From: AAAI Technical Report SS-93-07. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.
AnIntelligent Interface to a DatabaseSystem*
Alfredo
Fernandez-Valmayor
(*)
Carlos
Villarubia
(**)
Manuel Buenaga
(*) Dep. Inform:~tica - Fac. Fisicas - UniversidadComplutensede Madrid
28040 Madrid, Spain.
valmayor@fis.ucm.es - mbuenaga@dia.ucm.es
(**) Dep. Informfitica - E.U. Inforrmttica - Universidadde Castilla-La Mancha
13004Ciudad Real, Spain.
carlos@eui.uucp
Abstract
In this work, we describe the architecture of an
intelligent interface that improvesthe effectiveness of
full text retrieval methods through the semantic
interpretation of user’s queries in natural language
(NL). This interface comprises a user-expert module
that integrates a dynamic model of human memory
with a NL parser. This paper concentrates on the
problem of the elaboration of index patterns out of
specific cases or it/stances. The structure of the
dynamic memoryof eases and parsing techniques are
also discussed.
IntrodUction.
Manyresearchers have focused on the study of human
memoryprocesses to find a modelto improve the way
information is stored and retrieved from computers
(Foltz, 1991). They believe that better understanding
of how human memory indexes and searches
information would facilitate the developmentof new
and more effective methods of building databases.
Our approach deviates from this. Wetry to use this
better understanding of how human memoryworks,
not to build newand moreefficient databases, but to
build tools that will make possibly to use more
effectively pre-existent ones. The basis for this
improvementis the construction of an intelligent
interface. An interface that integrates a model of
humanmemorywith a natural language (NL) parser.
In somedatabase systems (DBS)the user, if he is not
an experienced one, can choose to interact with the
* Acknowledgment: This work is supported in
part by CICYTgrant TIC91-0217-C02-02:
"Arquitectura para Interfaces en LenguajeNatural con
Modeladode Usuario" and CICYTgrant TIC92-9958:
"Metodologiade disefio de herramientas de desarrollo
para sistemas expertos basados en casts"
45
computer through a humanintermediary, which helps
him to find the information he wants. Interacting
with a humanexpert has the additional advantage that
the user does not need to learn any formal language;
however, even with humanintermediaries information
retrieval of textual databases may have a very poor
performance. As an example, Blair and Maron(1985)
report on a precision index of 80%and a recall index
of less than 20%in a legal information retrieval
system.
Several researchers have discussed the possibility of
integrating an expert module in a DBSto automate
the role of the humanexpert (Borgman,Belkin, Croft,
Lesk, & Landauer, 1988). This has been our
approach. The interface we present here would
eliminate in most cases the need for this human
expert. In this interface a user-expert module
improves
the effectiveness of the retrieval proces
s
through the semantic interpretation of user’s naturallanguage query formulations. For that reason, NL
processing capabilities are at the core of the system.
In the user-expert moduleNLprocessing capabilities
are founded on memorybased parsing techniques; an
anticipated improvement of conceptual parsers
01iesbeck, 1986; Riesbeck & Schank, 1989).
In this paper, we begin with an analysis of the
process of information retrieval.
This will be
followed by an overview of our system and a
discussion of the main issues motivated by the
development of the prototype: the knowledge
representation and knowledgeacquisition techniques
used to modelexpertise; the structure and indexing of
cases in the memorymodel and finally the memory
based parsing techniques based on this model.
The system we have had in mind when designing our
prototype is a full-text retrieval system. A system
designed to search a file of unstructured, or partially
unstructured,
textual information (FerngndezValmayor& Villarrubia, 1991). Information retrieval
(IR) fromthis kind of systemis to someextent more
complex
thanretrieval fromstructureddatabases,since
retrieval cues can be matchedby manywordsin the
documents.
Avariety of methodsexist to recover information
from textual databases (Salton & McGilI, 1983).
Booleanexpressions of terms used in the documents
are the most simple approach. Other more
sophisticated methodsassociate vectors of weighted
terms with documents(Salton, 1989). In these later
methods,documentsare represented as vectors with
weighted components. User’s queries are also
representedas vectors in the samespace, therefore
distance betweenvectors representing queries, and
documentscan be computed.The documentsretrieved
wouldbe those with a distance less than a threshold
quantityas specifiedbythe user.
The effectiveness of a retrieval system can be
evaluatedin termsof the so called recall andprecision
indexes. Recall index measuresthe proportion of
relevant documentsout of all relevant documentsin
the database, while precision index measuresthe
proportionof the recovereddocuments
that are found
to be relevantto the user’s needs.In practice, recall
andprecisionindexestend to vary inversely;i.e., very
specific queryformulationsretrieve fewnon-relevant
items but also relatively few relevant ones. Query
formulationscan be altered to reachthe desiredrecall
and precision levels, usingrecall-enhancingdevices;
e.g., term truncation, and precision-enhancing
devices;e.g., termweighting(Salton, 1986).
System overview
The system we are working on is an interface to
existing databases. Accordingly,in designing our
system, we mainly deal with the the strategic
elaborationandverification stages of the information
retrieval process. In the first stage, the user can
express his information needs in natural language.
The user-expert modulewill translate these input
queries into a network of conceptual dependency
structures (CDs)that later it will use to generate
optimizedretrieval cues andto makethe mapping
into
the specific querylanguageof the accesseddatabase.
Oncethe user has queriedthe systemandretrieved the
information, he can inform to the system on the
appropriatenessof the retrieveddocuments.
Thetwo maincomponentsof the user-expert module
are a dynamicknowledgebase and a NLparser. We
conceive this knowledgebase as a modelof human
memorythat incorporates knowledge about the
database’sdomainand aboutthe contextof individual
users.
The way in which the user-expert module lays
betweenthe users and the DBSis depicted in the
diagram. Fromthe diagram, wecan see that if the
user is an expert he can interact directly with the
DBS.If the user is not an expert, then he can express
his questions in natural language.TheNLinterface
parses the user querybuildinga conceptualdependency
46
structure (CD)that capturesthe meaning
of the input.
Parsing of user’s input is based on prior memory
structures. In the knowledgebase, the long term
memory
has at its top level the dictionary used by
the conceptual parser (Dyer, 1983; Birnbaum
Selfridge, 1981). Theshort term memory,also in the
knowledgebase, is a list of CDstructures under
construction that represents the user-computer
interaction duringthe session. TheCDstructures in
the short term memory
represent both the meaningof
the input alreadyseen, andthe expectationsthat will
guide the parser to translate next input and to
optimizethe queries.
Beforeit can be used, the systemneedsto be trained.
Thedifference betweentraining and using the system
is in the way humanscommunicatewith it. To be
sure that initially the system makesthe correct
connections betweenwordsand concepts, the system
manager,or instructor, will give the system the
appropriated information in the form of CD
structures. Thesestructures will be incorporatedinto
the long term memoryshaping up the initial
dictionary and relations. At any time, if the system
managerdetects that the systemis performingpoorly,
he can give the system additional information to
reinforce proper CDstructures and the connections
betweenthem.In this way,the correct links in the
long term memorystructure can be partially
controlled.
Knowledgerepresentation and
knowledge acquisition.
Thefirst issue whendesigninga knowledge
base is to
specify the formalism in which the part of the
universe that is to be representedin the knowledge
base is to be described. ConceptualDependency
(CD)
structures are the basic formalismwehaveselected to
represent information in our model. The key idea
behindCDtheory is that meaningcan be represented
in a canonical, language-freemanner(Schank,1972).
In our implementation a CDis composed of an
atomicheadconceptanda list of attribute-value (or
role-filler) pairs wherevaluesare againCDstructures.
Complexevents, concepts, or objects, can be
representedby meansof a networkof CDstructures,
connectedbydifferent types of links, suchas, causal,
temporalor intentional (Pazzani,1988).
For example, we can connect a CD structure
representingthe chemicalelementmercuryas a heavy
metal used on someindustrial processes, with other
CDstructure representingthe environmentalproblems
causedby the uncontrolleddisposal of heavymetals.
In this manner, the user-expert module can
hypothesize that a search of information about
mercury, must include a search of documentsabout
contaminationcausedby heavymetals.
In the knowledge
base, general concepts(or patterns)
are representedas a packed-CD
structure (as opposed
to a networkof CDstructures). These structures
represent high order memory
conceptualizations and
USERS
]
I
N.L.INTERFACE
KNOWLEDGE
BASE
MAPPINGFUNCTIONTO
SPECIFIC QUERYLANGUAGES
QUERY
LANGUAGE
DATA BASE SYSTEM
are generally knownas MemoryOrganization Packets
(MOP). MOPsare organized as a hierarchical and
modifiable network that index individual cases or
items of information (Schank, 1982; Kolodner1983).
Other characteristic of our systemis that most of the
knowledge in it must be acquired using its own
mechanismsand not pre-stored or hand-tailored. That
means that the system considers all external
information as cases (instances with no variables) and
that the system itself must generate using its own
learning mechanismsall the patterns that index those
casgs.
Inductive learning is the basic mechanismthe system
uses to create these new patterns. To learn new
patterns, our prototype uses the empiricist algorithm
making an input pattern more general or more
specific, in response to feedback about howwell that
pattern explains or classifies an instance. To makea
pattern more specific, our system indexes under the
same pattern (MOP-father),all the instances or cases
that are compatible with this MOP-father. A new
MOPis created from these events when someof them
share one or morefeatures that are not present in the
MOP father.
This new MOPis indexed as a
specialization of the MOP-father,and the events, used
to create it, are indexed under it (Kolodner, 1983;
Lebowitz, 1983).
47
To generalize its input, our program starts by
replacing constants by variables (Fernandez-Valmayor
& Fernandez Chamizo, 1992). SBL-vars (for
Similarity Based Learning) are used to generalize
input instances. These variables are created by the
system when two CDstructures have a commonslot
with values that are different, but with somepossible
commoncharacteristics. For example:
The two CDsbelow are two possible representations
of the chemical elements lead and mercury.
(Following the usual convention we represent CD’s
head concepts in capital letters and attribute’s names
in non-capital letters; PP stands for picture producer
following Shank’s nomenclature and SBL-variables
are written with namesstarting with char "/,").
(PP
(PP
type
is-a
INANIMATE-OBJECT
CHEMICAL-ELEMENT
name
MERCURY)
type
is-a
INANIMATE-OBJECT
CHEMICAL-ELEMENT
symbol
group
density
Hg
METAL
HIGH
symbol
group
density
Pb
METAL
HIGH
name LEAD)
Using the generalization process described above our
systemwill obtainthe followingMOP-pattern:
(PP t y pe INANIMATE-OBJECT
is-a
symbol
/,SYMBOL
/~ IS-A
METAL
group
density
HIGH
name /,NAME)
In this pattern, the SBL-varLNAME
can matchany
valuein the slot ’name’,but the SBL-var
LIS-Aonly
will matchvalues in the slot ’is-a’ withany head
conceptandattribute value pairs ’group-METAL’
and
’density-HIGH’.
Thus,the systemcanuse this pattern
to indexinformation
aboutall highdensitymetals.
Tocreatenewpatterns,our systemuses anothertype
of variables; EBL-vars(for Explanation Based
Learning).Thesevariables represent not merely
conjunctive
generalization
of severalinstancesbutthe
structural relation betweenthe componentsof a
unique instance. For example: The CDbelow
describesa possible episodeof water-contamination
causedby the disposalof mercury
waste
(ACTION
type
actor
object
from
located-in
PTRANS
CHEMICAL-WASTE-LTD.
MERCURY
FACTORY
FORESTLAND
manufacturer-of
BATTERIES
owned-by
CHEMICAL-WASTE-LTD.
ALMADEN
to
C4:IOJND located-in
date 1-5-8g
result
CONTAMINATION
type
WATERCONTAMINATION
agent
MERCURY)
TheEBL
processwill obtainthe followingpattern(in
this pattern EBL-vars
are markedwith the symbol
,,?,,):
(ACTION
type PTRANS
?ACTOR
actor
object
,"K)6JECT
FACTORY located-in
f r om
FORESTLAND
manufacturer-of
BATTERIES
owned-by
?ACTOR
located-in
ALMADEN
to
date 1-5-89
result
CONTAMINATION
WATER-CONTAMINATION
type
.’?OBJECT)
agent
causes contaminationcomesfrom. Specific domain
knowledge
is includedin this processto deal with
attributes (or roles) with special meaningin the
domain.
Finally,oursystemhasprocessesto changeits ’focus
of attention’ or ’view point’ about a given
information.
In the examplesabove we have used to represent
mercurythe CD:
(PP type INANIMATE-OBJECT
is-a CHEMICAL-ELEMENT symbol
group
density
name MERCURY)
Where
the "focusof attention"is the CD-head
concept
’PP’. Changing the "focus of attention" to
’MERCURY’,
the systemobtains the following CD:
(MERCURYname-of
PP
type NANIMATE-OBJECT
is-a CHEMICAL-ELEMENT symbol Hg
group
METAL
density HIGH)
This CDrepresents the sameinformationthat the
formerbut now,the focus is not a generic object
(Picture Producer); the focus is MERCURY.
This
processof changingthe "focusof attention"of a CD
hasthree important
consequences:
First, it makespossibly to matchCDsthat in other
case will not be comparable;
e.g., it makespossibly
to match the CDabove ’with the CD(MERCURY
<empty
list of attribute-valuepairs>)that appearin
the slots of the CDthat describes the watercontamination
case andas a consequence
to inheritthe
attribute-value
pairsof the former.
Second,these new’viewpoints’ can be re-indexedin
long termmemory
makingnewentries at the concept
dictionaryat the root of the memory.
For example,
changingthe ’focus of attention’ of the CDthat
describesmercury
as a PPwill createbetweenothers
the followingnewentries:
(MERCURY name-of
Withthis patternthe systemcapturesthe idea that the
actorof the action,that resultin watercontamination,
is the ownerof the factoryfromwhichthe objectthat
48
Pig
METAL
HIGH
PP)
(Hg
symbol-of
CHEMICAL-ELEMENT)
(METAL
group-of
CHEMICAL-ELEMENT)
In this manner,the dictionary at the root of the
memory
will compriseall the termsusedin the CDs
givento the system.
Finally, whenit is combinedwith the process of
variabilization, changingthe ’focus of attention’
resultsin patternsthatexpressdifferentabstractions
or
conceptualizations (although applied to the same
information). For example, if in the watercontamination case we change the attention from
ACTION to CONTAMINATION
we obtain the
followingCDfor the sameinformation:
(CONTAMINATION
result-of
PTRANS
ACTION
t yp e
CHEMICAL-WASTE-LTD.
actor
obiect
MERCURY
from
FACTORY
located-in FORESTLAND
manufacturer-of BATTERIES
owned-by
CHEMICAL-WASTE-LTD.
to
located-in
ALMADEN
date 1-5-89
type WATER-CONTAMINATION
agent MERCURY)
WhenEBLand SBLprocesses are applied to this CD
the followingpattern can be generated:
(CONTAMINATION result-of
ACTION
t y pe
date
type
agent
PTRANS
actor
/,ACTOR
object
?AGENT
from
~,~
manufacturer-of
BATTERIES
to
located-in
ALMADEN
/,DATE
WATER-CONTAMINATION
?AGENT)
This pattern expresses the abstraction that the
contamination
is causedby an agentthat is the object
translated from places wherebatteries are madeto
ground located in Almaden(Fernandez-Valmayor
FernandezChamizo,1992)
User-expert
module implementation
The NC prototype
Thecurrent implementation
of the user-expert module
is based on NC, a prototype written in Allegro
Common
Lisp. This prototype was first developedto
test memory
organization and learning algorithms
(Fernandez-Valmayor, 1990). In NC, learning
algorithmsare used to create the abstractions that
makethe discrimination network that indexes all
input information. Our systemuses NCto index, by
meansof a networkof patterns, the training cases and
the casesthat result of translatinguser’s queriesinto
CDsstructures.
Our working hypothesis is that some basic but
essential linguistic aspects of humanmemory
can be
modeledusing twosimple guide-lines:
49
¯ Memoryis a schematic process, in which new
informationis integrated into already established
structureof patterns(Barlett, 1932).
¯ There is a small set of processes responsible for
maintaining
this structure.
Memorystructure and organization
Froma structural perspectivememory
is divided into
long term memoryand short term memory.The long
termmemory
is a sequenceof trees, andthe list of the
roots of these trees constitutesits root. Asthe system
incorporates newstructures to long-term memory,
this list is reorganizedaddingto it the CDheadvalues
that still are not in it. Theprocessof changingthe
’viewpoint’ also addsnewentries to memory’s
root.
Thenodesof the trees are the patterns that the system
creates from input CDstructures. The links that
connectthe trees’ nodeshavea label that expressesthe
difference betweena child-nodeandits father anda
frequencyvalue; this frequencyvalue is augmented
each time the nodeis successfullyaccessedduring a
searchoperation.
Theleaves of the trees are the input CDstructures,
these CDstructures can be the outcomeof the
translation of the user’s input, or the CDstructuresby
the systemmanagerwhentraining the system.
Short term memory
is the sequenceof CDstructures
that result fromthe translation of user input during
the session. At the end of the session the short term
memoryis subsumedinto the long term memory.
Unlike the short term memory,which experiences
rapid and shifting changes, the long term memory,
thoughupdatedat the end of each session, is a more
stable representation of the domainand user’s
peculiarities.
Memory based parsing.
The informationretrieval problemis deeply related
with the problemof natural languageunderstanding.
In the past years researchersin the IR field havebeen
arguing that to build the next generation of
informationretrieval systems, weneed systemsable
to extract the semantic information from both
documents
anduser’s queries, and that the matchingof
queries and documentsmustbe doneat this semantic
level (VanRijsbergen,1987).
At the sametimeover the past years researchersin the
NLarea havedevelopeda variety of techniques for
natural languageanalysis (Gazdar & Melish, 1989;
Schank & Riesbeck, 1981; Riesbeck & Schank,
1989).Underlying
these techniquesthere are different
approaches to NLprocessing: declarative versus
procedural, or semantic versus pure syntactically
drivenparsing. However,all these approacheshavein
common
that they are basedon a search processplus a
pattern matchinganalysis. Whatis different is the
data structure they searchandthe kindof patternsthey
try to match. Thus in memorybased parsing,
understanding
is consideredas the processof searching
memoryfor related language structures (Schank,
SHORTTERM MEORY
LONG TERM MEMORY
~r=5¢[~
fr=6
I’~
fr=9
~ ........
_~1~
I~
.~
....
frecue;~T--A~Lr of times the node is succesfully accessed
gradedlist of expectations(alternative patterns) for the
present input word. With this purpose, we define two
functions, the ’weight’ function and the ’activation’
function.
The parsing algorithm
To develop a parsing algorithm based on our memory
modelwe have to take into account three different but
deeplyrelated issues:
¯ Howtraditional parsing algorithms behave and what
are the principles uponthey operate.
¯ Howto infer from examples the structure of the
language.
¯ Howto deal with the inherent ambiguity of natural
language.
1980, 1982; Martin, 1989). Memorybased parsing
can be also considered as a continuation of the work
initially developedin conceptual analyzers and script
appliers.
In conceptual analysis,
words are linked to
conceptualizations
in a dictionary.
These
conceptualizations are structures that represent the
meaning of the word and they have associated
procedural actions knownas ’demons’ or ’requests’.
Conceptualizations not only have semantic meaning
but also precise syntactic knowledge;for examplein
the conceptualization associated with the word"say"
(MTRANS
actor
object
from
X
X
X
to
X)
First issue. It is claimed (Tomita, 1986) that the
syntax of natural languagecan be realized as a context
free grammar. Moreover, semantic grammarsare also
context-free phrase structure grammars in which
semantics and syntaxis is encoded in form of
productions. Many general context-free parsing
algorithms, able to handle arbitrary context-free
grammars, have been formulated. Amongthem, the
most well-knownis Earley’s algorithm. All practical
algorithms have some similarity
with Earley’s
algorithm. General parsing algorithms are like
Earley’s algorithm, in that they all construct wellformedsub-string tables (Tomita, 1986). At any point
in its execution the parser can avoid to make
redundant analysis by looking for phrases already
recognized in its table of well-formed sub-strings.
Whenall the phrases of a given type are stored in the
we will have in the slot ’actor’ the demon:"expect to
have already hear of the humanwho is performing
this act", and in the slot TO’the demon:"expect that
the humanreceiving the messagewill be the object of
the preposition ’to’ " (Dyer, 1983).
In our approach, we try to use a more integrated and
less procedural control structure than the requestsstack structure used in McELI(Birnbaum& Selfridge,
1981) or the working memoryused in DYPAR
(Dyer,
1983). In our system, the system’s dictionary is the
root of the long term memory,the list of trees’ roots;
and wordsin the input text are the primary cues to the
information stored under them. The strength of
associations betweenwords and the information stored
under its corresponding entry are different and from
them the parser builds the short term memoryas a
50
table there is no longerneedto use the grammar.
We conceive our long term memory as a
discrimination networkof patterns equivalent to a
table of well-formed
sub-strings.Thesepatternsplay a
role comparableto that of the "items’ list" that
constitute the table in Earley’s algorithm (Aho
&Ullman,
1977). In our case, ’items’ are the patterns
indexedundereach root’s memory
entry, andthe long
term memoryis used by the parsing algorithm in a
waysimilar to that of ’traditional’ parsingalgorithms,
that is weuse themto alternate top-downprediction
with bottom-up production reduction (Pereira
Shieber, 1987).
Secondissue. The problemof learning a language
froma training set of samplesentencesis the problem
of grammaticalinference (Angluin& Smith, 1983).
In our system initial learning happens whenthe
instructor presents the systeman initial set of CD
structures. ACDis a positive instanceto be learned.
ACDis a tree structure (with labeled branches)that
weconsideras equivalentto the derivationtree of the
input sentence. Thus, the system must learn the
languagefrom structural presentation, a technique
used by pattern recognition researchers to aid
grammatical inference. As we have seen above,
parsing algorithms only need to use the grammarto
build the well-formed
sub-string table, so our problem
is not to infer the grammar
but to built the equivalent
of the well-formedsub-string table. Thatis whatthe
implementedlearning mechanisms
do.
Third issue. Usually natural languagegrammarsare
ambiguous;a sentence can have multiple parses. In
the case wherethere exists two or morepossible
parses, it is not acceptablefor a practical natural
languageparser to produceonly onearbitrary parse.
All possible parses should be producedand stored
somewhere
for later disambiguation.Wheninferring a
grammarwe have another source of ambiguity: the
samelanguagemaybe generatedby several different
grammars.Thus, the question ’which grammarwill
the systemlearn?’ arise.
Aswehaveseen whatour systemdoes is to build the
equivalentof the parsinglist of itemsneededto parse
the input sentence.So if there are differentunderlying
grammarsin the training CDsinstances the wellformedsub-string table will correspondwith more
than one grammar.Therefore, whensearching the
memoryit will be founded expectations that
correspondnot only to different parses in the same
grammar
but to parses in different grammars.
Whatfollows is an outline of howthe parsing of an
input proceeds:
1- Systeminitializes short term memory
2- It searchesfor the first wordin the input in the
operational memory.It makesactive the patterns
underit by putting themin the short term memory.
3- Theactivated patterns generateexpectationsabout
wordsthat wouldbe seen next in the input. These
51
expectations will be tested against the patterns
activated by the followingwordinput. Activepatterns
in short term memory not compatibles with
followingwordsin the input are discharged.
4- If it doesn’tfind the word,it leavesa holefor that
word in the short term memoryand proceeds with
next input word.
Conclusions
and future work.
In this paper we have seen that a humanmemory
model is not a good candidate to replace actual
databasemodels. Its potential lies in its merit to
modela humanexpert that helps the user to formulate
better queries.
Natural languageprocessingcapabilities are at the
core of the systemdescribed here. Thefocus of our
paper has been to develop further memorybased
parsing techniques. If wecomparethe techniqueswe
use in our system with previous work, wecan say
that what wehave madeis to substitute the static
dictionary (whichguidesthe parsing) and muchof its
hand-tailored procedural knowledgeby a dynamic
structureanda set of genericprocesses.
Learningpatterns of well-formedsub-strings from
positive exampleshas been the other focal point of
our paper. The representation schemeand learning
mechanisms
describedhere can be consideredclose to
those of Winston’slearning blocksworldconcepts,as
they are describedin Michalski,Carbonell&Micheli
(1983). TheCDnetworkused in our system and the
semantic networkused by Winstonare similar, but
in our systemthe networkof conceptsandrelations is
morehomogeneous
and it is processedin a less handcrafted way.Thegeneral process that in our system
changes the ’point of view’ over a CDcan be
consideredrelated to the negative-satellitelinks used
by Wiston,but in our case this is not a particular
information introduced by the user, but a general
process that expands the learning power of the
system.
Other question that has been arisen is the need to
comparethe representational powerof CDsstructures
with logic formulas. External links betweenCDsand
slot namescan be thought as two-argument
predicates
but in our system wehave also incorporated logic
operators(’or’, ’and’,’not’) as links that are processed
in a special mannerby functionslike the functionthat
changesthe focusof attention.
Textualanalysis of accurately retrieved information
can be the next goal to improvethe context in which
user questions and commentsmust be interpreted.
Evaluation of text-analysis technologies has
demonstrated that text-analysis techniques,
incorporating natural language-processing,are more
effective than traditional information-retrieval
techniquesbasedon statistical classification when
applicationsrequire structured representationsof the
information present in texts (Lehnert & Sundheim,
1991)
LanguageText. Cognitive Science vol 7, pp.l-40.
Michalsky, R.S., Carbonell, J.G., Mitchell, T.M.
(1983). MachineLearning:an Artificial Intelligence
Approach. Morgan Kaufmann.
Martin,C.E.(1989). Case-basod Parsing in Riesbeck,
C.K. &Schank, R.C. Inside Case-BasedReasoning.
Hillsdale, NJ. LawrenceEflbaum Associates.
Pazzani, M.,(1988). Learning Causal Relationships:
AnIntegration of Empirical and Explanation-Based
learning Methods. PhD. Dissertation, Technical
Report UCLA-A/-88-10.
Pereira F.C.N. & Shieber S.M. (1987). Prolog and
Natural-LanguageAnalysis. CSLIlecture notes;
no.10
Riesbeck, C.K., (1986). FromConceptual analyzer
Direct MemoryAccess Parsing: An Overview. In
Advancesin Cognitive Science, Sharkey N.E. (Ed.).
John Wiley & Sons.
References
Aho, A.V. & Uliman, J.D. (1977). Principles of
Compiler Design. Addison Wesley.
Angluin, D. &Smith, C.H. (1983). Inductive
Inference: Theory and Methods, ComputingSurveys,
Vol. 15,No.3, pp. 237-269.
Barlett, F.C., (1932). Remembering.A Study in
Experimental and Social Psychology. Cambridge
University Press. (2a. reimpresi6n, 1977).
Bimbaum,L. & Selfridge, M.,(1981). Conceptual
analysis of natural language. En Schank, R.C. &
Riesbcck C.K. ,(Ed.),1981. Inside Computer
Understanding, LawrenceErlbaumAssociates, Inc.,
NewJersey.
Borgman,C.L., Belkin, N.J., Croft, W.B., Lesk,
M.E., Landauer, T.K. (1988). Retrieval systems for
the information seeker: Canthe role of the
intermexliaty be automated?Panel Session Human
Factors in Computing Systems, CHI’ 88 Conference
Proceedings (Washington, D.C.), ACM,NewYork,
May 1988, 51-53.
Riesbeck, C.K. & Schank, R.C., (1989). Inside CaseBased Reasoning. Hillsdale, NJ. Lawrence Erlbaum
Associates.
Salton, G. (1986) Anotherlook at automatic textrea’ieval systems. CommunicationsofACM,Vol 29,
No.7, 648-656.
Salton, G. (1989) Automatic Text Processing: The
Transformation, Analysis, and Retrieval of
Information by Computer. Reading, MA:AddisonWesley.
Blair, D.C., & Maron,M.E. (1985). An evaluation
retrieval effectivenessfor a full-text document
retrieval
system. Communications of ACM,28, 289-299.
Dyer, M.,(1983). In Depth Understanding. MITPress.
Fern~mdez-Valmayor,
A. (1990). Diseflo de una Base
de Conocimientos Dimimicay su Aplicacidn en un
EntomoEducativo. Doctoral thesis. Universidad
Complutense Madrid.
Ferruindez-Valmayor, A. & Fern~dez Chamizo, C.
(1992). Educationaland ResearchUtilization of
Dynamic KnowledgeBase. Computers & Education,
Vol.18, No. 1-3, pp. 51-61.
Fern~dez-Valmayor,A. & Villarrubia, C. (1991).
Propuesta de Sistema Unificado de Consulta de Bases
de Datos documentalespara el Instituto de la Mediana
y Pequefla EmpresaIndustrial, IMPI. Inform Ttcnico.
Departamentode !nform~tica, Universidadde CastillaLa Mancha,Ciudad Real.
Foltz, P.W. (1991). Models of HumanMemoryand
ComputerInformation Retrieval: Similar Approaches
to Similar Problems. ICS Tech Report 91-03.
University of Colorado.
Gazdar, G. & Mellish, C., (1989). Natural Language
Processing in Lisp. Addison-Wesley.
KolodnerJ. L.,(1983). MaintainingOrganization in
Dynamic Long-Term Memory, Cognitive Science,
Vol.7, pp.243-280.
KoiodnerJ.L., (1983). "Reconstructive Memory:
ComputerModel". Cognitive Science, Vol.7, pp.281328.
l.,¢hnert, W. &Sundheim,B. (1991). A Performance
Evaluation of Text-Analysis Technologies. A/
Magazine, Fall 1991, pp. 81-94.
Salton, G., &McGill, M. J. (1983). Introduction to
ModernInformation Retrieval. McGraw-Hill:New
York.
Schank, R.C., (1982). Dynamic Memory:A theory
of remindingand learning in computersand people.
CambridgeUniversity Press.
Schank, R.C., (1980). Languageand Memory.
Cognitive Science, 4, 243-284
Schank, R.C., (1972). "Conceptual Dependency:
Theory of Natural LanguageUnderstanding",
Cognitive Psychology. 1972, 552-631.
Schank, R.C. & Riesbeck C.K.. Ed. (1981). Inside
Computer Understanding. Lawrence Erlbaum
Associates, Inc., NewJersey.
Tomita, M. (1986). Efficient Parsing for Natural
Language:A Fast algorithm for Practical Systems.
Kluwer AcademicPublishers
Van Rijsbergen, C.J. (1987). A new theoretical
framework for information retrieval. SIGIR FORUM
Vol. 21, No 1-2 Autum86/Winter 87
Lebowitz, M., (1983). Generalization from Natural
52