritance and ROLE-PLAY, viewI'associates the giver of the hug-giving 
action with the source of the hug-transfer event. 
The views view£ and view3 represent metaphorical associa- 
tions that actions may be VIEWed as transfer-events from 
the actor to the object. View3 indicates the correspondence 
between the object of an action or event and the recipient 
of the transfer, and between the event and the object of the 
transfer. View2 represents the relationship between source 
and actor. The association between the hug-transfer event and the 
hugging action, labeled viewg,3" allows the inheritance of the 
knowledge that the source of the hug-transfer corresponds to 
the actor of the hugging action, that the object of the hug- 
transfer corresponds to the hugging action itself, and that the recipient 
of the hug-transfer corresponds to the object of the hugging. 
The view2,3' association is DOMINATEd by both view2 
and view3 and thus inherits their ROLE-PLAYs. 
Knowledge about specialized constructs, such as "giving a 
hug", makes use of abstract structured associations between 
language and meaning just as knowledge about selling makes 
use of knowledge about giving. In this way the Ace repre- 
sentation is used to take advantage of generalizations in the 
encoding of specialized knowledge. 
Figure 6: Knowledge used for "giving a hug" 
5 Summary and Conclusion 
The knowledge framework for language generation presented 
here shows how generalizations may be effectively exploited in 
the representation of linguistic knowledge. The application of 
the Ace framework to the representation of knowledge about selling 
shows how knowledge about giving may be used to 
represent information used in constructing utterances with 
the verb "sell". The example of "g!ving a hug" is used to 
demonstrate how the notion of a mew captures knowledge about metaphorical constructs. The representation embodies 
a uniform, parsimonious encoding of conceptual and linguis- 
tic knowledge which seems to promote the extensibility of 
natural language systems. 

References 

\[1\] D. Appelt. TELEGRAM: a grammar formalism for lan- 
guage planning. In Proceedings of the £Ist Annual Meet- ing of the Association. for Computational Linguistics, 
Cambridge, Massachusetts, 1983. 

\[2\] D. Bobrow and T. Winograd. An overview of KRL, a 
knowledge representation language. Cognitive Science, 
1(I), 1977. 

\[3\] R. Brachman. et. al. Research in Natural Language Un- 
derstanding. Technical Report 4274, Bolt Beranek and Newman, 1979. 

\[4\] E. Buchberger and H. Horacek. The generation compo- 
nent of the system VIE-LANG. In First International Workshop on Language Generation, 
Burg Stettenfels, West Germany, 1983. 

\[5\] S. Busemann. Topicalization and pronominalization: ex- 
tending a natural language generation system. In Pro- ceedings of the Sixth European Conference on Artificial 
Intelligence, Pisa, Italy, 1984. 

\[6\] A. Davey. Discourse production. Edinburgh University 
Press, Edinburgh, 1979. 

\[7\] C. J. Filhnore. Topics in lexical semantics. 1977. 

\[8\] D. Gentner. Structure-mapping: a theoretical frame- 
work for analogy. Cognitive Science, 7, 1983. 

\[9\] P. Jacobs. Generation in a natural language interface. In 
Proceedings of the Eighth International Joint Conference on Artificial Intelligence, 
Karlsruhe, Germany, 1983. 

\[10\] P. Jacobs. A knowledge-based approach to language 
production. PhD thesis, University of California, 
Berkeley, 1985. Computer Science Division Report 
UCB/CSD86/254. 

P. Jacobs. PHRED: a generator for natural language 
interfaces. Computational Linguistics, 11(4), 1985. 

P. Jacobs and L. Ran. Ace: associating language with 
meaning. In Proceedings of the Sixth European Confer- ence on Artificial Intelligence, 
Pisa, Italy, 1984. 

M. Kay. Fanctional Unification Grammar: a formal- 
ism for machine translation. In Proceedings of the Tenth International Conference on Computational Linguistics, 
Stanford, California, 1984. 

K. Kukich. Knowledge-Based Report Generation: A 
Knowledge-Engineering Approach to Natural Language Report Generation~ 
PhD thesis, University of Pitts- 
burgh, 1983. 

K. Kukich. Semantic and syntactic constraints in natural 
language stock reports. In proceedings of the $Sth annual meeting of ASIS, 
1982. 

G. Lakoff and D. Johnson. Metaphors we Live By. Uni- 
versity of Chicago Press, Chicago, 1980. 

J. Moore and A. Newell. How can MERLIN understand? 
In L. Gregg, editor, Knowledge and Cognition, Erlbaum 
Associates, Halsted, New Jersey, 1974. 

M. R. Quillian. Semantic Memory. Bolt Beranek and 
Newman, Cambridge, Massachusetts, 1966. 

R. B. Roberts and I. P. Goldstein. The FRL Manual. 
Technical Report AIM-408, MIT AI Lab, 1977. 

N. Sondheimer, R. Weischedel, and R. Bobrow. Seman- 
tic interpretation using KL-ONE. In Proceedings of the 
Tenth International Conference on Computational Lin- guistics, 
Palo Alto, 1984. 

R. Wilensky. KODIAK- a knowledge representation lan- 
guage. In Proceedings of the Sizth Annual Conference of the Cognitive Science Society, 
Boulder, Colorado, 1984. 

R. Wilensky, Y. Are, s, and D. Chin. Talking to UNIX 
in English: an overview of UC. Communications of the 
Association for Computing Machinery, 27(6), 1984. 
