From: AAAI Technical Report SS-97-02. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved.
REASON:NLP-based Search System for the WWW
Natalia Anikina, Valery Golender, Svetlana Kozhukhina, Leonid Vainer, Bernard Zagatsky
LingoSenseLtd.,
P.O.B. 23153,Jerusalem91231,Israel
valery@actcom.co.il
http://www.lingosense.co.il/ai_symp_paper.html
Abstract
This paper presents REASON,
an NLPsystem
designedfor knowledge-based
informationsearch in
large digital libraries and the WWW.
Unlike
conventionalinformationretrieval methods,our
approachis basedon using the contentof natural
languagetexts. Wehaveelaborateda searching
strategyconsistingof three steps: (1) analysis
natural languagedocuments
and queries resulting in
adequateunderstandingof their content; (2) loading
the extracted informationinto the knowledge
base and
in this wayformingits semanticrepresentation;(3)
query matchingproperconsisting in matchingthe
contentof the queryagainst the content of the input
documentswith the aim of finding relevant
documents.Theadvancedsearchingcapabilities of
REASON
are provided by its mainconstituents: the
LanguageModel, the KnowledgeBase, the
Dictionaryand the Softwarespecially developedfor
content understandingand content-basedinformation
search. Wecharacterizethe particularities of the
Systemand describe the basic components
of the
formallinguistic model,the architectureof the
knowledge
base, the set-up of the Dictionary,andthe
mainmodulesof the System- the moduleof analysis
and the logical inference module.REASON
is an
importantcontributionto the development
of efficient
tools for informationsearch.
Introduction
The WWW
is rapidly expandingand becominga source of
valuableinformation.At the sametime this informationis
poorlyorganized,mixedwith noise and dispersed over the
Intemet. Attemptsto find the informationusing currently
existing search enginessuchas Alta Vista, Lycos,
Infoseek,etc. lead to unsatisfactoryresults in manycases.
Usually,the responseto the searchcontainsa large
amountof documents,manyof whichare totally
irrelevant to the subjectof interest. Onthe other hand,
documents
that are relevant maybe missingif they do not
contain the exact keywords.Refiningthe searchby using
"advanced"
Booleanoperationsusually results in a few, or
evenno, documents.Serious problemsare causedby
homonymy,
ambiguity, synonymy,ellipsis, complex
semanticstructure of natural language.
Let us considersuchrelatively simplequeriesas:
Q1. Supercomputer
marketdata
Q2. Digital users groupin England
Q3.parallelizationtools
It is verydifficult to find documents
relevantto these
queries using traditional keywordsearch. Evenmore
sophisticatedapproachesbasedon different thesaurusesdo
not allowmatchingsemanticrelations betweendifferent
conceptsin queriesandtexts.
In order to solve the above-mentioned
search problems
we designed an NLP-basedsearch system REASON.
This
systemunderstandsfree-text documents,stores their
semanticcontent in the knowledgebase and performs
semanticmatchingduringthe processingof queries. This
paperbriefly describesthe mainprinciples of the System.
General Principles
and System Architecture
REASON
is an NLP-based information search system
whichis capableof:
¯ understandingthe content of natural language
documents
extracted fromlarge information
collections, such as the WWW;
¯ understandingthe content of the users’ query;
¯ building a knowledgebase in whichthe extracted
informationcontainedboth in the documents
and the
queryreceives an adequatesemanticrepresentation;
¯ performinginformation search whichconsists in
matchingthe contentof the queryagainst the content
of the documents
;
¯ finding relevant documents
or giving anotherkind of
response(e.g. answeringquestions).
REASON
does not use the keywordmatchingmethod.
Instead, the Systemapplies moresophisticated matching
techniqueswhichinvolve a numberof reasoning
operations(cf. Haase1995):
¯ establishing equivalenceof wordsand constructions,
¯ expansionof the queryon the basis of associations
betweenconcepts( definitions, "general-specific"
relations andso on),
¯ phrase transformations(based on "cause-result"
connections,temporalandlocal relations,
replacementby paradigmaticallyrelated words),
¯ logical inferences(basedon cross-referencesin the
text, restorationof impliedwords).
Themainconstituent parts of REASON
are: the
LanguageModel,the Dictionary, the Knowledge
Base and
the Software. Thepresent version of REASON
processes
texts in the Englishlanguage.TheSystemsoftwareis
written in C and runs on different platforms.Theinterface
to the Webis implemented
in Java.
Language
Model
Theneedsof natural languageprocessingcall for a
consistent, unambiguous
and computer-oriented
modelof linguistic description. Anatural language,unlike
artificial languages,aboundsin ambiguities,
irregularities, redundancy,
ellipsis. In constructingthe
languagemodelweproceededfrom current linguistic
theories, in particular, semanticsyntax(Tesniere1959),
generativesemantics(Fillmore1969),structural
linguistics (Chomsky
1982,Apresianet al. 1989),
Meaning-TextModel(Melchuk& Pertsov 1987). Wealso
makeuse of the experienceof the "SpeechStatistics"
ResearchGroupheadedby Prof. Piotrowski(Piotrowski
1988).
Ourlanguagemodelincludes four levels: morphological,
lexical, syntactic, discourse.At eachlevel wedistinguish
two planes: the planeof expression(forms)and the plane
of content (meanings,concepts).Thelatter is considered
the semanticcomponentof language.
Morphological Level
In describing English morphology
weapplied the
techniquesuggestedby J.Apresian. For compact
presentation of morphologicalinformationweused charts
of standard paradigmsof morphologicalforms and masks
for non-standardstems. Everywordstemis ascribed a
certain morphological
class on the basis of three charts
for verbsandtwocharts for nounsandadjectives.
Weelaboratedrules for revealingdeepstructure
categories for verbs. Thus,wehavelearnedto determine
the real time and aspect meaningsguidedby the cooccurrenceof adverbialsof time (certain features are
ascribed to each groupof adverbsandprepositionsof
time), grammatical
features of verbs(stative verbs, verbs
of motion).This approachallowsto transfer to utterances
in other languages,wherethe expressionof tense-aspect
relationsis different~
OurModelincludes about 80 patterns whichestablish
correspondence
betweenformand content of verbal tenseaspect forms.
Lexical Level
All the wordsare divided into two kinds: majorand minor
classes. Theformerinclude conceptual(notional verbs
andnouns)andattributive classes (e.g. adjectivesand
adverbsof manner,denotingthe properties of concepts).
Majorwordclasses are opengroupsof words, while minor
classes are closedgroupsconsistingof a limited number
of
words.Majorandminorclasses are treated differently in
text processing.Wehavesingled out about 200 conceptual
wordclasses for verbs and220 - for nouns.Eachword
class is characterized by somecommon
semanticor
pragmaticfeature, e.g. Wordclass 187 - "Sphereof
human
activity" (science, business),204 - "Cessation
Possession"(lose). Wehave singled out about 150
attributive classes, e.g. Wordclass 407- "Color"(white),
480- "Characterof action accordingto intention"
(deliberate).
Wedifferentiate betweenthe namesof properties(color,
size, depth,price) andthe valueof these properties(red,
big, deep, cheap). Thenamesof properties (nouns), their
values (adjectives) andunits of measurement
(nouns)
linked in terms of lexical functions(Melchuk
&Pertsov
1987), whichenablesthe Systemto paraphrasedifferent
statementsand bring themto a canonicalform.
Minorwordclasses play an auxiliary role. Theyinclude
functional words,supra-phrasalclasses, wordsexpressing
modality,probability,the speaker’sattitude, andso on.
Supra-phrasalwordclasses comprisewordsindicating
referenceto the informationcontainedin other parts of the
sentence(also, neither), sentenceconnectorsandsentence
organizers:e.g. "to beginwith"marksthe starting point of
the sentence, "by the way","on the other hand"develop
the theme,"therefore" marksthe conclusion,andso on.
Specialplaceis occupiedbyintensifiers of properties(too
big, quite well), actions(likedvery much),etc.
OurLinguistic Modelreflects logical, extralingual
connectionsandassociations betweenwords. Dueto these
relations, it becomes
possibleto join wordsinto groupings
on the paradigmaticaxis (Miller 1990).
Wedistinguish several types of paradigmaticgroupings:
1. Quasi-synonymicgroups
2. "General-specific"groupings
3. "Whole-part"groupings
4. "Cause-result"groupings
5. Conversives
6. Definitions
7. Lexicalparadigms- "Nests".
Quasi-synonymicgroups include words which are close
in meaning and can be interchangeable in speech or, at
least, in documents pertaining to a certain domain of
science or technology. As an example, we could mention
such Quasi-synonymic groups as Qv3 (show, demonstrate,
present, describe, expose), Qnl0 (developer, author,
inventor).
"General-specific" groupings. Such a grouping consists
of a word with a generic meaning and words with more
specific meanings.For instance, "market" is a generalized
term for more specific "purchase", "sale", "orders",
"supply".
"Whole-part" groupings. The name of the whole object
can be used instead of the names of its components. For
example, "America" can be used instead of USA,Canada;
"processing" can be used instead of "analysis", "synthesis",
and so on.
"Cause-result" groupings. Many concepts, especially
actions and states, are connected by the "Cause-result"
relation. Wehave revealed a numberof such connections,
e.g.:
"produce" -~ "sell" -~ "order"
(If a companyproduces something, it can sell this product
and a customer can order it.)
Special rules are designedin order to reveal the role of
the participants of the situation.
Conversives. Manyverbs are associated on the basis of
conversion. Theyform pairs of concepts (take - give, sell
buy, have - belong). In real speech the statements
organized by these verbs are interchangeable if the first
and the second (or third) actants of the predicate are
reversed:
I gave him a book ~ He took a book from me.
They have a laboratory -~ A laboratory belongs to
them.
The company sold us a computer -~ We bought a
computer from the company.
There are rules stating correspondences betweenactants.
Definitions. Paradigmatic associations between concepts,
based on the above-mentioned groupings, are most
efficiently realized in definitions. Everydefinition states a
certain equivalence of concepts. If definitions are loaded
into the knowledge base, the associations between the
components of the definition help identify statements
expressing the same idea. Wesingled out a number of
types of definitions:
1. Incompletedefinition of the type Ns is Ng(specific is
generic): DECUS
is a worldwideorganization.
2. Completedefinition of the type Ns is Ngthat
(which)... : DECUS
is a worldwideorganization
Information Technologyprofessionals that is
interested in the products, services, and technologies
3.
4.
5.
6.
7.
of Digital EquipmentCorporation and related
vendors.
Incompletedefinition of the type Attribute + Ngis
Ng: A parallel computeris a supercompnter.
Completedefinition of the type Attribute Ngis Ng
that... : A parallel computeris a supercompnterthat is
based on parallelization of sequential programs.
A list of junior concepts- Ngis Nls, N2s, N3s... :
Marketis sale,purchase, orders, supply ...
N1is the sameas N2: Identifier is the sameas
symbol.
N1 is abbreviation of N2: COBOL
is the abbreviation
of Common
Business-Oriented Language.
Nests. One of the most ingenious and useful ways of
grouping words is their organization into lexical
paradigms which we call "Nests". Every Nest contains
words of different parts of speech, but united around one
basic concept (e.g. invest, investment, make (an
investment), investor, etc.). Membersof the Nest may
derived from different root morphemes,e.g. buy, buyer,
customer, purchase, shopping. The main criterion is that
the members of the Nest stand in certain semantic
relations to the head of the Nest. Every Nest is viewedas
a frame and its membersoccupy certain slots in it. Each
slot corresponds to a specific semantic relation with the
head of the Nest or sometimeswith other slot-fillers. The
relations between the Nest memberscan be described in
terms of lexical functions - LF (according to I.Melchuk).
Wehave developed and adjusted the system of LF to the
peculiarities of our LanguageModel. Wedesigned several
types of Nests. Transitive verbs of action, for example,
form a frame which includes 25 slots. In each particular
case someof the slots mayremain empty.
Belowwe illustrate the Nest formedby the verb
"produce":
Slot 1 - the verb of action: produce
Slot2 - nominaactionis: production, manufacture
(LF S0)
Slot 3 - perform the action of Slot 2: carry out (LF
Operl)
Slot 4 - the performerof the action: producer,
manufacturer
Slot 7 - the generalized name ofactant 2: product
Slot 13 - able to performthe action of Slot 1: productive
(LF A0).
Semantic Features. The mechanismof semantic features
(SF) plays an important role in the semantic and syntactic
description of language. Within our Modelevery lexeme (
hence, a concept) is assigned a specific set of SF which
characterize their semantic content and often determine
their syntactic behavior. The co-occurrence of words is
often determined by their semantic agreement.
Wedistinguish about 200 SFfor nouns,180 SFfor verbs,
50SFfor adjectives.
Someexampleswouldsuffice.
SFfor nouns:"Relative" (son), "Economic
organization"
(company),"Inanimate"(house, book), "Transparent"
kindof, a part of), etc.
SFfor verbs: "Causative"(compel),"Possession"(own),
"Movement"
(go), etc.
SFfor adjectives: "Measurable
property"(cold),
"Relationto time"(past), etc.
SFare used in government
modelsto specify the
semanticrestrictions on the choiceof actants. Theyhelp
identify the actants andtheir semanticroles, resolve
ambiguity, reveal second-level modelscommonly
formed
by a limitedcircle of predicatesandso on.
Hereis oneexample
to illustrate the assertion:
Thehomonyms
"raise" (growplants), "raise03" Coring
up), "raise01"(lift), "raise02(gather money),"raise04"
(producean action) are differentiated only dueto the SF
their first andsecondactants.
Syntactic Level
Thebasic syntacticunit is a predicativeconstruction.The
predicate governsthe dependentsas the headof the
construction. Themainformof the predicative
constructionis a simplesentence,but the predicatecan
govern,in its turn, other predicativeconstructions.Among
dependentstructures wedistinguish subordinateclausesof
all kindsandpredicativegroupsformedby verbals. So it is
possible to gradepredicativeconstructionsby the rank of
their dependence
(first, secondor third levels).
At the deepstructure level relations betweenwordsare
viewedin terms of semanticroles, whena dependent("a
Servant")playsa definite semanticrole in relation to the
head("the Master").Thusa predicativeconstruction
representedas a role relation (Allen1995).In our
LanguageModelthe Syntactic Leveldistinguishes about
600semanticroles of the first andsecondlevels (including
subroles).
Hereare someexamplesof roles.
#1
Addressee
#9
Instrument
#35.1 Agentof the action
#35.2 Experiencer
##24.6.aContentof completeinformationat the 2nd
level (that-clause),etc.
Otherparts of speechcan also formrole relations. Oneof
the mostfrequentheadsof a predicativeconstructionis a
predicativeadjective( "Heis afraid of progress").But
treat suchadjectivesas verbswith auxiliary"be", andin
the courseof analysis the combination
"be + adjective"is
reducedto a one-member
verb (Lakoff 1970). Besides,
manynounscan haveservants linked to themby certain
semanticroles. Aspecial place among
potential mastersis
occupiedby prepositionsand conjunctions.
If the servantscanbe predicted,the combinability
of the
masteris describedin terms of models.Wehavesingled
out about 260 patterns of government,or Government
models.Everymodelis a frame.Theslot-fillers are
actants, eachof whichis linked to the masterby a certain
semanticrole. In typical government
modelsthe actants
are describedin termsof word-classes,but in local models,
characteristicof concretepredicates,the SFof actants are
usuallyspecified. Thenumber
of actants varies from1 to
7. For example,the verb "buy"has five actants. In the
followingsentenceall of themare realized:
"Thecustomerbought a computerfor his son from IBM
for 2000dollars".
Act 1 is linked by role #35.1 (Agent). RequiredSF N13
(Legalperson).
Act2 is linked by role #60.1(Objectof Action).
RequiredSF N4(Inanimatenon-abstract thing),
(Animals),N87(Services).
Act3 is linked by role #2 (Beneficiary).Required
N13, N3.
Act4 is linked by role #55 (Anti-Addressee).Required
SF N13.
Act5 is linked by role #12 (Measure).RequiredSF N51
(Money).
Themaster can haveservants the occurrenceand
selection of whichis not determinedby the predicate.
Theirpositioning,structural andsemanticfeatures are less
fixed than those of the actants. Theyare considerednonactant servants. Lexemes
functioningas non-actant
servantsare ascribed reverse modelswhichindicate in
whatrelation they stand in referenceto their master.
Second-levelgovernment
modelsare also classified into
twotypes: actant andnon-actant.For instance, the verbs
"tell" and"inform",besidesfirst-level models,can form
modelson the secondlevel, e.g.:
1) an objectclause linkedby"that" (Hetold us that
livedthere).
2) an objectclause introducedby relative conjunctive
words(Hetold us howhe workedthere).
Discourse
Animportantcomponent
of syntactic and lexical analysis
is processingtextual units larger than a sentence(Allen
1995).Thisis the level of discourse.
Oneof the mostsalient featuresthat characterizetext is
cohesion.OurSystemdeals with different types of
sentencecohesion.Theyincludethe use of substitutes,
correlativeelements,all typesof connectionsignals (like
wordorder, functional words,sentenceorganizersof the
"yet", "besides","to beginwith"types). Semantic
interdependencebetweensentences involves naming
identicalreferents(objects, actions,facts, properties)
different ways.Thus,parts of a broadcontextare knit by
variouskinds of cross-reference.For instance, pronouns
andtheir antecedentsfoundin the precedingparts of the
passageco-refer. OurLanguage
Modelhas special rules
for identifyingsuchcasesof co-reference.
Among
other meansof co-reference wetake into account
the following:
a) the repetition of the samenounwithdifferent articles or
other determiners;
b) the use of synonyms
or other semanticallyclose words
(company
- firm, parallelizer - parallelizationtool);
c) the use of generictermsinstead of morespecific in the
samecontext(vessel - ship, warship,freighter);
d) the use of common
and proper nounsfor denotingthe
samereferent (Washington
- the capital of the USA);
e) the use of semanticallyvagueterms instead of more
concrete denominations(company- Cray Research);
f) the occurrenceof wordswith relative meanings
(sister,
murderer,friend), the full understandingof whichis
possibleonlyif discourseanalysishelps to find their
actual relationships(say, whosesister, the mudererof
whom,a friend to whom).
Thesemanticrepresentationof a sentenceis often
impossiblewithouttaking into considerationthe
representationof other parts of the context.In particular, it
appearsnecessaryto considerthe syntactic structure and
role relations of precedingsentences.Specialrules helpto
restore ellipsis, incomplete
role relations with missing
obligatoryor essentially importantactants (key actants),
identify semanticallyclose syntacticstructures. Proper
discourseanalysis is an importantprerequisitefor natural
text comprehension
and informationsearch.
Dictionary
The main componentsof the LanguageModelare
mirroredin the Dictionaryentries. Ourdictionary
currentlyconsists of about20,000items. Further
expansionof the Dictionarycan be automated.New
lexemescan be addedby analogyif their semanticand
morphological
characteristics coincidewith those of the
previouslyintroducedentries. Anappropriate mechanism
for expandingthe Dictionaryhas beenelaboratedand
tested.
Everyentry contains exhaustivelexicographical
informationabout the word:its semantic,morphological,
syntactic, collocational, phraseological,paradigmatic
characteristics.Eachentry has 20 fields. Hereis part of the
entry"sell":
fl. Thestem: sell
12. Thecodeof the word-class:204
f3. Thecodeof the wordwithin the word-class:7
f4. Thecodeof the typical government
model:4.3.1
f5. Morphologicalclass: 42
f7. Syntactic features: G1, G9, G18
t’8. Semanticfeatures: V6, V86,V19,V53
f9. Grammaticalfeatures: GI2; Tv3
f15. Lexicalfunctions: Lf60= "sellS"
Theverb"sell" has twostems:"sell" and"sold"; both of
themshare the samefields - 2 and3 and constitute one
lexeme.Someother fields maydiffer. Thus,in the entry
"sold"wesee:
f5. 48
f9. G12;Tv3;Passl
In the entry "inform"wesee:
fl0. Paradigmaticconnections: Qvl3; Lv23
In the entry "abolish"wesee:
f13. Style: St3, St4
Forthe verb "delete":
f14. Domainof usage: Cf45
Somecommentson the meaningof the symbolsin the
fields are presentedbelow:
GI: possibleto use absolutely,withoutanyobjects;
G18:possibleto use witha direct object;
G12:usedwith prepositionlessdirect object;
G9:usedonly with adverbialsof place, andnot those of
direction;
Tv3:verb of a dual character(can be either resultative or
durative);
Lf60:"sellS: the verb formsa Nestof type 1;
Passl: the stemformsPassiveof type 1, whenactant 2
becomesthe subject (the left-handdependent);
Qvl3: the lexemebelongsto a quasi-synonymic
group
No13(alongsidewith "say", "report", etc);
Lv23:the verb can governa second-levelmodelof a
certain type;
St3:literarystyle;
St4: style of official documents;
CF45: the domainof usage "Computers".
Lackof space doesnot permitus to dwell on manyother
subtle wordcharacteristicscontainedin the entry.
There are somemorepoints concerningdictionary
entries. Specialentries are compiled
for set expressions.
Fixedand changeableset expressionsare processed
differently (compare
"of course","by bus", on the one
hand,and"vanitybag", "bringup", on the other).
Thedictionary entry also containsfull informationabout
typical and local government
models.
Someentries are formednot for concretewords,but for
representativesof groupsof words:e.g. conceptualword
classes or quasi-synonymic
groups.
Knowledge Base
The System’sKnowledge
Base (KB)serves for
representingthe meaningof natural languagetexts - both
documents
and queries. In our Systemthe term
Knowledge
base is used in a slightly different sencethan
in the theoryof expert systems.TheKBstructure is
specially orientedtowardsextractingfromthe text a
maximum
of knowledgethat is necessaryfor further
searchof documents
on the basis of the query.
SimpleRoleRelation. Thebasic construction of the KBis
a so-called simple Role Relation (simple RR). A simple
RRis a KBrepresentation of a simple sentence in the
formof a tree. Theroot of the tree is the predicateof the
sentence.All nodesof the tree are classified into twotypes
- 0-ranknodeswhichrepresent conceptinstances (objects,
actions, states) and lst-rank nodes whichrepresent the
values of the properties of these conceptinstances. Anarc
from a 0-rank parent node to a 0-rank child node
represents a role the latter performsin relation to the
former. Anarc from a 0-rank parent node to a lst-rank
child noderepresents an instance-property link whichis
nottreatedas a role.
Toillustrate a simpleRR,let us take a natural text
sentence:
S 1: TheCompany
provides customerswith the newest
technology.
For describingRoleRelations, weuse our own
metalanguagein whichthe RRof $I looks as follows:
provide => Company
< lact, #35.1 >, customers
< 2act, #2 >, technology(newest)< 3act, #60.1>
This is interpreted in the followingway:
Thepredicate "provide"- Master- has the following
Servants:
"Company"
as the 1st actant linked by role #35.1
("Agent"),
"customer"as the 2rid actant linked by role #2
("Beneficiary"),
"technology"
as the 3rd actant linked by role #60.1
("Objectof Action");
"technology"is specifiedby a propertyhavingthe value
"new".
Characteristics of the Elements of a Simple Role
Relation. Each node - an element of the RR- is
representedin the KBby a set of characteristics derived
fromthe text in the course of processing. Thus,for each
conceptinstance, there is a KBcorrespondence- a 0-rank
node which is an RR element having about 20
characteristics,e.g.:
¯ the class and the ordinal number,
¯ the morpho-syntacticcharacteristics,
¯ the referenceto the properties of the element,
¯
the role of the givenelementin referenceto the
master.
Likewise,for eachpropertyof the conceptinstance, there
is a KBcorrespondence
- a lst-rank nodewhichis an RR
elementhavingabout30 characteristics, e.g.:
¯ the value of the property (can be expressedby
number,a wordfrom the Dictionaryor a word
unknown
to the Dictionary),
¯ the nameof the property ("price"),
measurement
units associatedwith the property(as,
for example,"dollars"for "price"),
¯ the intensity and degreeof intensity for the property
value,
¯ the numericalcoefficientof fuzzinessusedfor
processingfuzzy values.
ComplexRole Relation. Simple RRsserve as building
material for three morecomplicatedtypes of knowledge
structures whichrepresent any natural languagesentence
in the KB,that is, for Systemof Facts, Task,andRule.
Thesethree types reflect the followingcommunicative
typesof utterances.
Systemof Facts correspondsto narrative sentencesthat
denotefacts, events, situations, etc andcannotbe defined
as Tasksor Rules.
Tasksrefer to sentenceswith hortatorymeaning(e.g.
requests, orders, commands)
that impel manor machineto
performa certain action (Directive)undercertain Starting
Conditions.Directiveis expressedby meansof the
Imperativemood,different formsof modality,questions,
andso on (cf. "Helpme!","I’d like youto help me",
"Couldyouhelp me, please?"; "Whatdid you see at the
exhibition?"- "Tell mewhatyou sawat the exhibition").
Rulesreflect varioustypes of relationships,dependencies,
regularities, objectivelaws, etc that do not impelmanor
systemto action. Rulesare mainlyexpressedby "if- then",
"in order to" andsimilar constructions.
For formaldifferentiation betweenSystemsof Facts,
Tasksand Rules, a numberof linguistic meansare
involved,such as mood,tense, modality,evaluative and
emotionalexpressions,interrogative forms,andothers.
Specificcombinations
of these factors also predetermine
quite definite semanticroles characteristicof eachtype.
Thefollowingsentencesare examplesof the three types
of knowledge
structures respectively:
$2: Convexannouncedthe total investmentprogram
whichpreserves customerinvestments.(Systemof Facts)
$3: As soonas a newissue of the "ComputerMagazine"
appears, find mearticles on supercomputers.
(Task)
$4: If a company
producesanyproduct, it evidently
receivesordersfor this product.(Rule)
Eachstructure is presentedin the formof a complexRR
whichis a tree of simpleRRs.
ASystemof Facts has the followingKBcharacteristics:
the moment
of registration; the author or sourceof
information;the reliability coefficientof the sourceof
information;the referencesto the simpleRRs,each of
thempresentingoneseparatefact of the systemof facts.
In sentence$2, for example,the simple RR1represents
the mainclause of the complexsentence, while the simple
RR2represents its subordinateclause manifestinginterpredicate(2nd-level)role ##80.1.1("DefiningRelative
Clause"):RR1=>RR2<##80.1.1>.
Naturallanguagedocuments
to be searchedare, as a rule,
systemsof facts.
Note: Asimplesentencecorrespondsto a systemof facts
containingonly onefact.
A Taskincludes a Directive andStarting Conditions.The
System’soperation, on the whole,consists in monitoring
the starting conditionsand, on their fulfillment,in
executingthe corresponding
directives.
A Taskhas the followingKBcharacteristics: the moment
of registration; the authoror sourceof information;the
referencesto the RRs,each of thempresentingone
separatedirectiveor starting conditionof the task.
In sentence$3, for example,the simpleRR1represents
the mainclause of the complexsentence, while the simple
RR2represents its subordinateclause manifesting2ndlevel role #4.1.2.2.1.3.1.b("Just after"): RR1=>RR2
< ##4.1.2.2.1.3.1.b>.
Naturallanguagequestionsand queries are particular
cases of Task,with starting conditionsoften omitted,
whichmeans"Now"by default (e.g.: Tell meabout the
supercomputer
exhibition in London).
ARulehas the followingKBcharacteristics: the
moment
of registration; the authoror sourceof
information;the reliability coefficientof the sourceof
information;
the typeof the rule: inferencerule, identity,
cause-effectrelationship,condition,etc; the probabilityof
the antecedent-consequent
relation; the referencesto the
antecedentandconsequentof the rule, eachof them
presentedby a RR.
In sentence$4, for example,the consequentis
representedby the simpleRR1,while the antecedentis
representedby the simpleRR2with role ##30.1.a
("Condition"):
RR1=>RR2< ##30.1.a >.
Thus,anynatural languagesentenceis presented,in a
generalcase, by a complexRR- a tree of simpleRRs,
wherethe root represents the mainclause, the other nodes
representsubordinateclauses, andthe arcs representthe
corresponding
roles.
Note:Wetreat coordinationof clausesas a special case
of subordination.
Special Search-Oriented KBConstructions. Alongside
with the registration of the above-mentionedknowledge
constructions, the Systemalso registers special searchoriented constructions in the KB. They are files of
conceptsandfiles of conceptinstances.
In the file of concepts,eachconcepthas the following
characteristics:the class of the concept;the list of the
elementsof RRswhichinclude the instances of the given
concept,i.e. the invertedlist for the concept;the list of
definitions for the conceptsandthe conceptinstances
whichprovidea powerfultool for the searchof documents.
Main Modules of the System
TheSystem’ssoftwarehas a modularstructure. The
Modulesrun asynchronously, exchangingmessages,
whichallowsto paraUelizetheir operationson a
multiprocessor.
TheSystemconsists of the followingmainmodules:
(a) the moduleof graphicaluser interface,
(b) the moduleof analysis whichforms the KBfrom
natural languagetexts,
(c) the moduleof generation(synthesis) whichproduces
natural languagetexts fromthe KB,
(d) the KB-accessmodulewhichprovidesaccess to the
KBfor analyzingFacts, Tasksand Rules,
(e) the moduleof logical inferencewhichhandles
document
searchingon the basis of a set of equivalent
transformationsover the KB.
Modules(b) and(e) are the mostcomplicatedin
Systemand play a mostimportantrole in the document
search.
Moduleof Analysis. Theprocess of analysis consists of a
numberof steps, each of themprovidingknowledge
for the
final KBrepresentation. The process begins with
morphologicalanalysis, after whichsemantic-syntactic
parsingis performed.Theparsingfalls into twostages.
Thefirst stage accessesthe input text in its surfaceform.
This stage consistsof the steps whichcarry out, in the
appropriateorder, the processingof wordsand word
groupings(set expressions,articles, personalpronouns,
nounconstructions(NNand the like), negations,etc)
attach 0- and1-st rank servantsto their masters- verbs,
predicativeadjectives, nouns.Thesearchingof 0-rank
servantsis conductedaccordingto the master’s
Government
model(for the actant servants) and by means
of special Prepositionalmodels(for the non-actant
prepositionalservants).
Thereare special rules that limit the area of searching.
TheSystemalso makesa canonical reduction of certain
constructionson the basis of equivalenttransformationsof
differenttypes.
Thefirst stage results in the draft variant of parsingat
the level of simpleRRs.
Thesecondstage of parsing, whichdeals with RRsonly,
producesfinal variants of the simpleRRs,establishes 2ndlevel relations (conjunctionalandnon-conjunctional),
forms the final complexRR,computesthe deeptenseaspectcharacteristics, andregisters all the knowledge
obtainedin the KB.
In the processof analysis, the Systemdetects various
typesof conflict situations, e.g.: a servantis linkedto more
than onemaster;a servant is ambiguous;
a predicateis
ambiguous,etc.
Toresolvethe conflicts the Systemexploits
a special inventoryof rules and the mechanism
of
penaltiesandawards.If a conflict cannotbe resolved,the
Systemregisters morethan one variant.
It shouldbe noted that word-sensedisambiguation
is
carried out at various steps includingmorphological
analysis, set expressionandarticle processingand
especially whenusing such powerfulfilters as Government
and Prepositional modelswith their strong semantic
restrictions. In addition, word-orderrules anddomain
filters are applied.
Moduleof Logical Inference for Document
Search. The
moduleof logical inference, together with the KB-access
module,makesall special transformations over the KB
whichare necessaryfor querymatchingon the basis of the
KBrepresentations of the documentand the query. The
techniqueof keywordmatchingis not used at all. Thus,
the one-sentence document "The developers of this
supercomputer
marketa variety of electronic devices"will
be recognizedby our Systemas non-relevant to the query
"Supercomputer
market".
Themodulepossessesthe full inventoryof the System’s
search techniques,bringing theminto play when
necessary. Theuse of these techniquesmaybe both
sequentialandparallel.
Thesearch processconsists of twomainstages:
Stage1. Theexpansionof the input query(i.e. its KB
representation)by use of quasi-synonyms,
definitions,
"General-specific"
relations andconceptualNestrelations
(cf. Vorhees1994).
Theexpansionof the queryenables the Systemto take
into accountnot only the explicit occurrencesof the query
elementsin the RRs,but also the implicit, expandedones.
Dueto the query expansion,the documentD1, for
example,will be recognizedas relevant to the queryQ1,
giventhe set of definitionsE1:
E 1: Saleis a particularcaseof market.
AnExemplarsystemis a supercomputer.
D 1: The sales of high-performance
Exemplarsystems
increasedthreefoldin the last 3 years, to
$6 billion in 1995.
QI: Supercomputermarket.
Stage2. Theuse of the rules of logical inference.
Theyare put into effect if stage 1 is not sufficient for
makingthe final decision. Thelogical inferencerules
allow to matchthe querywith separate pieces of
knowledge
whichare often scattered over the document.
Sometechniquesof this stage can be illustrated by the
followingsearch microsituation, whereE2presents
"encyclopedic"worldknowledge
(definitions andother
natural languagestatements entered into the KB
beforehand),D2is the document,Q2is the query.
E2: Crayproducessupercomputers.
Saleis a particular caseof market.
D2: Crayproductsare widelyused for high-performance
computing.
In the last quarter, the sales totaled $266
million.
Q2: Supercomputermarket.
Theelementsof the query(i.e. its RR)are checkedfor
their co-occurrence
(explicit or implicit) in oneof the
document’s
sentences(i.e. in the corresponding
RR).(It
shouldbe noted that here wemeansuch co-occurrenceof
the queryelementsthat preservesthe structure of the query
RoleRelation).
Therearen’t such sentencesin D2, so the Systemselects
the sentencecontainingonly the predicateof the query"market"(renderedby its implicit occurence"sale"). Then
the previoussentencesof D2are searchedthroughfor
findingthe objectof "sale". Forthis, specialobjectsearchingrules are applied.Theobjectis found- it is
"products".
"Products"is not the secondelementof the query
("supercomputer"),
but it belongsto the groupof the socalled non-terminalwordswhichsuggest further
reasoning.For the non-terminalwords,the conceptual
Nest rules are attached whichactivate the conceptualNest
relations of "products".This leads to generatinga new
inner subqueryin the formof the sentence"Crayproduces
XI", and, throughthe use of the encyclopedicfact "Cray
producessupercomputers",
results in recognizingthe
relevanceof D2to the queryQ2.
Besidesthe rules mentioned,the Systemuses a wide
range of knowledge-based
search techniques, such as
¯ various inference rules basedon different
paradigmaticwordgroupings,
¯ mechanisms
for searching antecedents, including not
only antecedentsfor pronouns,wordsof vague
meaning,
etc., but also for text lacunas,
¯ local and temporal transformations,
¯ "cause-result"transfer, etc.
Thesetechniquesallow, in particular, to realize query
matchingfor the followingsearch microsituation:
E3: Orderis a particular case of market.
Craysells supercomputers.
Franceis in Europe.
D3: Cray Reports the Results of the Fourth Quarter and
the Full Year. In France, orders totaled $66 million in the
quarter comparedwith $47 million for the third quarter of
1995.
Q3: Supercomputer market in Europe.
The query matching process uses, among others, the
inference rules
"produce" --~ "sell" ~ "order"
described in the LanguageModelsection.
Conclusion
At the present stage our Systemhandles documents
containing commonlyused, non-specialized vocabulary.
Besides, we process documentsbelonging to computer
science and related domains. Further expansion of the
Dictionary will enable the Systemto process documents
related to other domains. Experiments with the System
showits efficiency in refining results of the search
performed by traditional search engines on the WWW.
The Web-resident version of the System is currently under
development. REASON
possesses great potentialities for
various practical applications involving man-machine
communication
and artificial intelligence (e.g. expert
systems, machinetranslation, documentsummarization
and filtering, information extraction, educational software,
control of industrial processes).
References
Allen, J. 1995. Natural LanguageUnderstanding.
USA/Canada: The Benjamin/Cummings Publishing
Company,Inc.
Apresian, J.; Boguslavski,I.; Iomdin, L.; Lazurski, A.;
Pertsov, N.; Sannikov, V.; and Zinman, L. 1989.
Linguistic Support of ETAP-2.Moscow:Nauka.
Chomsky,N. 1982. SomeConcepts and Consequences of
the Theory of Governmentand Binding. Cambridge: MIT
Press.
Fillmore, Ch. J. 1969. Towarda ModernTheory of Case.
ModernStudies in English. N.Y.: Prentice Hall: 361-375.
Haase, K. B. 1995. MappingTexts for Information
Extraction. Proceedingsof SIGIR’95.
Lakoff, G. 1970. Irregularity in Syntax. N.Y.: Holt.
Melchuk,I.; and Pertsov, N. 1987. Surface Syntax of
English. Amsterdam/Philadelphia.
Miller, G. 1990. Wordnet: An On-Line Lexical Database.
International Journal of Lexicography3(4).
Piotrowski, P. 1986. Text Processing in the Leningrad
Research Group"SpeechStatistics" - Theory, Results,
Outlook. Literary and Linguistic Computing:Journal of
ALLC1(1).
Tesnlere, L. 1959. Elementsde syntaxe structurale. Pads.
Vorhees, E. M. 1994. Query Expansion Using LexicalSemantic Relations. Proceedingsof S1GIR’94.