KNOWLEDGE BASED QUESTION ANSWERING 
Michael J. Pazzani and Carl Engelman 
The MITRE Corporation 
Bedford, MA 01730 
ABSTRACT 
The natural language database query system 
incorporated in the KNOBS interactive planning 
system comprises a dictionary driven parser, 
APE-II, and script interpreter which yield a 
conceptual dependency conceptualization as a 
representation of the manning of user input. A 
conceptualization pattern matching production 
system then determines and executes a procedure for 
extracting the desired information from the 
database. In contrast to syntax driven Q-A 
systems, e.$., those based on ATH parsers, AFE-II 
ia driven bottom-up by expectations associated with 
word ~eanings. The procesain K of a query is based 
on the contents of several knowledge sources 
including the dictionary entries (partial 
conceptualizations and their expectations), frames 
representing conceptual dependency primitives, 
scripts which contain stereotypical knowledge about 
planning tasks used to infer states enabling or 
resulting from actions, and two production system 
rule bases for the inference of implicit case 
fillers, and for determining the responsive 
database search. The goals of this approach, all 
of which are currently at least partially achieved, 
include utilizing similar representations for 
questions with similar meanings but widely varying 
surface structures, developing a powerful mechanism 
for the disambiguatiou of words with multiple 
meanings and the determination of pronoun 
referents, answering questions which require 
inferences to be understood, and interpreting 
ellipses and unBra--natical utterances. 
THE SETTING 
The KNOBS \[Engelman, 1980\] demonstration 
system is an experimental expert system providing 
consultant services to an Air Force tactical air 
mission planner. The KNOBS database consists of 
several nets of frames, implemented within an 
extension of FRL \[Roberts, 1977\], representing both 
individual and generic classes of targets, 
resources, and planned missions. The KNOBS system 
supports a planner by checking the consistency of 
plan components, enumerating or ranking possible 
choices for plan components, or automatically 
generating a complete plan. Because these 
activities are accomplished by means of rules and 
constraints expressible in English, KNOBS will 
hopefully be a relatively easy system to learn. 
For the same reasons, it is also being considered 
as an aid to train mission planners. The natural 
language subsystem of KNOBS plays several roles 
including those of database query, database update, 
co~uand language, plan definition, and the addition 
or modification of production system rules 
representing domain knowledge. The moat developed 
of these is database query, upon which this paper 
will focus. 
The balance of this paper will first outline 
the use of conceptual dependency and mention some 
prior related work and then describe the several 
knowledge sources and the parts they play in the 
parsing of the input query. Finally, it will 
describe the method of deriving the appropriate 
database search and output response as well as a 
script-based approach to interpretting COmmands. 
USE OF CONCEPTUAL DEPENDENCY 
APE-If utilizes Conceptual Dependency theory 
\[Schank, 1972\] to represent the meaning of 
questions. Once the meaning of a question has been 
found, the question is answered by a rule based 
system whose teats are CD patterns and whose 
actions execute database queries. 
We feel it is important to represent the 
meaning in this manner for several reasons. First, 
the canonical meaning representation enables 
questions which have different surface expressions, 
but the same meanins, to be answered by the same 
mechanikm. This is not only of theoretical 
sisnificance, but is also a practical matter as it 
requires less effort to produce a robust system. 
Because people do not always say precisely 
what they mean, inferences may be required to 
explicate missing information. This inference 
process can also utilize the canonical meaning 
representation. Finally, finding the referent Qf a 
nominal which is modified by a relative clause is, 
in some cases, similar to question answering 
although the syntactic constructions used differ. 
As a result of this similarity, the question 
answering productions can also be used for 
determining the referents of a relative clause. 
The conversation with KNOBS (whose database is 
fictional) in Fig. 1 illustrates these points. 
The first question is represented in the same 
manner as "Does Ramstein have F-4G's?" and would 
be answered by the same rule. The second question, 
73 
USER: Are there F-4G's at Ramatein? 
KNOBS: RAMSTEIN has F-4Ga. 
USER: Can its fighters reach the target? 
KNOBS: F-15e can reach SE50301 from RA~SIEIN. 
F-4Ge and F-dCa can not reach BEb0301 from RA~STEIN. 
USER: Which SCL which are carried by an F-dC contain ECM? 
KNOBS: Sl, S7 and BB. 
F~guve i. A Question Answering Interchange withi, KN08S. 
after resolving the pronominal reference, requires 
an inference to find the location from which the 
F-4G's will be leaving. This inference states that 
if the source of the object of a physical transfer 
is missing, then the source could be the initial 
location of the object. The third question can be 
thought of as two questions: "Which SCL (Standard 
Configuration Load - a predefined weapons package) 
are carried by an F-dC?" and "Which of those 
contain ECM (Electronic Counter Measures - radar 
jamming equipment)?". The first part requires a 
script based inference: In order for an SCL to be 
carried by an aircraft, the aircraft must be 
capable of having the SCL as a part. After the 
first part is answered as a question, the second 
part is answered as a second question to discover 
which contain ECM. 
The system of representation used for nominals 
(or picture producers) differs from that normally 
present in a CD system. Typically, an object such 
as an F-4C would be represented as a picture 
producer with a TYPE case filled by VEHICLE, a 
SUBTYPE case filled by aircraft, and, perhaps, a 
MODEL case filled by F-4C. In KNOBS, the meaning 
representation produced by the parser is F-dC, the 
name of a frame. The set membership of this frame 
is indicated by links to other frames. F-dC is a 
kind of FIGHTER which is a kind of AIRPLANE which 
is an AIRCR~T which is a VEHICLE which is a 
PICTURE PRODUCER. We feel that representing 
nominals in this manner allows a finer degree of 
discrimination than explicitly labeled cases to 
denote a conceptual hierarchy. 
Many of the attributes of objects in the 
database (which are stored as value facets of slots 
in FRL) are represented as kinds of RELATIONS in 
the KNOBS system. For example, the representation 
of "Hahn's Latitude" is (LATITUDE ARGUMENT (HAHN)). 
Note, however, chat the representation of "Hahn's 
aircraft" is (AIRCRAFT LOC (AT PLACE (HAHN))). 
PREVIOUS WORK 
We would like to distinguish the KNOBS natural 
language facility from such familiar natural 
language query systems as LADDER \[Hendrix, 1978\] 
and LUNAR \[Woods, 1972\] in both function and 
method. The functional model of the above systems 
is that of someone with a problem to solve and a 
database containing information useful in its 
solution which he can access via a natural language 
interface. KNOBS, by contrast, integrates the 
natural language capability with multi-faceted 
problem solving support including critiquing and 
Benerating tactical plans. Our approach differs in 
method from these previous systems in its 
bottom-up, dictionary driven parsing which results 
in a canonical representation of the meaning of the 
query, its ability to perform context dependent 
inferences with this representation during question 
answering, and the use of a declarative 
representation of the domain to assist parsin S, 
question answering, plan updating, and inferencing. 
A system similar to APE-If in both its 
diccionarydriven approach to parsins and ice 
direct attack on word sense disambiguation is the 
Word Expert Parser (WEP) \[Small, 1980\]. This 
parser associates a discrimination net with each 
word to guide the meanin 8 selection process. Each 
word in a sentence is a pointer to a coroutine 
called a word expert which cooperates with 
neighboring words to build a meanin S representation 
of the sentences in a bottom-up, i.e., data driven, 
fashion. At each node in the discrimination net a 
multiple-choice test is executed which can query 
the lexical properties or expectations, 
(selectional restrictions \[Katz, 1963\]) of 
neighboring words, or proposed FOCUS, ACTIVITY, and 
DISCOURSE modules. The sense selection process of 
WEP requires that each word know all of the 
contexts in which its senses can occur. For 
example, to find the meaning of "pit", the pit 
expert can ask if a MINING-ACTIVITY, EATING-ACTION, 
CAR-RACINC, or MUSIC-CONCERT-ACTION is active. 
APE-II evolved from APE (A Parsing 
Experiment), a parser used by the DSAM 
(Distributable Script Applying Mechanism) and ACE 
(Academic Counseling Expert) projects at the 
University of Connecticut \[Cullingford, 1982\]. APE 
is based on the CA parser \[Birnbaum, 1981\] with the 
addition of a word sense disambiguation algorithm. 
In CA, word definitions are represented as 
requests, a type of test-action pair. The test 
part of a request can check lexical and semantic 
features of neighboring words; the actions create 
or connect CD structures, and activate or 
deactivate other requests. 
The method available to select the appropriate 
meaning of a word in CA is to use the test part of 
separate requests to examine the meanings of other 
words and co build a meaning representation as 
function of this local context. For example, if 
the objeet of "serve" is a food, the meaning is 
"bring to"; if the object is a ball, the meaning is 
"hit toward". This method works well for selecting 
a sense of a word which has expectations. However, 
some words have no expectations and the intended 
sense is the one that is expected. For example, 
the proper sense of "ball" in "John kicked the 
ball." and "John attended the ball." is the sense 
which the central action expects. 
The word definitions of APE are also 
represented as requests. A special concept called a 
VEL is used to represent the set of possible 
meanings of a word. When searching for a concept 
which has certain semantic features, an expectation 
can select one or more senses from a VEL and 
74 
discard those that are not appropriate. In 
addition, APE can use expectations from a 
contextual knowledge source such as a script 
applier to select a word sense. Each script is 
augmented with parser executable expectations 
called named requests. For example, aCa certain 
point in understanding a restaurant story, leaving 
• tip for the waiter is expected. The parser is 
then given a named request which could help 
disambiguate the words "leave" and "tip", should 
they appear. 
APE-II 
A word definition in APE-II consists of the 
set of all of its senses. Each sense contains • 
concept, i.e., • partial CD structure which 
expresses the meaning of this sense, and a set of 
conceptual and lexical expectatious. 
A conceptual expectation instructs the parser 
to look for a concept in s certain relative 
position which meets a selectional restriction. 
The expectation also contains a selectional 
preference, a more specific, preferred category for 
the expected concept (cf. \[Wilkg, 1972\]). If such 
a concept is found, the expectation contains 
information on how it can be combined with the 
concept which initiated the expectation. A lexical 
expectation instructs the parser to look for a 
certain word and add a new, favored sense to it. 
This process is useful for predicting the function 
of a prepositiou \[Reisbeck, 1976\]. The definition 
of a pronoun utilizes a context and focus mechanism 
co find the set of possible referents which agree 
with it in number and gender. THE PRONOUN IS THEN 
TREATED LIKE A WORD WITH MULTIPLE SENSES. The 
definitions of the words "fly", "eat" and "A/C" are 
shown in Fig. 2. 
The definition of "A/C" states that it means 
AIRCRAFT or AIR-CONDITIONER. APE-If uses 
selectional restrictions to choose the proper sense 
of "A/C" in the question "What A/C can fly from 
Hahn?". On the other hand, in the sentence "Send 4 
A/C to BE70701.", APE-II utilizes the facts that 
the OCA script is active, and that sending aircraft 
to a target is a scene of that script, Co determine 
that "A/C" means AIRCRAFT. In the question "What 
is an A/C?", APE-II uses a weaker argument to 
resolve the potential ambiguity. It utilizes the 
fact that AIRCRAFT is an object that can perform a 
role in the OCA script, while an AIR-CONDITIONER 
cannot. 
The definition of "fly" states that it means 
FLY which is a kind of physical transfer. The 
expectations associated with fly state the 
actor of the sentence (i.e., a concept which 
precedes the action in a d~clarative sentence, 
follows "by" in a passive sentence, or appears in 
various places in questions, etc.) is expected to 
be an AIRCRAFT in which case it is the OBJECT of 
FLY or is expected to be a BIRD in which case it is 
both the ACTOR and the OBJECT of the physical 
transfer. This is the expectation which can select 
the intended sense of "A/C". If the word "~o" 
75 
appears, it might serve the function of indicating 
the filler of the TO case of FLY. The word "from" 
is given a similar definition, which would fill the 
FROM case with the object of the preposition which 
:should be a PICTURE-PRODUCER but is preferred to be 
a LOCATION. 
The definition of "eat" contains an 
expectation with s selectional preference which 
indicates that the object is preferred to be food. 
This preference serves another purpose also. The 
object will be converted to a food if possible. 
For example, if the object were "chicken" then this 
conversion would assert that it is a dead and 
cooked chicken. 
We vili first discuss the parsing process as 
if sentences could be parsed in isolation and then 
explain how it is augmented to account for context. 
The simplified parsing process consists of adding 
the senses of each word to an active memory, 
considering the expectations, and removin E concepts 
(senses) which are not connected to other concepts. 
Word sense disambiguation and the resolution 
of pronominal references are achieved by several 
mechanisms. Selectional restrictions can be 
helpful to resolve m-biguities. For example, many 
actions require an animate actor. If there are 
several choices for the actor, the inanimate ones 
will be weeded out. Conversely, if there are 
several choices for the main action, and the actor 
has been established as animate, then ~hose actions 
which require an inanimate actor will be discarded. 
Selectional preferences are used in addition to 
selectioual restrictions. For example, if "eat" 
has an object which is a pronoun whose possible 
referents are a food and a coin, the food will be 
preferred and the coin discarded as a possible 
referent. 
A conflict resolution mechanism is invoked if 
more than one concept satisfies the restrictions 
and preferences. This consists of using 
"conceptual constraints" to determine if the CD 
structure which would be built is plausible. These 
constraints are predicates associated with CD 
primitives. For example, the locational specifier 
INSIDE has a constraint which states that the 
contents must be smaller than the container. 
The disnmbiguation process can make use of the 
knowledge structures which represent stereotypical 
domain information. The conflict resolution 
algorithm also determines if the CD structure which 
would be built refers to a scene in an active 
script and prefers to build this type of 
conceptualization. At the end of the parse, if 
there is an ambiguous nominal, the possibilities 
are matched against the roles of the active 
scripts. Nominals which can be a script role are 
preferred. 
A planned extension to the parsing algorithm 
consists of augmenting the definition of a word 
sense with information about whether it is an 
uncommonly used sense, and the contexts in which i¢ 
could be used (see \[Charniak, 1981\]). Only some 
senses will be added to the active memory and if 
(DEF-WORD A/C (SENSE (AIRCRAFT)) 
(SENSE (AIR-CONDITIONER))) 
(DEF-WOED EAT (SENSE \[EAT ACTOR (NIL) 
OBJECT (NIL) 
TO (*INSIDE I PLACE (~STOMACN ~ PART (NIL\] 
EXPECTATIONS (\[ IF (IN-ACT-SPOT #ANI}~TE) 
THEN ((SLOTS (TO PLACE PART) 
(ACTOR\] 
\[IF (IN-OBj-SPOT *PP*) 
PREFER (#~OOO) 
THEN ((SLOTS (OBJECT\])) 
\[DEF-WORD FLY (SENSE (FLY OBJECT (N~L) 
ACTOR (NIL) 
INSTE~NT ($IIY) 
TO (*PROX* PLACE (NIL)) 
FROM (*PROX* PLACE (NIL))) 
EXPECTATIONS (\[IF (IN-ACT-SPOT AIRCRAFT) 
THEN ((SLOTS (OSJECT))) ~ 
ELSE (IF (IN-ACT-SPOT BIRD) 
THEN ((SLOTS (ACTOR) (OBJECT)\]) 
LEXICAL-EXPECTATIONS ((TO (MAKE-DEF (OB-PEEP ~ppw) 
(TO PLACE) 
(*~.oc*))) 
(FROM (MAKE-DEF (OS-Pg~P *PP*) 
(FROH PLACE) 
(*LOt*)))))) 
I Figure 2. APE-\[I Dictionary Definitions. 
none of those concepts can be connected, other 
senses will be added. A similar mechanism can be 
used for potential pronoun referents, organizing 
concepts according to implicit or explicit focus in 
addition to their location in active or open focus 
spaces (see \[Grosz, 1977\]). 
Another extension to APE-II will be the 
incorporation of a mechanism similar to the named 
requests of APE. However, because the expectations 
of APE-II are in a declarative format, it is hoped 
that these requests can be generated from the 
causally linked scenes of the script. 
QUESTION ANSWERING 
After the meaning of a question has been 
represented, the question is answered by means of 
pattern-invoked rules. Typically, the pattern 
matching process binds variables to the major 
nominals in a question conceptualization. The 
referents of these nominals are used in executing a 
database query which finds the answer to the user's 
question. Although the question conceptualization 
and the answer could be used to generate a natural 
language response \[Goldman, 1975\], the current 
response facility merely substitutes the answer and 
referents in a canned response procedure associated 
with each question answering rule. 
The question answering rules are organized 
according to the context in which they are 
appropriate, i.e., the conversational script 
\[Lehnert, 1978\], and according to the primitive of 
the conceptualization and the "path to the focus" 
of the question. The path to the focus of a 
question is considered to be the path of conceptual 
cases which leads to the subconcept in question. 
A question answering production is displayed 
in Fig. 3. It is a default pattern designed to 
answer questions about which objects are at a 
location. This pattern is used to answer the 
question "~hat fighters do the airbasee in West 
Gerlmny have?". In this example, the pattern 
variables &LOC is bound to the meaning 
representation of "the airbases in West Germany" 
and &OBJECT is bound to the meaning representation 
of "fighters". The action is then executed and the 
referent of &OBJECT is found to be (FIGHTER) and 
the referent of &LOC is found to be (HAHN SEMBACH 
BITBURG). The fighters at each of these locations 
is found and the variable ANSWER is bound to the 
value of MAPPAIR: 
((HAHN . (F-4C F-15)) (SEMBACH . NIL) 
(BITBURG . (F-~ F-15))). 
The response facet of the question answering 
production reformats the results of the action to 
merse locations with the same set of objects. The 
answer "There are none at Sembach. Hahn and 
Bitburg have F-4Cs and F-15s." is printed on 
successive iteratione of PMAPC. 
The production in Fig. 3 is used to answer 
most questions about objects aC a location. It 
invokes a general function which finds the subset 
of ~he parts of a location which belong to a 
certain class. The OCA (offensive counter air) 
script used by the KNOBS system contains a more 
specific pattern for answering question about the 
defenses of a location. This production is used to 
answer the question "What SAMe are at BE70701?". 
The action of this production executes a procedure 
which finds the subset of the surface to air 
missiles whose range is greater than the distance 
to the location. 
76 
(DEF-Q-PAT PAT (*EXISTS w OBJECT &OBJECT 
LOt (*PIOX* PLACE &LOt)) 
ACTION {MAPPAIR (FIND-REFEEEMTS &LOt) 
(FUNCTION (LAMBDA (LOt) 
(MAPCONC (FII;D-LZFERZNTS &OBJECT) 
(FUNCTION (LAMBDA (TYPE)' 
(FIND-OEJECTS-AT LOC TYPE\] 
RESPONSE \[PMAPC (MEEGEPAIRS ANSI~lt) 
(FUNCTION (LA~SDA (LOt ITZMS) 
(CO~D ((NULL II7~S) 
(MSG "There ere none se " 
(~aMZ LOC) ".')) 
(TIII~J~-PERSON? "have" LOC) 
(~U~ ZTZMS) m.N\] 
q-l, OCUS (o~zcT Is-A)\] 
Ftoure 3. A OuestHon Answertno Production. 
In addition to executing a database query, the 
action of a rule can racureively invoke other 
queJCion answering rules. For example, to answer 
the question '*Row many airbasaJ have F-At'e?", a 
general rule converts the conceptualization of the 
question to that of '~hich airbaees have F-At°e? " 
and counts the result of answering the larger. The 
question answering rules can also be used to find 
the referent of complex nominals such as "the 
airbases which have F-AC'e". The path to the focus 
of the "question" is indicated by the conceptual 
case of the relative pronoun. 
INFERENCE 
when important roles are not filled in a 
concept, "conceptual completion" inferences are 
required to infer the fillers of conceptual cases. 
Our conceptual completion inferences are expressed 
as rules represented and organized in a manner 
analogous to question answering rules. The path to 
the focus of a conceptual completion inference ie 
the conceptual case which it is intended co 
explioate. Conceptual completion inferences are 
run only when necessary, i.e., when required by the 
pattern m4tcher to enable a question answering 
pattern (or even another inference pattern) to 
match successfully, 
An example conceptual completion inference is 
illustrated in FiE. 4. It is designed to infer the 
missing source of a physical transfer. The pattern 
binds the variable &OBJECT co the filler of the 
OBJECT role and thq action executes a function 
which looks at the LOCATION case of &OBJECT or 
checks the database for the known location of the 
referent of &OBJECT. This inference would not be 
used in processin E the question "Which aircraft at 
Ramstein could reach the target from Hahn?" because 
the source has been explicitly stated. It would be 
used, on the other hand, in processing the 
question, "Which aircraft at Ramstein can reach the 
target?". Its effect would be to fill the FROM 
slot of the question conceptualization with 
RAMSTEIN. 
77 
(DEF-IHFERZNCE PAT (*PT~S* OBJECT &OBJECT) 
ACTION (F~MD-LOCATION &OBJECT) 
I}IlP~BJKNCB (FROM)) 
Ftgure 4; A Concelat Completion Inference. 
If a question answering production cannot be 
found to respond to a question, and the question 
refers Co a scene in an active script, causal 
inferences are used CO find an answerable question 
vhich can be constructed as a state or action 
~upliad by the original question. These inferences 
are represented by causal links \[CullinKford, 1978\] 
which connect the lCltel and actions of a 
stereotypical situation. The causal links used for 
this type of inference are RESULT (actions can 
result in state changes), ENABLE (states can enable 
action), and EESULT-ENA3LE (an action results in a 
state which enables an action). This last 
inference is so coumon that it is given a special 
link. In soma cases, the intermediate state is 
unimportant or unknown. In addition to causal 
links, temporal links are also represented to 
reason about the sequencing of actions. 
The causal inference process consists of 
locating a script paCtern of an active script which 
represents the scene of the script referred to by a 
question. The pattern matchfnE algorithm assures 
that the constants ~n the pattern are a super-class 
of the constants in the conceptual hierarchy of FRL 
frames. The variables in script patterns are the 
script roles which represent the common objects and 
actors of the script. The binding of script roles 
to subconcepts of a question conceptualization is 
subject to the recursive matching of patterns which 
indicate the common features of the roles. (This 
will be explained in more detail in the section on 
interactive script instantiation.) After the scene 
referenced by the user question is identified, a 
new question concept is constructed by substituting 
role bindings into patterus representing states or 
actions linked to the identified scene. 
Two script patterns from the OCA script are 
illustrated in Fig. 5. The script pattern named 
(DZF-SCRIPT-PAT 
(DEF-SCRIYT-PAT 
NAME At-FLY-TO-TARGET 
PAT (*PTRANS* OBJECT &OCA:AIRCRAFT 
TO (*FROX* PLACE &0CA:TARGET) 
FROM (*PROX* PLACE &OCA:AIRHASE)) 
SCRIPT OCA 
AFTER At-HIT-TARGET 
RESULT-ENABLE At-HIT-TARGET 
RESULT At-OVER-TARGET) 
NAME AC-HIT-TARCET 
PAT (*PROPEL* ACTOR &OCA:AIRCRAFT 
TO (*LOCSPEC* PLACE &0CA:TARGET) 
OBJECT &OCA:SCL) 
SCRIPT OCA 
RESULT TARGET-IS-DESTROYED 
AFTER At-FLY-BACK) 
Figure 5. Definitions of Script Patterns, 
AC-FLY-TO-TARCET matches the meaning of sentences 
which refer to the aircraft flying to the target 
from an airbase. It results in the aircraft being 
over the target which enables the aircraft to 
attack the target. The script pattern 
At-HIT-TARGET represents the propelling of a weapon 
toward the target. It results in the destruction of 
the target, and is followed by the aircraft flying 
back Co the airbase. 
The knowledge represented by these script 
patterns is needed to answer the question "What 
aircraft at Hahn can strike BE70701?". The answer 
produced by KNOBS, "Y-15s can reach BE70701 from 
Hahn.", requires a causal inference and a concept 
completion inference. The first step in producing 
this answer is to represent the meaning of the 
sentence. The conceptualization produced by APE-If 
is shown in Fig. 6a. A search for a question 
answering pattern to answer this fails, so causal 
inferences are tried. The question concept is 
identified Co he the AC-HIT-TARGET scene of the 0CA 
script, and the scene which RESULT-ENABLEs it, 
AC-FLY-TO-TARGET is instantiafied. This new 
question conceptualization is displayed in Fig 6b. 
A question answering pattern whose focus is (OBJECT 
IS-A) is found which could match the inferred 
question (Fig. 6c). To enable this pattern to match 
the inferred question, the FROM case must be 
inferred. This is accomplished by a concept 
completion inference which produces the complete 
conceptualization shown in Fig. 6d. Finally, the 
action and response of the question answering are 
executed to calculate and print ~n answer. 
INTERACTIVE SCRIPT INSTANTIATION 
The script patterns which describe the 
relationships among the scenes of a situation are 
also used by the KNOBS system to guide a 
conversation about that domain. The conversation 
with KNOBS in Fig. 7 illustrates the entering of 
plan components by interactively insCantiating 
script patterns. 
The first user sentence instantiaces two 
script patterns (the flying of aircraft, and the 
striking of a target) and binds the script roles: 
TARGET Co BE70501, WING to 109TFW, AIRCRAFT-NUMBER 
to 4, and TIME-OVER-TARGET to 0900. KNOB~ asks the 
user to select the AIRCRAFT. Because the user 
replied with a question whose answer is an 
aircraft, KNOBS asks if the user would like would 
like to use chat aircraft am a component of the 
developing plan. This is accomplished by a rule 
that is activated when KNOBS asks the user to 
specify a plan component. The interpretation of the 
user s negative answer is handled by s rule 
activated when KNOBS asks a yes-no question. KNOBS 
checks the consistency of the user's answer and 
explains a constrainc which has failed. Then, the 
user corrects this problem, and KNOBS processes the 
extra information supplied by matching the meaning 
of the user's input to a script pattern. 
(*PROPEL* TO (*PROX* PLACE (BE70701)) 
ACTOR (AIRCRAFT IS-A (*?*) 
LOC (AT PLACE (HAHN))) OBJECT (NIL) 
MODE (*POTENTIAL*)) 
Figure 6a. The Meaning Representation of 
"What ~ircraft at Hahn can strike BE70701?" 
(*PTRANS ~ OBJECT (AIRCRAFT IS-A (*?*) 
LOt (AT PLACE (HAHN))) 
TO (*PROX* PLACE (BE70701)) 
FROM (*PROX* PLACE (alL)) 
, MODE (*POTENTIAL*)) 
Fioure 6b. The Conceot Inferred from 6a. 
(What aircraft at Hahn can go to BE70701?) 
(*~I~S* TO (*PROX* PLACE &TARGET) 
FROM (*PROX* PLACE &AIRBASE) 
OBJECT &AIRCRAFT) 
Fiqure 6c. A Question Answering Pattern 
which could Match 6b. 
(*PTRANS* OBJECT (AIRCP.AFT IS-A (*?~) 
LOt (AT PLACE (HAHN))) TO (*PROX* PLACE (BET0701)) 
FROM (*PROX ~ PLACE (HAHN)) 
MODE (*POTENTIALS)) 
Fiqure 6d. The Complete Question Conceptualization after 
Inferring the Source. 
78 
USER: 
KNOBS: 
USER: 
KNOBS: 
USEI: 
EHOBS: 
USER: 
Send 4 aircraft from the Logcfv co sc:iks SE7050L at 0900. 
Whac aircraft do you vane to use7 
What alrcrafc are in the I09TI~T 
The I09TFW has F-4Cs. 
WouLd you Like to use F-4Cs for the aircraft? 
NO, F-4Gs. 
The 10~r~ does nOC co~tain F-4Gs. 
F17 the P-4Gs out of the 126TFW st Eamscsia. 
Ftaure 7. A Conversation with ~OSS. 
(DEF-ROLE-PAT PAT (AIRCR£YT L0C (*pgOXt PLACE &OCA:AI/t3ASE) 
PART (PART ARGUMENT &OCA:WING)) 
SCB.IPT-NOLZ &OCA: AIRCRAFT 
SCIIPT OCA) 
Figure 8. A Script Role Pattern. 
A script role can be bound by matching against 
patterns associated with other script roles in 
addition to matching against script patterns. Fig. 
8 shows a role pattern associated with the script 
role AIRCL~YT. This pattern serves two purposes: 
to prevent bindings to the script role vhichwould 
not make sense (i.e., the object which plays the 
AIRCRAFT role ~st be an aircraft) and to 
recursively bind other script roles to attached 
concepts. In this exemple, the AIRBASE or the ~NC 
could be attached to the AIRCRAFT concept, e.g., 
"F-4Cs from Hahn" or "F-dCa in the 126TFW". 
The interactive script interpreter is an 
alternative to the menu system provided by KNOBS 
for the entering of important components of a plan 
Co be checked for consistency. KNOBS also provides 
a means of automatically finishing the creation of 
a consistent plan. This can allow an experienced 
mission planner to enter a plan by typing one or 
two sentences and hitting a key which tells KNOBS 
co choose the unspecified components. 
TRANSFERRING DOMAINS 
To demonstrate their domain independence, the 
KNOBS System and APE-II have been provided with 
knowledge bases to plan and answer questions about 
naval "show of flag" missions. This version of 
KNOBS also uses FRL as a database language. 
A large portion of the question answering 
capability was directly applicable for a number of 
reasons. First of all, dictionary entries for 
frames are constructed automatically when they 
appear in a user query. The definitions of the 
attributes (slots) of a frame which are represented 
as RELATIONs are also constructed when needed. The 
definitions of many common words such as "be", 
"have", "a", "of", etc., would be useful in 
understanding questions in any domain. The 
question answering productions and concept 
completion inferences are separated into default 
and domain specific categories. Many of the simple 
but common queries are handled by default patterns. 
For example, "Which airbases have fighters?" and 
"What ports have cruisers?" are answered by the 
same default pattern. Currently, the Navy version 
of KNOBS has 3 domain specific question answering 
patterns, compared to 22 in the Air Force version. 
(There are 46 default patterns.) The most 
important knowledge structure missing in the Navy 
domain is the scripts which are needed to perform 
causal inferences and dialog directed planning. 
Therefore, the system can answer the question "What 
weapons does the Nimitz have?", but can't answer 
'~ihat weapons does the NimiCz carry?". 
CONCLUSION 
We have argued that the processing of natural 
languaae database queries should be driven by the 
meaning of the input, as determined primarily by 
the emaninss of the constituent words. The 
zuechanisms provided for word sense selection and 
for the inference of missing meaning elements 
utilize a variety of knowledge sources. It is 
believed Chat this approach will prove more general 
and extensible than those based chiefly on the 
surface structure of the natural language query. 
ACKNOWLZDGENEI~ S 
We would like to thank Tom Fawcett, Bud 
Frawley, Frank Jernigan, and Ethan Scarl for their 
CO1vementS. 
This work was supported by USAF Electronics 
System Division under Air Force contract 
F19628-82-C-0001. 

REFERENCES 

Birnbaum, L., and Selfridge, M., "Conceptual 
Analysis," in Ins%de Artificial Intelli2ence: Five 
Proera~# Plus Miniatures. Schank, R., Riesbeck, C. 
K. (was), Lawrence Erlbaum Associates, Nillsdale, 
NJ, 1981. 

Charniak, E., "Six Topics in Search of a Parser: An 
Overview of AI Language Research," in Proceedings of 
the 7th International Joint Conference on 
Artificial Intelli2ence, Vancouver, 1981. 

Cullingford, R., "Script Application: Computer 
Understanding of Newspaper Stories," Research 
Report I16, Department of Computer Science, Yale 
University, 1978. 

CullinKford, R. and Pazzani, M., "Word Meaning 
Selection in Multimodule Language-Processing 
Systems," TR-82-13, EE&CS Dept., University of 
Connecticut, 1982. 

Engelman, C., Scarl, E., and Berg, C., "Interactive 
Frame Instantiatlon," in Proc. of The Firs~ Annual 
Confere~c~ on Artifi;~al Intelligen;~, Stanford, 
1980. 

Goldman, N., "Conceptual Generation," in Conceptual 
Information Processing. Schank, R. (ed), 
Ninth-Holland Publishing Company, 1975. 

Grosz, B., "The Representation and Use of Focus in 
Dialog Understanding," SRI Technical Note 151, 
1977. 

Hendrix, G. G., Sacerdoti, E. D., Sagalowicz, D., 
and Slocum, J., '*Developing a Natural Language 
Interface to Complex Data." Association for 
Computing Machinery Transactions on Database 
Systems. Volume 3, Number 2, June 1978. 

Katz, J. S. and Fodor, J. A., "The Structure of 
Semantic Theory," Language. 39, 1963. 
Lehnert, W., Th_...ee Process of Question Answering. A 
Comvuter Simulation of Cognition. Lawrence Erlbaum 
Associates, Inc., 1978. 

Reisbeck, C., and Schank, R. Co, "Comprehension by 
Computer: Expectation Based Analysis of Sentences 
in Context," Research Report #78, Department of 
Computer Science Yale University, 1976. 

Roberts, R. Bruce, and Goldstein, Ira P., "The FRL 
Manual," MIT AI Lab. Memo 409, September 1977. 

Schank, R., "Conceptual Dependency: A Theory of 
Natural Language Understanding," Co2nitive 
PsvcholoxT. Vol. 3, No. 4, 1972. 

Small, S., "Word Expert Parsing: A Theory of 
Distributed Word-Based Natural Language 
Understanding," TR-954, University of Maryland 
1980. 

Wilks, Y., Grammar. Meaning ~ The Machine Analysis 
of LanguaRe. London, 1972. 

Woods, W. A., Kaplan, R. M., and Nash-Webber, B., 
"The Lunar Sciences Natural Language Information 
System." BBN Report 2378, Bole, Beranek, and 
Newman Inc., Cambridge, MA, 1972. 
