THE TEXT SYSTEM FO~NATURAL LANGUAGE GENERATION: 
AN OVERVIEW* 
Kathleen R. M::Keown 
Dept. of Computer & Information Science 
The Moore School 
University of Pennsylvania 
Philadelphia, Pa. 19104 
ABSTRACT 
Computer-based generation of natural language 
requires consideration of two different types of 
problems: i) determining the content and textual 
shape of what is to be said, and 2) transforming 
that message into English. A computational 
solution to the problems of deciding what to say 
and how to organize it effectively is proposed 
that relies on an interaction between structural 
and semantic processes. Schemas, which encode 
aspects of discourse structure, are used to guide 
the generation process. A focusing mechanism 
monitors the use of the schemas, providing 
constraints on what can be said at any point. 
These mechanisms have been implemented as part of 
a generation method within the context of a 
natural language database system, addressing the 
specific problem of responding to questions about 
database structure. 
1.0 INTRODUCTION 
Deciding what to say and how to organize it 
effectively are two issues of particular 
importance to the generation of natural language 
text. In the past, researchers have concentrated 
on local issues concerning the syntactic and 
lexical choices involved in transforming a 
pre-determined message into natural language. The 
research described here ~nphasizes a computational 
Solution to the more global problems of 
determining the content and textual shape of what 
is to be said. ~re specifically, my goals have 
been the development and application of principles 
of discourse structure, discourse coherency, and 
relevancy criterion to the computer generation of 
text. These principles have been realized in the 
TEXT system, reported on in this paper. 
The main features of the generation method 
used in TEXT include I) an ability to select 
relevant information, 2) a system for pairing 
rhetorical techniques (such as analogy) with 
discourse purv~ses (such as defining terms) and 
3) a focusing mec~mnism. Rhetorical techniques, 
which encode aspects of discourse structure, guide 
the selection of information for inclusion in the 
text from a relevant knowledge poq~l - a subset of 
*This work was partially supported by National 
Science ~Dundation grant #MCS81-07290. 
the knowledge base which contains information 
relevant to the discourse purpose. The focusing 
mechanism helps maintain discourse coherency. It 
aids in the organization of the message by 
constraining the selection of information to be 
talked about next to that which ties in with the 
previous discourse in an appropriate way. These 
processes are described in more detail after 
setting out the framework of the system. 
2.0 APPLICATION 
In order to test generation principles, the 
TEXT system was developed as part of a natural 
language interface to a database system, 
addressing the specific problem of generating 
answers to questions about database structure. 
Three classes of questions have been considered: 
questions about information available in the 
database, requests for definitions, and questions 
about the differences between database entities 
\[MCKE(3WN 80\]. In this context, input questions 
provide the initial motivation for speaking. 
Although the specific application of 
answering questions about database structure was 
used primarily for testing principles about text 
generation, it is a feature that many users of 
such systems would like. Several experiments 
(\[MALHOTRA 75\], \[TENNANT 79\]) have shown that 
users often ask questions to familiarize 
themselves with the database structure before 
proceeding to make requests about the database 
contents. The three classes of questions 
considered for this system were among those shown 
to be needed in a natural language database 
system. 
Implementation of the TEXT system for natural 
language generation used a portion of the Office 
of Naval Research (ONR) database containing 
information about vehicles and destructive 
devices. Some examples of questions that can be 
asked of the system include: 
> What is a frigate? 
> What do you know about submarines? 
> What is the difference between a 
and a kitty hawk? 
whisky 
113 
The kind of generation of which the system is 
capable is illustrated by the response it 
generates to question (A) below. 
A) ~at kind of data do you have? 
All entities in the (INR database have DB 
attributes R~MARKS. There are 2 types of 
entities in the ONR database: destructive 
devices and vehicles. The vehicle has DB 
attributes that provide information on 
SPEED-INDICES and TRAVEL-MEANS. The 
destructive device has DB attributes that 
provide information on LETHAL-INDICES. 
TEXT does not itself contain a facility for 
interpreting a user's questions. Questions must 
be phrased using a simple functional notation 
(shown below) which corresponds to the types of 
questions that can be asked . It is assumed that 
a component could be built to perform this type of 
task and that the decisions it must make would not 
affect the performance of the generation system. 
I. (definition <e>) 
2. (information <e>) 
3. (differense <el> <e2>) 
where <e>, <el>, <e2> represent entities in the 
database. 
3.0 SYSTEM OVERVIEW 
In answer ing a question about database 
structure, TEXT identifies those rhetorical 
techniques that could be used for presenting an 
appropriate answer. On the basis of the input 
question, semantic processes produce a relevant 
knowledge pool. A characterization of the 
information in this pool is then used to select a 
single partially ordered set of rhetorical 
techniques from the various possibilities. A 
formal representation of the answer (called a 
"message" ) is constructed by selecting 
propositions from the relevant knowledge pool 
which match the rhetorical techniques in the given 
set. The focusing mechanism monitors the matching 
process; where there are choices for what to say 
next (i.e. - either alternative techniques are 
possible or a single tec~mique matches several 
propositions in the knowledge pool), the focusing 
mechanism selects that proposition which ties in 
most closely with the previous discourse. Once 
the message has been constructed, the system 
passes the message to a tactical component 
\[BOSSIE 81\] which uses a functional grammar 
\[KAY 79\] to translate the message into English. 
4.0 KNOWLEDGE BASE 
Answering questions about the structure of 
the database requires access to a high-level 
description of the classes of objects ino the 
database, their properties, and the relationships 
between them. The knowledge base used for the 
TEXT system is a standard database model which 
draws primarily from representations developed by 
Chen \[CHEN 76\], the Smiths \[SMITH and SMITH 77\], 
Schubert \[SCHUBERT et. al. 79\], and Lee and 
Gerritsen \[LEE and GERRITSEN 78\]. The main 
features of TEXT's knowledge base are entities, 
relations, attributes, a generalization hierarchy, 
a topic hierarchy, distinguishing descriptive 
attributes, supporting database attributes, and 
based database attributes. 
Entities, relations, and attributes are based 
on the Chen entity-relationship model. A 
generalization hierarchy on entities 
\[SMITH and ~94ITH 77\], \[LEE and GERRITSEN 78\], and 
a to~ic hierarch Y on attributes 
\[SCHUBERT et. al. 79\] are also used. In the topic 
hierarchy, attributes such as MAXIMUM SPEED, 
MINIMUMSPEED, and ECONOMIC SPEED are gene?alized 
as SPEED INDICES. In -the general ization 
hierarchy,--entities such as SHIP and SUBMARINE are 
generalized as WATER-GOING VEHICLE. ~he 
generalization hierarchy includes both 
generalizations of entities for which physical 
records exist in the database (database entity 
classes) and sub-types of these entities. The 
sub-types were generated automatically by a system 
developed by McCoy \[MCCOY 82\]. 
An additional feature of the knowledge base 
represents the basis for each split in the 
hierarchy \[LEE and GERRITSEN 78\]. For 
eneralizations of the database entity classes, 
partltlons are made on the basis of different 
attributes possessed, termed sup\[~or tin~ db 
attributes. For sub-t~pes of the database entit-y 
classes, partitions are made on the basis of 
different values possessed for given, shared 
attributes, termed based db attributes. 
~dditional d esc r i pt ive--"--in fo"~a t ion that 
distinguishes sub-classes of an entity are 
captured in ~ descriptive attributes 
(DDAs). For generalizati6ns Of 6he database 
entity classes, such DDAs capture real-world 
characteristics of the entities. Figure 1 shows 
the DDAs and supporting db attributes for two 
generalizations. (See \[MCCOY 82\] for discussion 
of information associated with sub-types of 
database entity classes). 
114 
(ATER-VEHIC 9 
'rP&VEI-MEDIUM 
/ ~DE~A~R (DDA) SURFACE (DDA) 
-DRAFT,DISPLACEMENT -DEPTH, MAXIMHM 
(s~rting dbs) SUBM<GED SPEED 
(supporting dbs) 
FIGURE i DDAS and supporting db attributes 
5.0 SELECTING RELEVANT INFOPJ~ATION 
The first step in answering a question is to 
circumscribe a subset of the knowledge base 
containing that information which is relevant to 
t~ question. This then provides limits on what 
information need be considered when deciding what 
to say. All information that might be relevant to 
the answer is included in the partition, but all 
information in the partition need not be included 
in the answer. The partitioned subset is called 
the relevant ~ow~l~_~e pool. It is similar to 
what Grosz has called mglo-6~ focus" \[GROSZ 77\] 
since its contents are focused throughout the 
course of an answer. 
The relevant knowledge pool is constructed by 
a fairly simple process. For requests for 
definitions or available information, the area 
around the questioned object containing the 
information immediately associated with the entity 
(e.g. its superordinates, sub-types, and 
attributes) is circumscribed and partitioned from 
the remainir~ knowledge base. For questions about 
tk~ difference between entities, the information 
included in the relevant knowledge pool depends on 
how close in the generalization hierarchy t~ two 
entities are. For entities that are very similar, 
detailed attributive information is included. For 
entities that are very different, only generic 
class information is included. A combination of 
this information is included for entities falling 
between t~se two extremes. (See \[MCKEOWN 82\] for 
further details). 
6.0 R~LETORICAL PREDICATES 
~%etorical predicates are the means which a 
speaker has for describing information. ~hey 
characterize the different types of predicating 
acts s/he may use and delineate the structural 
relation between propositions in a text. some 
examples are "analogy" (comparison with a familiar 
object), "constituency" (description of sub-parts 
or sub-types), and "attributive" (associating 
properties with an entity or event). Linguistic 
discussion of such predicates (e.g. \[GRIMES 75\], 
\[SHEPHERD 26\]) indicates that some combinations 
are preferable to others. Moreover, Grimes claims 
that predicates are recursive and can be used to 
identify the organization of text on any level 
(i.e. - proposition, sentence, paragraph, or 
longer sequence of text), alti~ugh he does not 
show how. 
I have examined texts and transcripts and 
have found that not only are certain combinations 
of rhetorical tec~miques more likely than others, 
certain ones are more appropriate in some 
discourse situations than others. For example, I 
found that objects were frequently defined by 
employing same combination of the following means: 
(i) identifying an item as a memDer of some 
generic class, (2) describing an object's 
function, attributes, and constituency (either 
physical or class), (3) making analogies to 
familiar objects, and (4) providing examples. 
These techniques were rarely used in random order; 
for instance, it was common to identify an item as 
a member of some generic class before providing 
examples. 
In the TEXT system, these types of standard 
patterns of discourse structure have been captured 
in schemas associated with explicit discourse 
purposes. The schemas loosely identify normal 
patterns of usage. The~ are not intended to serve 
as grammars of text. The schema shown be-~ 
~rves the purposes o~ providing definitions: 
Identification Schema 
identification (class&attribute/function) 
\[analogy~constituency~attributive\]* 
\[particular-illustration~evidence\]+ 
{amplification~analogy~attributive} 
{particular-illustration/evidence} 
Here, "{ \]" indicates optionality, "/" 
indicates alternatives, "+" indicates that the 
item may appear l-n times, and "*" indicates that 
the item may appear O-n times. The order of the 
predicates indicates that the normal pattern of 
definitions is an identifying pro~'~tion followed 
by any number of descriptive predicates. The 
speaker then provides one or more examples and can 
optionally close with some additional descriptive 
information and possibly another example. 
TEXT's response to the question "What is a 
ship?" (shown below) was generated using the 
identification schema. ~e sentences are numbered 
to show the correspondence between each sentence 
and the predicate it corresponds to in the 
instantiated schema (tile numbers do not occur in 
the actual output). 
115 
(definition SHIP) 
Schema selected: identification 
i) identification 
2) evidence 
3) attributive 
4) particular-illustration 
I) A ship is a water-going vehicle that 
travels on the surface. 2) Its surface-going 
capabilities are provided by the DB attributes 
DISPLACEMENT and DRAFT. 3) Other DB 
attributes of the ship include MAXIMUM_SPEED, 
PROPULSION, FUEL (FUELCAPACITY and 
FUEL_TYPE), DIMENSIONS, SPEED DEPENDENT RANGE 
and OFFICIAL NAME. 4) The-- ~ES,-- for 
example, has MAXIMUM SPEED of 29, PROPULSION 
of STMTURGRD, FUEL~f 810 (FUEL CAPACITY) and 
BNKR (FUEL TYPE), DIMENSIONS of ~5 (DRAFT), 46 
(BEAM), and 438 (LENGTH) and 
SPEED DEP~DENT RANGE of 4200 (ECONOMIC_RANGE) 
and 2~00 (ENDUP~NCE_RANGE). 
Another strategy commonly used in the 
expository texts examined was to describe an 
entity or event in terms of its sub-parts or 
sub-classes. This strategy involves: 
I) presenting identificational or attributive 
information about the entity or event, 
2) presenting its sub-parts or sub-classes, 
3) discussing attributive or identificational 
information with optional evidence about each of 
the sub-classes in turn, and 4) opt--'l-6~al~'y 
returning to the orig-{nal-~ity with additional 
attributive or analogical information. The 
constituency schema, shown below, encodes the 
techniques used in £his strategy. 
The Constituency Schema 
attributive/identification (entity) 
constituency (entity) 
{ attributive/identification 
(sub-classl, sub-class2,..) 
{evidence 
(sub-classl, sub-class2, ...)} }+ 
{attributive/analogy (entity) } 
TEXT'S response to the question "What do you 
know about vehicles?" was generated using the 
constituency schema. It is shown below along with 
the predicates that were instantiated for the 
answer. 
(information VEHICLE) 
J 
Schema selected: constituency 
i) attributive 
2) constituency 
3) attributive 
4) attributive 
5) attributive 
i) The vehicle has DB attributes that 
provide information on SPEED INDICES and 
TRAVEL MEANS. 2) qhere are 2- types of 
vehicl~s in the ONR data~\]se: aircraft and 
water-going vehicles. 3) The water-going 
vehicle has DB attributes that provide 
information on TRAVEL MEANS and 
WATER GOING OPERATION. 4) The ~ircraft has DB ° 
attributes -- that provide information on 
TRAVEL MEANSf FLIGHT RADIUS, CEILING and ROLE. 
Other DB attributes -of the vehicle include 
FUEL( FUEL_CAP~EITY and FUEL_TYPE) and FLAG. 
Two other strategies were identified in the 
texts examined. These are encoded in the 
attributive schema, which is used to provide 
detailed information about a particular aspect of 
an entity, and the compar e and contrast schema, 
which encodes a strategy --~r contrasting two 
entities using a description of their similarities 
and their differences. For more detail on these 
strategies, see \[MCKEGWN 82\]. 
7.0 USE OF THE SCHEMAS 
As noted earlier, an examination of texts 
revealed that different strategies were used in 
different situations. In TEXT, this association 
of technique with discourse purpose is achieved by 
associating the different schemas with different 
question-types. For example, if the question 
involves defining a term, a different set of 
schemas (and therefore rhetorical techniques) is 
chosen than if the question involves describing 
the type of information available in the database. 
The identification schema can be used in 
response to a request for a definition. The 
purpose of the attributive schema is to provide 
detailed information about one particular aspect 
of any concept and it can therefore be used in 
response to a request for information. In 
situations where an object or concept can be 
described in terms of its sub-parts or 
sub-classes, the constituency schema is used. It 
may be selected in response to requests for either 
definitions or information. The compare and 
contrast schema is used in response ~o a questl'i'~ 
about the difference between objects. A surmary 
of the assignment of schemas to question-types is 
shown in Figure 2. 
116 
Schemas used for TEXT 
i. 
2. 
3. 
4. 
identification 
-requests for definitions 
attributive 
-requests for available information 
constituency 
-requests for definitions 
-requests for available information 
compare and contrast 
-requests about the differenceS 
between objects 
FIGURE 2 
Once a question has been posed to TEXT, a 
schema must be selected for the response structure 
which will then be used to control the decisions 
involved in deciding what to say when. On the 
basis of the given question, a set of schemas is 
selected as possible structures for the response. 
This set includes those sch~nas associated with 
the given question-type (see Figure 2 above). A 
single schema is selected out of this set on the 
basis of the information available to answer the 
question. 
For example, in response to requests for 
definitions, the constituency schema is selected 
when the relevant knowledge pool contains a rich 
description of the questioned object's sub-classes 
and less information about the object itself. 
When this is not the case, the identification 
schema is used. The test for what kind of 
information is available is a relatively simple 
one. If the questioned object occurs at a higher 
level in the hierarchy than a pre-determined 
level, the constituency schema is used. Note that 
the higher an entity occurs in the hierarchy, the 
less descriptive information is available about 
the entity itself. More information is available 
about its sub-parts since fewer common features 
are associated with entities higher in the 
hierarchy. 
This type of semantic and structural 
interaction means that a different schema may be 
used for answering the same type of question. An 
earlier example showed that the identification 
schema was selected by the TEXT system in response 
to a request for a definition of a ship. In 
response to a request for a definition of a guided 
projectile (shown below), the constituency schema 
is selected since more information is available 
about the sub-classes of the guided projectile 
than about the guided projectile itself. 
(definition GUIDED) 
Schema selected: Constituency 
i) identification 
2) constituency 
3) identification 
4) identification 
5) evidence 
6) evidence 
7) attributive 
I) A guided projectile is a projectile 
that is self-propelled. 2) There are 2 types 
of guided projectiles in the ONR database: 
torpedoes and missiles. 3) The missile has a 
target location in the air or on the earth's 
surface. 4) The torpedo has an underwater 
target location. 5) The missile' s target 
location is indicated by the DB attribute 
DESCRIPTION and the missile' s flight 
capabilities are provided by the DB attribute 
ALTITUDE. 6) The torpedo' s underwater 
capabilities are provided by the DB attributes 
under DEPTH ( for exampl e, 
MAXIMUM OPERATING DEPTH). 7) The guided 
proj ec t~-i e ~as DB attributes 
TIME TO_TARGET & UNITS, HORZ RANGE_& UNITS and 
NAME. 
Once a schema has been selected, it is filled 
by matching the predicates it contains against the 
relevant knowledge pool. The semantics of each 
predicate define the type of information it can 
match in the knowledge pool. The semantics 
defined for TEXT are particular to the database 
query dumain and would have to be redefined if the 
schemas were to be used in another type of system 
(such as a tutorial system, for example). The 
semantics are not particular, however, to the 
domain of the database. When transferring the 
system from one database to another, the predicate 
semantics would not have to be altered. 
A proposition is an instantiated predicate; 
predicate arguments have been filled with values 
from the knowledge base. An instantiation of the 
identification predicate is shown below along with 
its eventual translation. 
Instantiated predicate: 
(identification (OCEAN-ESCORT CRUISER) 
(non-restrictive TRAVEL-MODE SURFACE)) SHIP 
Eventual translation: 
The ocean escort and the cruiser are surface 
ships. 
The schema is filled by stepping through it, 
using the predicate s~nantics to select 
information which matches the predicate arguments. 
In places where alternative predicates occur in 
the schema, all alternatives are matched against 
the relevant knowledge pool producing a set of 
propositions. The focus constraints are used to 
select the most appropriate proposition. 
i17 
The schemas were implemented using a 
formalism similar to an augmented transition 
network (ATN). Taking an arc corresponds to the 
selection of a proposition for the answer. States 
correspond to filled stages of the schema. The 
main difference between the TEXT system 
implementation and a usual ATN, however, is in the 
control of alternatives. Instead of uncontrolled 
backtracking, TEXT uses one state lookahead. From 
a given state, it explores all possible next 
states and chooses among them using a function 
that encodes the focus constraints. This use of 
one state lookahead increases the efficiency of 
the strategic component since it eliminates 
unbounded non-determinism. 
8.0 FOCUSING MECHANISM 
So far, a speaker has been shown to be 
limited in many ways. For example, s/he is 
limited by the goal s/he is trying to achieve in 
the current speech act. TEXT's goal is to answer 
the user's current question. To achieve that 
goal, the speaker has limited his/her scope of 
attention to a set of objects relevant to this 
goal, as represented by global focus or the 
relevant knowledge pool. The speaker is also 
limited by his/her higher-level plan of how to 
achieve the goal. In TEXT, this plan is the 
chosen schema. Within these constraints, however, 
a speaker may still run into the problem of 
deciding what to say next. 
A focusing mechanism is used to provide 
further constraints on what can be said. The 
focus constraints used in TEXT are immediate, 
since they use the most recent proposition 
(corresponding to a sentence in the ~glish 
answer) to constrain the next utterance. Thus, as 
the text is constructed, it is used to constrain 
what can be said next. 
Sidner \[SIDNER 79\] used three pieces of 
information for tracking immediate focus: the 
immediate focus of a sentence (represented by the 
current focus - CF), the elements of a sentence 
~---I~hare potential candidates for a change in 
focus (represented by a potential focus list - 
PFL), and past immediate focY \[re--pr--esent--'-~--6y a 
focus stack). She showed that a speaker has the 
3~6~win-g'~tions from one sentence to the next: 
i) to continue focusing on the same thing, 2) to 
focus on one of the items introduced in the last 
sentence, 3) to return to a previous topic in 
~lich case the focus stack is popped, or 4) to 
focus on an item implicitly related to any of 
these three options. Sidner's work on focusing 
concerned the inter~\[e__tation of anaphora. She 
says nothing about which of these four options is 
preferred over others since in interpretation the 
choice has already been made. 
For generation, ~.~ver, a speaker may have 
to choose between these options at any point, 
given all that s/he wants to say. The speaker may 
be faced with the following choices: 
i) continuing to talk about the same thing 
(current-focus equals current-focus of the 
previous sentence) or starting to talk about 
something introduced in the last sentence 
(current-focus is a member of potential-focus-list 
of the previous sentence) and 2) continuing to 
talk about the same thing (current focus remains 
the same) or returning to a topic of previous 
discussion (current focus is a member of the 
focus-stack). 
When faced with the choice of remaining on 
the same topic or switching to one just 
introduced, I claim a speaker's preference is to 
switch. If the speaker has sanething to say about 
an item just introduced and does not present it 
next, s/he must go to the trouble of 
re-introducing it later on. If s/he does present 
information about the new item first, however, 
s/he can easily continue where s/he left off by 
following Sidner's legal option #3. ~qus, for 
reasons of efficiency, the speaker should shift 
focus to talk about an item just introduced when 
s/he has something to say about it. 
When faced with the choice of continuing to 
talk about the same thing or returning to a 
previous topic of conversation, I claim a 
speaker's preference is to remain on the same 
topic. Having at some point shifted focus to the 
current focus, the speaker has opened a topic for 
conversation. By shifting back to the earlier 
focus, the speaker closes this new topic, implying 
that s/he has nothing more to say about it when in 
fact, s/he does. Therefore, the speaker should 
maintain the current focus when possible in order 
to avoid false implication of a finished topic. 
These two guidelines for changing and 
maintaining focus during the process of generating 
language provide an ordering on the three basic 
legal focus moves that Sidner specifies: 
I. 
2. 
3. 
change focus to member of previous 
potential focus list if possible - 
CF (new sentence) is a member of PFL 
(last sentence) 
maintain focus if possible - 
CF (new sentence) = CF (last sentence) 
return to topic of previous discussion - 
CF (new sentence) is a member of 
focus-stack 
I have not investigated the problem of 
incorporating focus moves to items implicitly 
associated with either current loci, potential 
focus list members, or previous foci into this 
scheme. This remains a topic for future research. 
Even these guidelines, however, do not appear 
to be enough to ensure a connected discourse. 
Although a speaker may decide to focus on a 
specific entity, s/he may want to convey 
information about several properties of that 
entity. S/he will describe related properties of 
the entity before describing other properties. 
118 
Thus, strands of semantic connectivity will occur 
at more than one level of the discourse. 
An example of this phenomenon is given in 
dialogues (A) and (B) below. In both, the 
discourse is focusing on a single entity (the 
balloon), but in (A) properties that must be 
talked about are presented randomly. In (B), a 
related set of properties (color) is discussed 
before the next set (size). (B), as a result, is 
more connected than (A). 
(A) The balloon was red and white striped. 
Because this balloon was designed to carry 
men, it had to be large. It had a silver 
circle at the top to reflect heat. In fact, 
it was larger than any balloon John had ever 
seen. 
(B) The balloon was red and white striped. It 
had a silver circle at the top to reflect 
heat. Because this balloon was designed to 
carry men, it had to be large. In fact, it 
was larger than any balloon John had ever 
seen. 
In the generation process, this phenomenon is 
accounted for by further constraining the choice 
of what to talk about next to the proposition with 
the greatest number of links to the potential 
focus list. 
8.1 Use Of The Focus Constraints 
TEXT uses the legal focus moves identified by 
Sidner by only matching schema predicates against 
propositions which have an argument that can be 
focused in satisfaction of the legal options. 
Thus, the matching process itself is constrained 
by the focus mechanism. The focus preferences 
developed for generation are used to select 
between remaining options. 
These options occur in TEXT when a predicate 
matches more than one piece of information in the 
relevant knowledge pool or when more ~,an one 
alternative in a schema can be satisfied. In such 
cases, the focus guidelines are used to select the 
most appropriate proposition. When options exist, 
all propositions are selected which have as 
focused argument a member of the previous PFL. If 
none exist, then 
whose focused 
current-focus. 
propositions are 
is a member of 
filtering steps 
possibilities to 
proposition with 
all propositions are selected 
argument is the previous 
If none exist, then all 
selected whose focused argument 
the focus-stack. If these 
do not narrow down the 
a single proposition, that 
the greatest number of links to 
the previous PFL is selected for the answer. Tne 
focus and potential focus list of each proposition 
is maintained and passed to the tactical component 
for use in selecting syntactic constructions and 
pronominalization. 
Interaction of the focus constraints with the 
schemas means that although the same schema may be 
selected for different answers, it can be 
instantiated" in different ways. Recall that the 
identification schema was selected in response to 
the question "What is a ship?" and the four 
predicates, identification, evidence, attributive, 
and ~articular-illustrati0n, were instantiated. 
Tne identification schema was also selected in 
response to the question "What is an aircraft 
carrier?", but different predicates were 
instantiated as a result of the focus constraints: 
(definition AIRCRAFT-CARRIER) 
Schema selected: identification 
I) identification 
2) analogy 
3) particular-illustration 
4) amplification 
5) evidence 
i) An aircraft carrier is a surface ship 
with a DISPLACEMENT between 78000 and 80800 
and a LENGTH between 1039 and 1063. 
2) Aircraft carriers have a greater LENGTH 
than all other ships and a " greater 
DISPLACEMENT than most other ships. 3) Mine 
warfare ships, for example, have a 
DISPLACF24ENT of 320 and a LENGTH of 144. 
4) All aircraft carriers in the ONR database 
have REMARKS of 0, FUEL TYPE of BNKR, FLAG of 
BLBL, BEAM of 252, ENDU--I~NCE RANGE of 4000, 
ECONOMIC SPEED of 12, ENDURANCE SPEED of 30 
and PRO~LSION of STMTURGRD. 5)--A ship is 
classified as an aircraft carrier if the 
characters 1 through 2 of its HULL NO are CV. 
9.0 FUTURE DIRECTIONS 
Several possibilities for further development 
of the research described here include i) the use 
of the same strategies for responding to questions 
about attributes, events, and relations as well as 
to questions about entities, 2) investigation of 
strategies needed for responding to questions 
about the system processes (e.g. How is 
manufacturer ' s cost determined?) or system 
capabilities (e.g. Can you handle ellipsis?) , 
3) responding to presuppositional failure as well 
as to direct questions, and 4) the incorporation 
of a user model in the generation process 
(currently TEXT assumes a static casual, naive 
user and gears its responses to this 
characterization). Tnis last feature could be 
used, among other ways, in determining the amount 
of detail required (see \[ MCKEOWN 82\] for 
discussion of the recursive use of the sch~nas). 
119 
10.0 CONCLUSION 
The TEXT system successfully incorporates 
principles of relevancy criteria, discourse 
structure, and focus constraints into a method for 
generating English text of paragraph length. 
Previous work on focus of attention has been 
extended for the task of generation to provide 
constraints on what to say next. Knowledge about 
discourse structure has been encoded into schemas 
that are used to guide the generation process. 
The use of these two interacting mechanisms 
constitutes a departure from earlier generation 
systems. The approach taken in this research is 
that the generation process should not simply 
trace the knowledge representation to produce 
text. Instead, communicative strategies people 
are familiar with are used to effectively convey 
information. This means that the same information 
may be described in different ways on different 
occasions. 
The result is a system which constructs and 
orders a message in response to a given question. 
Although the system was designed to generate 
answers to questions about database structure (a 
feature lacking in most natural language database 
systems), the same techniques and principles could 
be used in other application areas (for example, 
computer assisted instruction systems, expert 
systems, etc.) where generation of language is 
needed. 
~owl~~ 
I would like to thank Aravind Joshi, Bonnie 
Webber, Kathleen McCoy, and Eric Mays for their 
invaluable comments on the style and content of 
this paper. Thanks also goes to Kathleen Mccoy 
and Steven Bossie for their roles in implementing 
portions of the sys~om. 
References 
\[BOSSIE 82\]. Bossie, S., "A tactical model for 
text generation: sentence generation using a 
functional grammar," forthcoming M.S. thesis, 
University of Pennsylvania, Philadelphia, Pa., 
1982. 
\[CHEN 76\]. Chen, P.P.S., "The 
entity-relationship model - towards a unified view 
of data." __ACM Transactions --°n Database Svstems, 
Vol. I, No. I. (1976). 
\[GRIMES 75\]. Grimes, J.E. The Thread of 
Discourse. Mouton, The Hague, Par-~. (1975). 
\[GROSZ 77\]. Grosz, B. J., "The representation and 
use of focus in dialogue understanding." Technical 
note 151, Stanford Research Institute, Menlo Park, 
Ca. (1977). 
\[LEE a\[~ GERRITSEN 78\]. Lee, R.M. end 
R. Gerritsen, "Extended semantics Lot 
generalization hierarchies", in Proceedings of the 
1978 ACM-SIGMOD International Conference on 
Management of Data, Aus£1n, Tex., 1978. 
\[KAY 79\]. Kay, M. "Functional grammar." 
Proceedings of the 5th ;~inual Meetin~ of the 
Berkele Z Ling~\[stl-l~Soc--\[~ty. (1979). 
\[MALHOTRA 75\]. Malhotra, A. "Design criteria for 
a knowledge-based English language system for 
management: an experimental analysis." MAC 
TR-146, MIT, Cambridge, Mass. (1975). 
\[MCCOY 82\]. McCoy, K. F., "Augmenting a database 
knowledge representation for natural language 
generation," in Proc. of the 20th Annual 
Conference of the ~soc-~tion~or Com~utatlo-~ 
Linguistics , Toronto, Canada, 1982. 
\[MCKEOWN 80\]. McKeown, K.R., "Generating 
relevant explanations: natural language responses 
to questions about database structure." in 
Proceedinss of AAAI, Stanford Univ., Stanford, Ca. 
(1980). pp. 306-9. 
\[MCKEOWN 82\]. McKeown, K. R., "Generating natural 
language text in response to questions about 
database structure." Ph.D. dissertation, 
University of Pennsylvania, Philadelphia, Pa. 
1982. 
\[SHEPHERD 26\]. Shepherd, H. R., Tne Fine Art of 
Writinc/, The Macmillan Co., New York, N. Y., 1926. 
\[SIDNER 79\]. Sidner, C.L., "Towards a 
computational theory of definite anaphora 
comprehension in English discourse." Ph.D. 
dissertation, MIT AI Technical Report #TR-537, 
Cambridge, Mass. (1979). 
\[SMITH and SMITH 77\]. Smith, J.M. and 
Smith, D. C.P., "Database abstractions: 
aggregation and generalization." University of 
Utah, ACM Transactions on Database Systems, Vol. 
2, #2, June 1977, pp. 105-33. 
\[TENNANT 79\]. Tennant, H., "Experience with the 
evaluation of natural language question 
answerers." Working paper #18, Univ. of Illinois, 
Urbana-Champaign, Ill. (1979). 
120 
