perm filename PAPER.DOC[4,KMC]1 blob
sn#049382 filedate 1973-06-14 generic text, type T, neo UTF8
␈↓ ↓-IDIOLECTIC LANGUAGE-ANALYSIS FOR
␈↓␈↓ mUNDERSTANDING DOCTOR-PATIENT DIALOGUES*
␈↓ ↓>Horace Enea and Kenneth Mark Colby
␈↓ ↓↑Department of Computer Science
␈↓ αAStanford University
␈↓ αDStanford, California
␈↓---------------------------------------------------------
␈↓*␈α This␈α research␈α is␈α supported␈α by␈α Grant␈α PHS␈α MH␈α 06645-12␈α from
the␈α⊂National␈α⊃Institute␈α⊂of␈α⊂Mental␈α⊃Health,␈α⊂by␈α⊂(in␈α⊃part)␈α⊂Research
Scientist␈α≤Award␈α≤(No.␈α≤ 1-K05-K14,433)␈α≤from␈α≤the␈α≤National
Institute␈α⊃of␈α⊂Mental␈α⊃Health␈α⊂to␈α⊃the␈α⊂second␈α⊃author␈α⊂and␈α⊃(in␈α⊂part)
by␈α⊃the␈α⊃Advanced␈α⊃Research␈α∩Projects␈α⊃Agency␈α⊃of␈α⊃the␈α∩Office␈α⊃of
the␈α Secretary␈α of␈α Defense(SD-183).
␈↓---------------------------------------------------------
␈↓␈↓ α\␈↓
ABSTRACT␈↓
␈↓ A␈α∂programming␈α∂language␈α⊂is␈α∂described␈α∂which␈α∂is␈α⊂designed␈α∂to
simplify␈α↔the␈α⊗construction␈α↔of␈α⊗computer␈α↔programs␈α↔to␈α⊗analyze
English.␈α⊃This␈α⊃system␈α⊂attempts␈α⊃to␈α⊃merge␈α⊂the␈α⊃best␈α⊃features␈α⊂of
pattern␈α≤matchers␈α≤and␈α≥the␈α≤phrase␈α≤structure␈α≥approach␈α≤to
language␈α⊂analysis.␈α⊃ Several␈α⊂practival␈α⊃problems␈α⊂which␈α⊃occur␈α⊂in
dealing␈α with␈α such␈α a␈α system␈α are␈α described.
␈↓ α)␈↓
INTRODUCTION␈↓
␈↓ Why␈α∃is␈α∃it␈α∃so␈α∃difficult␈α∃for␈α∃machines␈α∃to␈α∃understand
natural␈α→language?␈α_Perhaps␈α→it␈α→is␈α_because␈α→machines␈α→do␈α_not
simulate␈α∃sufficiently␈α∀what␈α∃humans␈α∀do␈α∃when␈α∃humans␈α∀process
language.␈α∞ Several␈α
years␈α∞of␈α
experience␈α∞with␈α∞computer␈α
science
and␈α~linguistic␈α≠approaches␈α~have␈α~taught␈α≠us␈α~the␈α≠scope␈α~and
limitations␈α of␈α syntactic␈α and␈α semantic␈α
parsers.(Schank,Tesler␈α and
Weber,␈↓↓8␈↓␈α Simmons,␈↓↓9␈↓␈α Winograd,␈↓↓13␈↓␈α Woods␈↓↓14␈↓).␈α!While␈α extant
linguistic␈α∪parsers␈α∩perform␈α∪satisfactorily␈α∩with␈α∪carefully␈α∩edited
text␈α∂sentences␈α∞or␈α∂with␈α∂small␈α∞dictionaries␈α∂,␈α∞they␈α∂are␈α∂unable␈α∞to
deal␈α∂with␈α∂everyday␈α∞language␈α∂behavior␈α∂characteristic␈α∂of␈α∞human
conversation.␈α≡ In␈α≡a␈α∨rationalistic␈α≡quest␈α≡for␈α∨certainty␈α≡and
attracted␈α⊂by␈α⊂an␈α⊂analogy␈α⊂from␈α∂the␈α⊂proof␈α⊂theory␈α⊂of␈α⊂logicians␈α∂in
␈↓which␈α∂provability␈α∞implied␈α∂computability,␈α∂computational␈α∞linguists
hoped␈α∪to␈α∩develop␈α∪formalisms␈α∩for␈α∪natural␈α∪language␈α∩grammars.
But␈α⊂the␈α⊃hope␈α⊂has␈α⊃not␈α⊂been␈α⊂realized␈α⊃and␈α⊂perhaps␈α⊃in␈α⊂principle
cannot␈α∀be.␈α∀ (It␈α∀is␈α∀difficult␈α∀to␈α∀formalize␈α∀something␈α∀which␈α∪can
hardly␈α be␈α formulated).
Linguistic␈α~parsers␈α~use␈α~morphographemic␈α~analyses,
parts-of-speech␈α'assignments␈α'and␈α'dictionaries␈α'containing
multiple␈α≥word-senses␈α≡each␈α≥possessing␈α≡semantic␈α≥features,
programs␈α⊗or␈α⊗rules␈α⊗for␈α⊗restricting␈α⊗word␈α↔combinations.␈α⊗ Such
parsers␈α∩perform␈α⊃a␈α∩detailed␈α⊃analysis␈α∩of␈α⊃every␈α∩word,␈α⊃valiantly
disambiguating␈α↔at␈α↔each␈α↔step␈α↔in␈α↔an␈α↔attempt␈α↔to␈α↔construct␈α⊗a
meaningful␈α≤interpretation.␈α≤ While␈α≤it␈α≤may␈α≤be␈α≠sophisticated
computationally,␈α∪a␈α∀conventional␈α∪parser␈α∀is␈α∪quite␈α∀at␈α∪a␈α∀loss␈α∪to
deal␈α∩with␈α⊃the␈α∩caprice␈α∩of␈α⊃ordinary␈α∩conversation.␈α∩ In␈α⊃everyday
discourse␈α⊂people␈α⊂speak␈α⊃colloquially␈α⊂and␈α⊂idiomatically␈α⊃using␈α⊂all
sorts␈α_of␈α_pat␈α_phrases,␈α_ slang␈α_and␈α_cliches.␈α_ The␈α_number␈α_of
special-case␈α↔expressions␈α_is␈α↔indefinitely␈α↔large.␈α_ Humans␈α↔are
cryptic␈α∞and␈α∂elliptic.␈α∞ They␈α∞lard␈α∂even␈α∞their␈α∂written␈α∞expressions
with␈α≤meaningless␈α≤fillers␈α≤and␈α≤fragments.They␈α≤convey␈α≠their
intentions␈α∀and␈α∀ideas␈α∀in␈α∀idiosyncratic␈α∀and␈α∃metaphorical␈α∀ways,
blithely␈α
violating␈α∞rules␈α
of␈α
'correct'␈α∞grammar␈α
and␈α∞syntax.␈α
Given
these␈α
difficulties,␈α∞how␈α
is␈α∞it␈α
that␈α∞people␈α
carry␈α∞on␈α
conversations
easily␈α∂most␈α∞of␈α∂the␈α∞time␈α∂while␈α∞machines␈α∂thus␈α∞far␈α∂have␈α∂found␈α∞it
extremely␈α∃difficult␈α∃to␈α∀continue␈α∃to␈α∃make␈α∃appropriate␈α∀replies
indicating␈α some␈α degree␈α of␈α understanding?
It␈α_seems␈α→that␈α_people␈α→`get␈α_the␈α→message'␈α_without
always␈α∩analyzing␈α∩every␈α∩single␈α∩word␈α∩in␈α∩the␈α∩input.␈α∩They␈α∩even
ignore␈α_some␈α↔of␈α_its␈α_terms.␈α↔People␈α_make␈α_individualistic␈α↔and
idiosyncratic␈α⊃selections␈α⊃from␈α⊃highly␈α⊃redundant␈α⊃and␈α⊂repetitious
communications.␈α⊃ These␈α⊃personal␈α⊃selective␈α⊃operations,␈α⊃ based
on␈α∃idiosyncratic␈α∃intentions,␈α∃produce␈α∃a␈α∃transformation␈α∃of␈α∃the
input␈α∂by␈α∂destroying␈α∂and␈α∂even␈α∂distorting␈α∂information.␈α∂ In␈α∞speed
reading,␈α∪for␈α∪example,␈α∪only␈α∪a␈α∪small␈α∪percentage␈α∪of␈α∪contentive
words␈α≠on␈α≠each␈α≠page␈α~need␈α≠be␈α≠looked␈α≠at.␈α≠ These␈α~words
somehow␈α→resonate␈α~with␈α→the␈α→readers␈α~relevant␈α→conceptual-
belief␈α⊃structure␈α⊃whose␈α∩processes␈α⊃enable␈α⊃him␈α∩to␈α⊃`understand'
not␈α⊂simply␈α⊂the␈α⊂language␈α⊂but␈α⊂all␈α⊂sorts␈α⊂of␈α⊂unmentioned␈α∂aspects
about␈α↔the␈α↔situations␈α↔and␈α↔events␈α↔being␈α↔referred␈α↔to␈α_in␈α↔the
language.␈α∂ Normal␈α∞written␈α∂English␈α∞text␈α∂is␈α∞estimated␈α∂to␈α∂be␈α∞5/6
redundant␈α9 (Rubenstein␈α8 and␈α9 Haberstroh␈↓↓7␈↓).␈α8Spoken
conversations␈α∨in␈α∨English␈α∨are␈α∨probably␈α∨better␈α∨than␈α≡50%
redundant(Carroll␈↓↓1␈↓).␈α≡Words␈α≡can␈α≡be␈α≡garbled␈α≡and␈α≥listeners
nonetheless␈α∂get␈α∂the␈α⊂gist␈α∂or␈α∂drift␈α∂of␈α⊂what␈α∂is␈α∂being␈α⊂said.␈α∂ They
see␈α the␈α "pattern"␈α and␈α thus␈α can␈α supply␈α much␈α of␈α what␈α is␈α missing.
To␈α"approximate␈α!such␈α"human␈α"achievements␈α!we
require␈α_a␈α_new␈α_perspective␈α_and␈α_a␈α_practical␈α_method␈α_which
differs␈α≥from␈α≥that␈α≡of␈α≥current␈α≥linguistic␈α≡approaches.␈α≥ This
alternate␈α∨approach␈α∨should␈α∨incorporate␈α∨those␈α∨aspects␈α∨of
parsers␈α⊃which␈α⊃have␈α⊃been␈α∩found␈α⊃to␈α⊃work␈α⊃well,␈α∩e.g.,␈α⊃detecting
embedded␈α⊂clauses.␈α⊂ Also␈α⊂individualistic␈α⊂features␈α∂characteristic
of␈α an␈α∨idiolect␈α should␈α∨have␈α dominant␈α emphasis.␈α∨Parsers
represent␈α∞complex␈α∞and␈α∞refined␈α∞algorithms.␈α∞ While␈α∞on␈α∞one␈α
hand
they␈α≡subject␈α≡a␈α∨sentence␈α≡to␈α≡a␈α≡detailed␈α∨and␈α≡sometimes
overkilling␈α≡analysis,␈α≥on␈α≡the␈α≥other,␈α≡they␈α≥are␈α≡finicky␈α≥and
oversensitive.␈α⊗ A␈α⊗conventional␈α⊗parser␈α⊗may␈α⊗simply␈α⊗halt␈α⊗if␈α⊗a
word␈α∂in␈α⊂the␈α∂input␈α∂sentence␈α⊂is␈α∂not␈α∂present␈α⊂in␈α∂its␈α⊂dictionary.␈α∂It
␈↓finds␈α∪ungrammatical␈α∩expressions␈α∪such␈α∩as␈α∪double␈α∩prepositions
(`Do␈α~you␈α~want␈α~to␈α≠get␈α~out␈α~of␈α~from␈α~the␈α≠hospital?')␈α~quite
confusing.␈α_ Parsers␈α→constitute␈α_a␈α→tight␈α_conjunction␈α→of␈α_tests
rather␈α⊂than␈α∂a␈α⊂loose␈α⊂disjunction.␈α∂ As␈α⊂more␈α∂and␈α⊂more␈α⊂tests␈α∂are
added␈α∂to␈α∞the␈α∂conjunction,␈α∂the␈α∞parser␈α∂behaves␈α∞like␈α∂a␈α∂finer␈α∞and
finer␈α≡filter␈α≡which␈α∨makes␈α≡it␈α≡increasingly␈α≡difficult␈α∨for␈α≡an
expression␈α⊂to␈α∂pass␈α⊂through.␈α∂ Parsers␈α⊂do␈α∂ not␈α⊂ allow␈α⊂ for␈α∂ the
exclusions␈α typical␈α of␈α everyday␈α human␈α dialogues.
Finally,␈α∂it␈α∂is␈α⊂difficult␈α∂to␈α∂keep␈α∂consistent␈α⊂a␈α∂dictionary
of␈α≡over␈α≡500␈α≡multiple-sense␈α≡words␈α≡classified␈α≡by␈α≡binary
semantic␈α∂features␈α∂or␈α∂rules.␈α⊂For␈α∂example,␈α∂suppose␈α∂a␈α⊂noun␈α∂(Ni)
is␈α∀used␈α∀by␈α∀some␈α∀verbs␈α∪as␈α∀a␈α∀direct␈α∀object␈α∀in␈α∀the␈α∪semantic
sense␈α⊂of␈α⊃a␈α⊂physical␈α⊃object.␈α⊂Then␈α⊂it␈α⊃is␈α⊂noticed␈α⊃that␈α⊂Ni␈α⊃is␈α⊂also
used␈α∞by␈α∞other␈α∞verbs␈α∞in␈α∞the␈α∞sense␈α∞of␈α∞a␈α∞location␈α∞so␈α∞`location'␈α
is
added␈α⊗to␈α⊗Ni's␈α⊗list␈α⊗of␈α⊗semantic␈α⊗features.␈α⊗ Now␈α↔Ni␈α⊗suddenly
qualifies␈α
as␈α a␈α
direct␈α object␈α
for␈α
a␈α lot␈α
of␈α other␈α
verbs.␈α
But␈α some
of␈α_the␈α_resultant␈α_combinations␈α_make␈α_no␈α_sense␈α_even␈α_in␈α_an
idiolect.␈α⊃If␈α⊂a␈α⊃special␈α⊂feature␈α⊃is␈α⊂then␈α⊃created␈α⊂for␈α⊃Ni,␈α⊃then␈α⊂one
loses␈α_the␈α↔power␈α_of␈α_general␈α↔classes␈α_of␈α_semantic␈α↔features.
Adding␈α∞a␈α∞single␈α∞semantic␈α∞feature␈α∞can␈α∞result␈α∞in␈α∞the␈α
propagation
of␈α∃hidden␈α∀inconsistencies␈α∃and␈α∀unwanted␈α∃side-effect..␈α∃as␈α∀the
dictionary␈α∞grows␈α∂it␈α∞becomes␈α∂increasingly␈α∞unstable␈α∂and␈α∞difficult
to␈α control.
Early␈α"attempts␈α!to␈α"develop␈α"a␈α!pattern-matching
approach␈α)using␈α)special-purpose␈α)heuristics␈α*have␈α)been
described␈α↔by␈α⊗Colby,␈α↔Watt␈α⊗and␈α↔Gilbert,␈↓↓2␈↓␈α↔Weizenbaum␈↓↓11␈↓␈α⊗and
Colby␈α∩and␈α∩Enea.␈↓↓3␈↓␈α∩The␈α⊃limitations␈α∩of␈α∩these␈α∩attempts␈α∩are␈α⊃well
known␈α∪to␈α∪workers␈α∪in␈α∪artificial␈α∪intelligence.␈α∪The␈α∪man-machine
conversations␈α∪of␈α∀such␈α∪programs␈α∪soon␈α∀becomes␈α∪impoverished
and␈α∩boring.␈α∩Such␈α⊃primitive␈α∩context-restricted␈α∩programs␈α⊃often
grasp␈α∪a␈α∩topic␈α∪well␈α∪enough␈α∩but␈α∪too␈α∩often␈α∪do␈α∪not␈α∩understand
quite␈α∃what␈α∃is␈α∃being␈α∃said␈α∃about␈α∃the␈α∃topic,␈α∃with␈α∃amusing␈α∃or
disastrous␈α consequences.␈α This␈α
shortcoming␈α is␈α a␈α
consequence␈α of
the␈α⊂limitations␈α⊂of␈α⊂a␈α⊂pattern-␈α⊂matching␈α⊂approach␈α⊂lacking␈α⊂a␈α⊂rich
conceptual␈α∩structure␈α∩into␈α∪which␈α∩the␈α∩pattern␈α∪abstracted␈α∩from
the␈α∀input␈α∃can␈α∀be␈α∀matched␈α∃for␈α∀inferencing.␈α∃ These␈α∀programs
also␈α∀lack␈α∀a␈α∀subroutine␈α∀structure,␈α∀both␈α∀pattern␈α∀directed␈α∀and
specific,␈α desirable␈α for␈α generalizations␈α and␈α further␈α analysis.
The␈α⊃strength␈α⊃of␈α⊃these␈α⊃pattern␈α⊃matching␈α⊂approaches
lies␈α∞in␈α
their␈α∞ability␈α
to␈α∞ignore␈α
some␈α∞of␈α
the␈α∞input.␈α
They␈α∞look␈α
for
patterns,␈α∃which␈α∃means␈α∃the␈α∀emphasis␈α∃of␈α∃some␈α∃detail␈α∃to␈α∀the
exclusion␈α∞of␈α∞other␈α∞detail.␈α∞ Thus␈α∞they␈α∞can␈α∞get␈α∞something␈α∞out␈α∞of
nearly␈α every␈α sentence--␈α sometimes␈α more,␈α sometimes␈α less.
An␈α*interesting␈α)pattern-matching␈α*approach␈α)for
machine␈α≤translation␈α≤has␈α≤been␈α≤developed␈α≤by␈α≥Wilks.␈↓↓12␈↓␈α≤His
program␈α∂constructs␈α∂a␈α∂pattern␈α∞from␈α∂English␈α∂text␈α∂input␈α∂which␈α∞is
matched␈α∪against␈α∪templates␈α∪in␈α∪an␈α∪interlingual␈α∪data␈α∪base␈α∪from
which,in␈α⊗turn,␈α∃ French␈α⊗output␈α⊗is␈α∃generated␈α⊗without␈α⊗using␈α∃a
generative␈α grammar.
In␈α$the␈α%course␈α$of␈α$constructing␈α%an␈α$interactive
simulation␈α∀of␈α∀paranoia␈α∀we␈α∀were␈α∀faced␈α∀with␈α∀the␈α∀problem␈α∀of
dealing␈α with␈α
unedited␈α and␈α unrestricted␈α
natural␈α language␈α as␈α
it␈α is
used␈α$in␈α$the␈α#doctor-patient␈α$situation␈α$of␈α$a␈α#psychiatric
interview.(Colby,␈α⊃Hilf,␈α⊃Weber,␈α⊃and␈α⊃Kraemer,␈↓↓4␈↓␈α⊃Colby␈α⊃and␈α⊂Hilf␈↓↓5␈↓).
␈↓This␈α+domain␈α+of␈α+discourse␈α+admittedly␈α,contains␈α+many
psychiatrically␈α∩stereotyped␈α⊃expressions␈α∩and␈α⊃is␈α∩constrained␈α⊃in
topics␈α∃(Newton`s␈α∃laws␈α∃are␈α∃rarely␈α∃discussed).␈α∃But␈α∃it␈α⊗is␈α∃rich
enough␈α∪in␈α∪verbal␈α∀behavior␈α∪to␈α∪be␈α∀a␈α∪challenge␈α∪to␈α∀a␈α∪language
understanding␈α∂algorithm␈α∂since␈α∂a␈α∂variety␈α∂of␈α∂human␈α∂experiences
are␈α
discussed␈α domain␈α
including␈α the␈α
interpersonal␈α
relation␈α which
develops␈α⊃between␈α⊃the␈α⊃interview␈α⊃participants.␈α⊃ A␈α⊃look␈α∩at␈α⊃the
contents␈α∞of␈α∞a␈α∞thesaurus␈α∞reveals␈α∞that␈α∞words␈α∞relating␈α∞to␈α∞people
and␈α their␈α interrelations␈α make␈α up␈α roughly␈α 70%␈α of␈α English␈α words.
The␈α∩diagnosis␈α∩of␈α∩paranoia␈α∩is␈α∩made␈α∪by␈α∩psychiatrists
relying␈α_mainly␈α↔on␈α_the␈α_verbal␈α↔behavior␈α_of␈α_the␈α↔interviewed
patient.␈α∞ If␈α
a␈α∞paranoid␈α∞model␈α
is␈α∞to␈α
exhibit␈α∞paranoid␈α∞behavior␈α
in
a␈α≠psychiatric␈α≠interview,␈α≠ it␈α≠must␈α≠be␈α≠capable␈α≠of␈α~handling
dialogues␈α⊗typical␈α↔of␈α⊗the␈α↔doctor-patient␈α⊗context.␈α↔ Since␈α⊗the
model␈α∩can␈α∪communicate␈α∩only␈α∩through␈α∪teletyped␈α∩messages,the
vis-a-vis␈α≠aspects␈α≠of␈α≠the␈α≠usual␈α≠psychiatric␈α≤interview␈α≠are
absent.␈α_ Therefore␈α_the␈α_model␈α↔must␈α_be␈α_able␈α_to␈α_deal␈α↔with
unedited␈α⊗typewritten␈α↔natural␈α⊗language␈α↔input␈α⊗and␈α↔to␈α⊗output
replies␈α∂which␈α∂are␈α∂indicative␈α∞of␈α∂an␈α∂underlying␈α∂paranoid␈α∞thought
process␈α during␈α the␈α episode␈α of␈α a␈α psychiatric␈α interview.
In␈α≠an␈α≠interview␈α≤there␈α≠is␈α≠always␈α≠a␈α≤who␈α≠saying
something␈α to␈α
a␈α whom␈α with␈α
definite␈α intentions␈α
and␈α expectations.
There␈α∩are␈α∩two␈α∪situations␈α∩to␈α∩be␈α∪taken␈α∩into␈α∩account,␈α∪the␈α∩one
being␈α~talked␈α~about␈α~and␈α→the␈α~one␈α~the␈α~participants␈α~are␈α→in.
Sometimes␈α↔the␈α↔latter␈α⊗becomes␈α↔the␈α↔former.␈α↔ Participants␈α⊗in
dialogues␈α∪have␈α∪intentions␈α∀and␈α∪dialogue␈α∪algorithms␈α∀must␈α∪take
this␈α⊃into␈α⊃account.␈α⊂ The␈α⊃doctor's␈α⊃intention␈α⊂is␈α⊃to␈α⊃gather␈α⊂certain
kinds␈α∀of␈α∪information␈α∀while␈α∪the␈α∀patient's␈α∪intention␈α∀is␈α∀to␈α∪give
information␈α∂and␈α∂get␈α∂help.␈α∂ A␈α∂job␈α∞is␈α∂to␈α∂be␈α∂done;␈α∂it␈α∂is␈α∂not␈α∞small
talk.␈α∩ Our␈α∩working␈α∩hypothesis␈α∩is␈α∩that␈α∩each␈α∩participant␈α∩in␈α∩the
dialogue␈α%understands␈α%the␈α%other␈α%by␈α&matching␈α%selected
idiosyncratically-␈α≤significant␈α≤patterns␈α≤in␈α≤the␈α≥input␈α≤against
conceptual␈α≡patterns␈α≥which␈α≡contain␈α≥information␈α≡about␈α≥the
situation␈α$or␈α$event␈α$being␈α$described␈α$linguistically.␈α# This
understanding␈α"is␈α"communicated␈α"reciprocally␈α#by␈α"linguistic
responses␈α(judged␈α)appropriate␈α(to␈α(the␈α)intentions␈α(and
expectations␈α∩of␈α∪the␈α∩participants␈α∩and␈α∪to␈α∩the␈α∪requirements␈α∩of
the␈α∞situation.␈α∂In␈α∞this␈α∂paper␈α∞we␈α∞shall␈α∂describe␈α∞only␈α∂our␈α∞current
input-analyzing␈α_processes␈α_used␈α_to␈α_extract␈α_a␈α_pattern␈α_from
natural␈α↔language␈α↔input.␈α_ In␈α↔a␈α↔later␈α↔communication␈α_we␈α↔shall
describe␈α"the␈α"inferential␈α!processes␈α"carried␈α"out␈α"at␈α!the
conceptual␈α
level␈α
once␈α
a␈α pattern␈α
has␈α
been␈α
received␈α
by␈α memory
from␈α the␈α input-analysing␈α processes.
Studies␈α↔of␈α↔our␈α↔1971␈α↔model␈α↔of␈α_paranoia␈α↔(PARRY)
indicated␈α∞that␈α∞about␈α∞thirty␈α∞percent␈α∞of␈α∞the␈α∞sentences␈α∞were␈α
not
understood␈α∩at␈α∪all␈α∩,␈α∩that␈α∪is,␈α∩no␈α∩concept␈α∪in␈α∩the␈α∪sentence␈α∩was
recognized.␈α_ In␈α_a␈α_somewhat␈α↔larger␈α_number␈α_of␈α_cases␈α↔some
concepts,␈α∂but␈α∂not␈α∂all,␈α⊂ were␈α∂recognized.␈α∂ In␈α∂many␈α⊂cases␈α∂these
partially␈α∂recognized␈α⊂sentences␈α∂lead␈α⊂to␈α∂a␈α⊂partial␈α∂understanding
that␈α∞was␈α∞sufficient␈α∂to␈α∞gather␈α∞the␈α∂intention␈α∞of␈α∞the␈α∂speaker␈α∞and
thus␈α≠lead␈α≠to␈α≠output␈α≠an␈α≠appropriate␈α≠response.␈α~ However,
misunderstandings␈α occurred␈α too␈α often.␈α For␈α example:
␈↓∞ DOCTOR: How old is your mother ?
PARRY: Twenty-eight
␈↓PARRY␈α⊃has␈α⊃interpreted␈α⊃the␈α∩question␈α⊃as␈α⊃referring␈α⊃to␈α∩his␈α⊃own
age␈α
and␈α∞answered␈α
by␈α
giving␈α∞his␈α
age.␈α
The␈α∞purpose␈α
of␈α∞our␈α
new
language␈α⊂analysis␈α⊂system␈α⊂is␈α⊃to␈α⊂significantly␈α⊂raise␈α⊂the␈α⊃level␈α⊂of
understanding␈α∞by␈α∂preventing␈α∞such␈α∞misunderstandings␈α∂while␈α∞not
restricting␈α∪what␈α∩can␈α∪be␈α∪said␈α∩to␈α∪PARRY.␈α∩ We␈α∪do␈α∪not␈α∩expect
complete␈α∃under-␈α∀standing␈α∃from␈α∃this␈α∀system␈α∃--␈α∃even␈α∀native
speakers␈α∪of␈α∪the␈α∪language␈α∩do␈α∪not␈α∪completely␈α∪understand␈α∩the
language.
By␈α
`understanding'␈α
we␈α
mean␈α the␈α
system␈α
can␈α
do␈α some
or␈α all␈α of␈α the␈α following:
1)␈α+Determine␈α*the␈α+intention␈α+of␈α*the
interviewer␈α1in␈α0making␈α1a␈α0particular
utterance.
2)␈α_Make␈α_common␈α_logical␈α_deductions␈α_that
follow␈α from␈α the␈α interviewers␈α utterance
3)␈α∞Form␈α∞an␈α∞idioletic␈α∂internal␈α∞representation
of␈α∂the␈α∂utterance␈α∂so␈α∂that␈α∂questions␈α∂may␈α∂be
answered,␈α∀commands␈α∀carried␈α∀out,␈α∃or␈α∀data
added␈α to␈α memory.
4)␈α∂Determine␈α∂references␈α∂for␈α∂pronouns,␈α∂and
other␈α anaphora.
5)␈α∃Deduce␈α∀the␈α∃tone␈α∀of␈α∃the␈α∀utterance,i.e.,
hostile,␈α insulting...
6)␈α!Classify␈α"the␈α!input␈α!as␈α"a␈α!question,
rejoinder,command,␈α ...
␈↓ The␈α∞approach␈α∂we␈α∞are␈α∞taking␈α∂consists␈α∞of␈α∂merging␈α∞the
best␈α⊃features␈α∩of␈α⊃pattern␈α∩directed␈α⊃systems␈α∩such␈α⊃as␈α∩the␈α⊃MAD
DOCTOR,␈↓↓2␈↓␈α⊃ELIZA␈↓↓11␈↓␈α⊃and␈α⊂parsing␈α⊃directed␈α⊃systems␈α⊃for␈α⊂example,
Winograd,␈↓↓13␈↓␈α⊃Woods.␈↓↓14␈↓.␈α⊃ By␈α⊂merging␈α⊃the␈α⊃BNF␈α⊃phrase␈α⊂structure
approach␈α≤ot␈α≤analyzing␈α≤English␈α≤with␈α≤the␈α≤pattern␈α≠matching
approach,␈α
with␈α
its␈α attendant␈α
emphasis␈α
of␈α some␈α
concepts␈α
to␈α the
exclusion␈α↔of␈α⊗others.␈α↔ The␈α⊗programs␈α↔to␈α⊗accomplish␈α↔this␈α⊗are
written␈α∂in␈α∂MLISP2,␈α⊂ an␈α∂extensible␈α∂version␈α∂of␈α⊂the␈α∂programming
language␈α∃MLISP,␈↓↓6,10␈↓␈α∀and␈α∃uses␈α∃an␈α∀interpreted␈α∃version␈α∃of␈α∀the
pattern␈α⊗matcher␈α∃designed␈α⊗for␈α∃a␈α⊗new␈α⊗programming␈α∃language
LISP70.
The␈α∩following␈α∩is␈α∩a␈α⊃basic␈α∩description␈α∩of␈α∩the␈α⊃pattern
matcher.␈α⊃ We␈α⊃shall␈α⊃illustrate␈α⊃its␈α⊃operation␈α⊃using␈α∩examples␈α⊃of
problems␈α common␈α to␈α teletyped␈α psychiatric␈α dialogues.
␈↓ ↓j␈↓
PATTERN MATCHING␈↓
␈↓ Pattern␈α∪directed␈α∀computation␈α∪involves␈α∪two␈α∀kind␈α∪of
operations␈α5on␈α4data␈α5structures:␈α5decomposition␈α4and
recomposition.␈α∀ Decomposition␈α∀breaks␈α∀down␈α∀an␈α∃input␈α∀stream
into␈α components␈α
under␈α the␈α direction␈α
of␈α a␈α
decompostion␈α pattern
("dec").␈α∪ The␈α∩inverse␈α∪operation,␈α∩recomposition,␈α∪constructs␈α∩an
␈↓output␈α⊂stream␈α∂under␈α⊂the␈α⊂direction␈α∂of␈α⊂a␈α⊂recomposition␈α∂pattern
("rec").
A␈α rewrite␈α rule␈α is␈α of␈α the␈α form:
␈↓∞ dec → rec
␈↓It␈α∞defines␈α
a␈α∞partial␈α∞function␈α
on␈α∞streams␈α∞as␈α
follows:␈α∞if␈α∞the␈α
input
stream␈α∂matches␈α∂the␈α∞dec,␈α∂then␈α∂the␈α∞output␈α∂stream␈α∂is␈α∞generated
by␈α∃the␈α∀rec.␈α∃The␈α∀following␈α∃rule␈α∀(given␈α∃as␈α∀an␈α∃example␈α∀only)
could␈α be␈α part␈α of␈α a␈α question␈α answering␈α function:
␈↓∞ How are you ? → Very well and you ?
␈↓If␈α the␈α input␈α stream␈α consists␈α of␈α the␈α four␈α tokens:
␈↓∞ How are you ?
␈↓the␈α output␈α stream␈α will␈α consist␈α of␈α the␈α five␈α tokens:
␈↓∞ Very well and you ?
␈↓␈↓
REWRITE FUNCTIONS ␈↓
␈↓ A␈α∨rewrite␈α≡rule␈α∨defines␈α≡a␈α∨partial␈α∨function,␈α≡for
example,␈α↔the␈α↔mapping␈α↔of␈α↔some␈α↔particular␈α↔token␈α↔into␈α⊗some
other␈α↔particular␈α↔token.␈α↔ A␈α↔broader␈α↔partial␈α↔function␈α↔can␈α⊗be
defined␈α⊗as␈α⊗the␈α⊗union␈α⊗of␈α⊗several␈α⊗rewrite␈α⊗rules.␈α⊗ A␈α∃rewrite
function␈α definition␈α is␈α of␈α the␈α form:
␈↓∞ RULES OF <name> =
dec1 → rec1,
dec2 → rec2,
...
decn → recn;
␈↓␈↓
VARIABLES ␈↓
␈↓ A␈α⊃function␈α⊃is␈α⊃difficult␈α⊃to␈α⊃define␈α⊃if␈α⊃every␈α∩case␈α⊃must
be␈α∩enumerated.␈α∩ Therefore,␈α∪rewrite␈α∩rules␈α∩allow␈α∪variables␈α∩to
appear␈α⊂in␈α⊂patterns.␈α⊂ The␈α⊂value␈α⊂of␈α⊂a␈α⊂variable␈α⊂can␈α⊂be␈α⊃either␈α⊂a
list␈α or␈α an␈α atom.␈α In␈α this␈α paper␈α the␈α notation:
␈↓∞ :X
␈↓where␈α∀X␈α∪ia␈α∀any␈α∪identifier,␈α∀will␈α∪denote␈α∀the␈α∪variable␈α∀X.␈α∪ The
variables␈α∩of␈α∩each␈α∪rule␈α∩are␈α∩distinct␈α∪from␈α∩the␈α∩variables␈α∪of␈α∩all
other␈α rules,␈α even␈α if␈α their␈α names␈α are␈α the␈α same.
The␈α
following␈α
definition␈α
has␈α
only␈α
three␈α
rewrite␈α
rules,
but␈α handles␈α an␈α unlimited␈α number␈α of␈α input␈α streams:
␈↓∞ RULES OF REPLY=
HOW ARE YOU '? → VERY WELL '?
HOW IS :X → I HAVEN''T SEEN :X ',
LATELY'.,
DID :X GO TO :Y '? →
WHY DON''T YOU ASK :X YOURSELF'?;
␈↓A␈α∞variable␈α
can␈α∞appear␈α∞more␈α
than␈α∞once␈α∞in␈α
a␈α∞single␈α∞dec␈α
pattern,
but␈α≡it␈α≡must␈α≡match␈α≥identical␈α≡items␈α≡at␈α≡each␈α≥appearance.
Example:
␈↓∞ RULES OF EQUAL =
(EQUAL :X :X) → TRUE;
␈↓␈↓
ELLIPSIS ␈↓
␈↓ To␈α↔make␈α⊗patterns␈α↔easier␈α⊗to␈α↔read␈α⊗and␈α↔write,␈α⊗the
ellipsis␈α_symbol␈α→...␈α_can␈α→be␈α_used␈α_to␈α→stand␈α_for␈α→an␈α_unnamed
variable.␈α Thus:
␈↓∞ IS ... COMING → NO, ... COULD NOT MAKE IT.
␈↓If␈α an␈α ellipsis␈α (...)␈α occurs␈α
several␈α times␈α on␈α a␈α side,␈α it␈α
designates␈α a
different␈α~variable␈α~each␈α~time.␈α~ The␈α~n'th␈α~ellipsis␈α~in␈α≠a␈α~dec
designates␈α the␈α same␈α variable␈α as␈α the␈α n'th␈α ellipsis␈α in␈α the␈α rec.
Ellipsis␈α_is␈α_one␈α_of␈α_the␈α_principle␈α_ideas␈α→of␈α_pattern
matching.␈α
It␈α
permits␈α
imprecise␈α
mathching;␈α
that␈α
is,␈α
the␈α emphasis
or␈α ignoring␈α of␈α items.
␈↓
AUTOMATIC ORDERING OF RULES ␈↓
␈↓ The␈α⊂order␈α⊂of␈α⊂rules␈α⊂in␈α⊂a␈α⊂function␈α⊂definition␈α⊂does␈α∂not
specify␈α⊃the␈α⊃order␈α⊃in␈α⊃which␈α⊃the␈α⊃system␈α⊃will␈α⊃attempt␈α⊃to␈α⊃apply
them.␈α∂ This␈α∂ordering␈α∂operation␈α∞is␈α∂handled␈α∂by␈α∂a␈α∂special␈α∞system
ordering␈α function.␈α Consider␈α the␈α rewrite␈α function:
␈↓∞ RULES OF REPLY =
I SEE :X → SO WHAT '?,
I SEE ANN → WOW '!;
␈↓Both␈α rules␈α would␈α match:
␈↓∞ I SEE ANN
␈↓In␈α∞such␈α∂cases␈α∞the␈α∂more␈α∞specific␈α∂rule␈α∞takes␈α∂precedence.␈α∞ Thus,
given:
␈↓∞ I SEE ANN
␈↓as␈α the␈α input␈α stream␈α ,␈α the␈α output␈α stream␈α would␈α be:
␈↓∞ WOW !
␈↓but␈α given:
␈↓∞ I SEE STARS
␈↓the␈α output␈α stream␈α would␈α be:
␈↓∞ SO WHAT ?
␈↓A␈α
literal␈α
is␈α∞more␈α
specific␈α
than␈α∞a␈α
variable.␈α
A␈α∞variable␈α
appearing
for␈α
the␈α∞second␈α
time␈α∞is␈α
more␈α
specific␈α∞than␈α
a␈α∞variable␈α
appearing
for␈α∪the␈α∩first␈α∪time␈α∩in␈α∪a␈α∩dec.␈α∪ This␈α∩is␈α∪so␈α∩because␈α∪the␈α∩second
occurence␈α
of␈α
the␈α
variable␈α
must␈α
match␈α
the␈α
same␈α
pattern␈α
as␈α
the
first␈α∪occurence.␈α∩The␈α∪precedence␈α∩function␈α∪is␈α∩itself␈α∪written␈α∩in
rewrites␈α
and␈α so␈α
is␈α both␈α
extendable␈α and␈α
changable␈α by␈α
the␈α user.
Currently␈α precedence␈α∨is␈α calculated␈α by␈α∨a␈α left␈α to␈α∨right
application␈α⊗of␈α⊗the␈α⊗above␈α⊗criteria.␈α⊗ Therefore,␈α↔the␈α⊗following
function␈α defines␈α the␈α LISP␈α function␈α EQUAL:
␈↓∞ RULES OF EQUAL =
(EQUAL :X :X) → T,
(EQUAL :X :Y) → NIL;
␈↓␈↓
SEGMENTS ␈↓
␈↓ Sometimes␈α⊗ it␈α↔ is␈α⊗ desirable␈α↔ for␈α⊗ a␈α↔ variable␈α⊗ to
match␈α an␈α indeterminate␈α number␈α of␈α items.␈α This␈α is␈α notated:
␈↓∞ ::X
␈↓Use␈α∞of␈α
the␈α∞double-colon␈α∞("::")␈α
means␈α∞that␈α
the␈α∞variable␈α∞(e.g.,␈α
X)
will␈α match␈α zero␈α or␈α more␈α items.␈α Example:
␈↓∞ RULES OF APPEND=
(APPEND (::X)(::Y)) → (::X ::Y);
␈↓or␈α if␈α the␈α input␈α stream␈α were:
␈↓∞ (APPEND (A B) (C D E))
␈↓the␈α output␈α stream␈α would␈α be:
␈↓∞ (A B C D E)
␈↓For␈α increased␈α readability␈α the␈α rule␈α could␈α also␈α be␈α written:
␈↓∞ RULES OF APPEND =
(APPEND (...) (...)) → (... ...);
␈↓Another␈α example:
␈↓∞ RULES OF REPLY =
WHERE DID ::X GO →
::X WENT HOME '.;
␈↓Therefore,
␈↓∞ WHERE DID THE CARPENTER GO →
THE CARPENTER WENT HOME.
␈↓␈↓
APPLICATION ␈↓
␈↓ One␈α∂of␈α∂the␈α∞main␈α∂deficiencies␈α∂of␈α∞the␈α∂system␈α∂in␈α∞which
the␈α⊂MAD␈α⊂DOCTOR␈α⊂was␈α⊃programmed␈α⊂was␈α⊂its␈α⊂lack␈α⊃of␈α⊂adequate
subroutining␈α∩capability.␈α∩ Subroutines␈α∩may␈α∩be␈α∩indicated␈α∩in␈α∩the
rewrite␈α system␈α as␈α follows:
␈↓∞ RULES OF LAST =
␈↓∞ () → (),
(:X) → :X,
(:X ...) → <LAST (...)>;
␈↓The␈α⊂"<>"␈α⊃surrounding␈α⊂a␈α⊃pattern␈α⊂means␈α⊃that␈α⊂the␈α⊃current␈α⊂input
stream␈α⊃is␈α⊃to␈α⊃be␈α⊃pushed␈α⊂down,␈α⊃that␈α⊃the␈α⊃function␈α⊃indicated␈α⊂by
the␈α∂first␈α∂token␈α∂within␈α∂the␈α⊂brackets␈α∂is␈α∂to␈α∂be␈α∂entered␈α⊂with␈α∂the
rest␈α
of␈α
the␈α
pattern␈α
appended␈α
to␈α
the␈α
front␈α
of␈α
the␈α∞input␈α
stream,
and␈α∂that␈α⊂the␈α∂output␈α∂stream␈α⊂is␈α∂to␈α∂be␈α⊂placed␈α∂into␈α⊂the␈α∂restored
current␈α↔input␈α_stream.␈α↔Note␈α↔that␈α_MLISP2␈α↔functions␈α_may␈α↔be
called␈α as␈α well␈α as␈α rewrite␈α functions.
␈↓
GOALS ␈↓
␈↓ To␈α→gain␈α→the␈α~advantage␈α→of␈α→goal␈α~directed␈α→pattern
matching␈α∂and␈α⊂computing,␈α∂as␈α⊂well␈α∂as␈α∂the␈α⊂full␈α∂power␈α⊂of␈α∂context
sensitive␈α grammars,␈α the␈α following␈α form␈α may␈α be␈α used:
␈↓∞ RULES OF PREPOSITIONAL_PHRASE =
<PREPOSITION>:P <NOUN_PHRASE>:N
→ (PREP_PH :P :N);
␈↓The␈α∃identifer␈α∃between␈α∃the␈α∃angled␈α∃brackets␈α∃("<>")␈α∃names␈α∀a
rewrite␈α∞function␈α∞the␈α∞rules␈α∞of␈α∞which␈α∞are␈α∞to␈α∞be␈α∞matched␈α∞against
the␈α∞input␈α∂stream.␈α∞ When␈α∂a␈α∞match␈α∞occurs␈α∂the␈α∞output␈α∂stream␈α∞of
the␈α goal␈α will␈α be␈α bound␈α to␈α the␈α associated␈α variable.␈α Example:
␈↓∞ RULES OF PREPOSITIONAL_PHRASE =
<PREPOSITION>:P <NOUN_PHRASE>:N
→ (PREP_PH :P :N);
RULES OF NOUN_PHRASE =
TOWN → (NOUN_PH TOWN),
PALO ALTO → (NOUN_PH PALO_ALTO);
RULES OF PREPOSITON =
IN → IN,
ON → ON;
␈↓and␈α the␈α input␈α stream:
␈↓∞ IN PALO ALTO
␈↓the␈α output␈α stream␈α would␈α be:
␈↓∞ (PREP_PH IN (NOUN_PH PALO_ALTO))
␈↓␈↓
OPTIONALS ␈↓
␈↓ Many␈α≠other␈α~shorthands␈α≠exist␈α~to␈α≠simplify␈α~writing
rules.␈α⊃One␈α⊃useful␈α⊃feature␈α⊃that␈α⊃will␈α⊃be␈α⊃mentioned␈α⊃here␈α⊃is␈α⊂the
optional.
␈↓∞ RULES OF AUXILARY_PHRASE =
<AUXILARY>:A [<NEGATIVE>:N]:N1 →
(AUX_PH :A [:N]:N1 );
␈↓If␈α⊗the␈α⊗optional␈α⊗pattern,␈α⊗enclosed␈α⊗in␈α⊗square␈α↔brackets␈α⊗("[]"),
occurs␈α∞in␈α∞the␈α∞input␈α∞stream␈α∞it␈α∂will␈α∞be␈α∞bound␈α∞to␈α∞:N.␈α∞ :N1␈α∂will␈α∞be
␈↓bound␈α⊃to␈α⊃2.␈α∩ If␈α⊃the␈α⊃<NEGATIVE>␈α⊃does␈α∩not␈α⊃occur,␈α⊃:N1␈α∩will␈α⊃be
bound␈α to␈α 1.␈α On␈α the␈α rec␈α side␈α
of␈α the␈α rules␈α if␈α :N1␈α is␈α 2␈α then␈α
:N␈α will
be␈α∀placed␈α∀in␈α∀the␈α∀output␈α∀stream.␈α∀ If␈α∀it␈α∀is␈α∀1␈α∀then␈α∀nothing␈α∪is
placed␈α∞in␈α∞the␈α∞output␈α∞stream␈α
at␈α∞that␈α∞point.␈α∞ Example,␈α∞given␈α
the
rule␈α above:
␈↓∞ DO → (AUX_PH DO)
DO NOT → (AUX_PH DO NOT)
␈↓␈↓
MORE EXAMPLES ␈↓
␈↓ We␈α have␈α
collected␈α a␈α large␈α
number␈α of␈α
dialogues␈α using
our␈α∩previous␈α⊃program␈α∩PARRY.␈α⊃ These␈α∩dialogues␈α⊃form␈α∩a␈α⊃large
body␈α∞of␈α
examples␈α∞of␈α
the␈α∞kind␈α
of␈α∞English␈α
which␈α∞we␈α∞can␈α
expect.
Martin␈α∨Frost,␈α∨a␈α≡graduate␈α∨student␈α∨in␈α∨Computer␈α≡Science,
Stanford␈α≥University,␈α≡ has␈α≥written␈α≡a␈α≥keyword␈α≡in␈α≥context
program␈α∀which␈α∀enables␈α∀us␈α∪to␈α∀isolate␈α∀examples␈α∀centered␈α∪on
particular␈α⊗words␈α⊗so␈α⊗that␈α∃uses␈α⊗of␈α⊗thoses␈α⊗words␈α⊗in␈α∃context
become␈α∩more␈α∪apparent.␈α∩ Our␈α∪general␈α∩approach␈α∪is␈α∩to␈α∪build␈α∩a
system␈α∀which␈α∀can␈α∀produce␈α∀desired␈α∀intreptations␈α∀from␈α∀these
examples␈α∂and␈α∞to␈α∂incrementally␈α∂add␈α∞to␈α∂the␈α∞rules␈α∂in␈α∂the␈α∞system
as␈α~new␈α~cases␈α→are␈α~discovered␈α~during␈α→the␈α~running␈α~of␈α→the
program.
␈↓ Following␈α!are␈α!some␈α examples␈α!of␈α!commonly␈α occuring
situations␈α∞and␈α
examples␈α∞of␈α∞the␈α
kind␈α∞of␈α∞rules␈α
we␈α∞use␈α∞to␈α
handle
them.
␈↓
QUESTION INTRODUCER ␈↓
␈↓ In␈α∃doctor-patient␈α⊗dialogues␈α∃it␈α⊗is␈α∃quite␈α⊗common␈α∃to
introduce␈α
a␈α∞question␈α
by␈α∞the␈α
use␈α
of␈α∞a␈α
command.␈α∞ The␈α
"question
introducer"␈α↔is␈α⊗followed␈α↔by␈α⊗either␈α↔a␈α⊗<NOUN_PHRASE>␈α↔or␈α⊗a
<DECLARATIVE_SENTENCE>.␈α For␈α example,
␈↓∞ COULD YOU TELL ME YOUR NAME?
␈↓Rather␈α⊃than␈α⊃attempt␈α⊂a␈α⊃literal␈α⊃analysis␈α⊂of␈α⊃this␈α⊃question,␈α⊂which
might␈α lead␈α to␈α the␈α interpretation:
␈↓∞ DO YOU HAVE THE ABILITY TO SPEAK YOUR NAME TO ME?
␈↓we␈α utilize␈α rules␈α like:
␈↓∞ RULES OF SENTENCE =
<QUESTION_INTRODUCER>:Q <NOUN_PHRASE>:N
→ (IS :N '*'?'* );
RULES OF QUESTION_INTRODUCER =
COULD YOU TELL ME → ,
WOULD YOU TELL ME → ,
PLEASE TELL ME → ;
␈↓ Although␈α⊂it␈α⊂is␈α⊂conceivable␈α∂that␈α⊂there␈α⊂are␈α⊂an␈α∂infinite
number␈α∩of␈α∩ways␈α∩to␈α∩introduce␈α∩a␈α∩question␈α∩in␈α∩this␈α∩manner,␈α∩we
have␈α⊂found␈α⊂only␈α⊂about␈α⊃six␈α⊂literal␈α⊂strings␈α⊂are␈α⊂actually␈α⊃used␈α⊂in
our␈α⊂data␈α⊂base␈α⊂of␈α⊂dialogues.␈α∂ When␈α⊂we␈α⊂discover␈α⊂a␈α⊂new␈α∂string
we␈α
incrementally␈α
add␈α a␈α
rule.␈α
When␈α we␈α
have␈α
enough␈α examples
to␈α∃dectect␈α∃a␈α⊗more␈α∃general␈α∃form␈α⊗we␈α∃replace␈α∃the␈α⊗rules␈α∃for
␈↓<QUESTION_INTRODUCER>␈α~by␈α~a␈α~more␈α~elegant␈α~and␈α→general
formulation.␈α⊗ This␈α⊗approach␈α⊗allows␈α⊗us␈α⊗to␈α↔process␈α⊗dialogues
before␈α⊂we␈α⊂have␈α⊂a␈α⊂complete␈α⊂analysis␈α⊂of␈α⊂all␈α⊃possible␈α⊂sentence
constructions,␈α∪and␈α∪it␈α∪allows␈α∪us␈α∪to␈α∪build␈α∪a␈α∀language␈α∪analyzer
based␈α on␈α actually␈α occurring␈α forms.
Notice␈α∀that␈α∀it␈α∀is␈α∪possible␈α∀to␈α∀make␈α∀more␈α∀than␈α∪one
analysis␈α∪of␈α∀any␈α∪given␈α∀sentence␈α∪depending␈α∀on␈α∪what␈α∀is␈α∪being
looked␈α⊗for.␈α↔A␈α⊗poet␈α↔might␈α⊗be␈α⊗interested␈α↔in␈α⊗the␈α↔number␈α⊗of
syllables␈α→per␈α→word␈α→and␈α~the␈α→patterns␈α→of␈α→stress.␈α~ A␈α→"full"
analysis␈α∪of␈α∪English␈α∀must␈α∪allow␈α∪for␈α∀this␈α∪possibility,␈α∪ but␈α∀it␈α∪it
clearly␈α
foolish␈α
to␈α produce␈α
this␈α
kind␈α of␈α
analysis␈α
for␈α
PARRY.␈α Our
analysis␈α⊃will␈α⊂be␈α⊃partial␈α⊂and␈α⊃idiosyncratic␈α⊂to␈α⊃the␈α⊂needs␈α⊃of␈α⊂our
program.␈α This␈α is␈α what␈α is␈α meant␈α by␈α idiolectic.
␈↓
FILLERS ␈↓
␈↓ It␈α⊗is␈α↔quite␈α⊗common␈α↔for␈α⊗interviewers␈α↔to␈α⊗introduce
words␈α∩of␈α∩little␈α∩significance␈α∩to␈α∩PARRY␈α∩into␈α∩the␈α∩sentence.␈α⊃For
example:
␈↓∞ WELL, WHAT IS YOUR NAME?
␈↓The␈α∀"well"␈α∀in␈α∀this␈α∀sentence␈α∀serves␈α∀no␈α∀purpose␈α∃in␈α∀PARRY's
analysis,␈α∞although␈α∞it␈α∞might␈α∞to␈α∞a␈α∞linguist␈α∞interested␈α∂in␈α∞hesitation
phenomena.␈α∞ These␈α∞fillers␈α∞can␈α∞be␈α∞ignored.␈α∞ The␈α∞following␈α
rules
accomplish␈α this:
␈↓∞ RULES OF SENTENCE =
<FILLERS>:F <SENTENCE>:S → :S;
RULES OF FILLERS =
WELL → ,
OK → ;
␈↓␈↓
PUNCTUATION ␈↓
␈↓ Interviewers␈α∂use␈α∂little␈α∂intra-sentence␈α∂punctuation␈α∞in
talking␈α∪to␈α∪PARRY.␈α∪ When␈α∪it␈α∪is␈α∪used␈α∪it␈α∪is␈α∪often␈α∀to␈α∪seperate
phrases␈α that␈α might␈α otherwise␈α be␈α ambiguous.␈α Example:
␈↓∞ WHY WEREN'T YOU VERY CLOSE, FRANK
␈↓Here␈α∪the␈α∪comma␈α∪clearly␈α∩puts␈α∪"CLOSE"␈α∪in␈α∪a␈α∪different␈α∩phrase
from␈α⊃"FRANK".␈α⊃Punctuation,␈α⊃ when␈α⊃used␈α⊃in␈α⊃PARRY's␈α⊃rules,␈α⊂ is
generally␈α∀enclosed␈α∪in␈α∀optional␈α∀brackets␈α∪("[]").␈α∀ This␈α∀has␈α∪the
effect␈α of␈α seperating␈α phrases␈α
when␈α punctuation␈α is␈α used,␈α
but␈α not
requiring␈α full␈α punctuation␈α for␈α the␈α system␈α to␈α work.␈α Example:
␈↓∞ RULES OF SENTENCE =
<SENTENCE>:S1 [',]:C <SENTENCE_CONNECTOR>:SC
<SENTENCE>:S2
→ (CONUNCTION :SC :S1 :S2);
␈↓␈↓
CLICHES AND IDIOMS ␈↓
␈↓ The␈α(English␈α)we␈α(encounter␈α)in␈α(doctor-patient
dialogues␈α∞is␈α∞made␈α∂up␈α∞of␈α∞a␈α∞great␈α∂number␈α∞of␈α∞cliches␈α∂and␈α∞idioms,
␈↓therefore␈α⊃we␈α∩anticipate␈α⊃a␈α∩large␈α⊃number␈α∩of␈α⊃rules␈α∩devoted␈α⊃to
them.␈α For␈α example:
␈↓∞ RULES OF TIME_PHRASES =
A COUPLE OF <TIME_UNIT>:T AGO
→ (TIME (RELATIVE PAST)(REF PRESENT) :T);
RULES OF TIME_UNIT =
SECONDS → (WITHIN CONVERSATION),
MOMENTS → (WITHIN CONVERSATION),
DAYS → (BEFORE CONVERSATION DAYS);
␈↓␈↓
REPRESENTATION CORRECTION ␈↓
␈↓ Intermediate␈α∪results␈α∪are␈α∩often␈α∪produced␈α∪which␈α∩are
misleading␈α∪in␈α∪meaning␈α∪or␈α∪are␈α∪in␈α∪the␈α∪wrong␈α∪form␈α∀for␈α∪further
processing.␈α∂We,␈α∂therefore,␈α∞incorporate␈α∂at␈α∂various␈α∂points␈α∞rules
which␈α≡detect␈α≥certain␈α≡undesired␈α≥intermediate␈α≡results␈α≥and
convert␈α them␈α to␈α the␈α desired␈α form.␈α Example:
␈↓∞ RULES OF CORRECT_FORM =
(QUESTION ... (SENTENCE ...)) →
(QUESTION ... ...);
␈↓␈↓
UNKNOWN WORDS ␈↓
␈↓ Rules␈α⊃can␈α⊃be␈α∩derived␈α⊃to␈α⊃handle␈α⊃words␈α∩which␈α⊃were
previously␈α unknown␈α to␈α the␈α system.␈α For␈α example:
␈↓∞ RULES OF UNKNOWN_WORD =
DR'. :X → <NEW_WORD NAME :X>,
THE :X <VERB_PHRASE>:V →
<NEW_WORD NOUN :X>,
I :X YOU → <NEW_WORD VERB :X>;
␈↓Here␈α "NEW_WORD"␈α is␈α
a␈α function␈α which␈α
adds␈α new␈α words␈α
to␈α the
dictionary.
␈↓
CONCLUSION ␈↓
␈↓ We␈α∞are␈α
faced␈α∞with␈α∞the␈α
problems␈α∞of␈α∞natural␈α
language
being␈α∩used␈α∩to␈α∩interview␈α∩people␈α∩in␈α∩a␈α∪doctor-patient␈α∩context.
We␈α⊃have␈α⊂developed␈α⊃a␈α⊃language␈α⊂processing␈α⊃system␈α⊃which␈α⊂we
believe␈α⊗is␈α⊗capable␈α∃of␈α⊗performing␈α⊗in␈α∃these␈α⊗interviews␈α⊗at␈α∃a
significantly␈α≥improved␈α≥level␈α≤of␈α≥performance␈α≥compared␈α≤to
systems␈α∀used␈α∀in␈α∀the␈α∀past.␈α∀ We␈α∀have␈α∃developed␈α∀techniques
which␈α⊂can␈α⊃measure␈α⊂performance␈α⊂in␈α⊃comparison␈α⊂with␈α⊃the␈α⊂ideal
of␈α⊗a␈α⊗real␈α↔human␈α⊗patient␈α⊗in␈α↔the␈α⊗same␈α⊗context.␈↓↓4,5,7␈↓␈α↔We␈α⊗are
designing␈α
our␈α
system␈α∞with␈α
the␈α
realization␈α∞that␈α
a␈α
long␈α∞period␈α
of
development␈α"is␈α"necessary␈α"to␈α"reach␈α"desired␈α"levels␈α!of
performance.␈α⊂ This␈α⊂is␈α⊃a␈α⊂system␈α⊂that␈α⊃can␈α⊂work␈α⊂at␈α⊃a␈α⊂measured
level␈α∪of␈α∪performance␈α∪and␈α∪be␈α∪improved␈α∪over␈α∪time␈α∀with␈α∪new
rules␈α⊃having␈α⊃minimum␈α⊃interaction␈α⊃with␈α⊃those␈α∩already␈α⊃existing.
Our␈α⊂system␈α⊂is␈α⊂designed␈α⊂so␈α∂that␈α⊂a␈α⊂complete␈α⊂analysis␈α⊂of␈α∂every
word␈α or␈α phrase␈α of␈α an␈α utterance␈α is␈α not␈α neceesary.
␈↓ The␈α∀basis␈α∀of␈α∀this␈α∀system␈α∀is␈α∀a␈α∀rewrite␈α∪interpreter
which␈α↔will␈α↔automatically␈α↔merge␈α↔new␈α↔rules␈α↔into␈α↔the␈α_set␈α↔of
already␈α⊗existing␈α∃rules␈α⊗so␈α∃that␈α⊗the␈α∃system␈α⊗will␈α⊗continue␈α∃to
␈↓handle␈α sentences␈α which␈α it␈α handled␈α in␈α the␈α past.
␈↓ α>␈↓
REFERENCES␈↓
␈↓␈↓↓1␈↓Carroll,␈α⊃J.B.␈α⊃Language␈α⊃and␈α⊃Thought.␈α⊃Prentice-Hall,␈α⊂Englewood
Cliffs,␈α New␈α Jersey,␈α p.␈α 59.
␈↓↓2␈↓Colby,␈α K.M.,␈α
Watt,␈α J.␈α
and␈α Gilbert,␈α
J.P.␈α A␈α
computer␈α method
of␈α⊃psychotherapy.␈α∩Journal␈α⊃of␈α∩Nervous␈α⊃and␈α∩Mental␈α⊃Disease,
142,148-152,1966.
␈↓↓3␈↓Colby,K.M.␈α∨and␈α∨Enea,H.␈α∨Heuristic␈α∨methods␈α for␈α∨computer
understanding␈α∩ of␈α∩natural␈α∩language␈α∩in␈α∪context␈α∩restricted
on-line␈α dialogues.␈α Mathematical␈α Biosciences,1,1-25,1967.
␈↓↓4␈↓Colby,␈α⊂K.M.,␈α⊃Hilf,␈α⊂F.D.,␈α⊃Weber,␈α⊂S.,␈α⊃and␈α⊂Kraener,␈α⊃H.␈α⊂Turing-like
indistinguishability␈α∃tests␈α∃for␈α∀the␈α∃validation␈α∃of␈α∃a␈α∀computer
simulation␈α?␈α⊂of␈α?␈α⊂paranoid␈α?␈α⊂processes.␈α?␈α⊂Artificial
Intelligence,3,199-221,1972.
␈↓↓5␈↓Colby,␈α!K.M.␈α!and␈α"Hilf,␈α!F.D.␈α!Multidimensional␈α"analysis␈α!in
evaluating␈α≥the␈α≤adequacy␈α≥of␈α≤a␈α≥simulation␈α≥of␈α≤paparnoid
processes.␈α∪ Memo␈α∀AIM-194.␈α∪Stanford␈α∀Artificial␈α∪Intelligence
project,␈α Stanford␈α University.
␈↓↓6␈↓Enea,␈α⊂H.␈α∂MLISP,␈α⊂Technical␈α∂report␈α⊂no.␈α∂CS-92,␈α⊂1968,␈α∂Computer
Science␈α Department,␈α Stanford␈α University.
␈↓↓7␈↓Rubenstein,␈α↔A.H.␈α↔and␈α↔Haberstroh,␈α⊗C.␈α↔J.,␈α↔Some␈α↔Theories␈α⊗of
Organization,␈α Dorsey␈α Press,␈α Homewood␈α Ill.,1960,␈α p.␈α 232.
␈↓↓8␈↓Schank,␈α∩R.C.,␈α∩Tesler,␈α∪L.␈α∩and␈α∩Weber,S.␈α∩Spinoza␈α∪ii:␈α∩Conceptual
case-based␈α natural␈α
language␈α analysis.␈α Memo␈α
AIM-109,␈α 1970,
Stanford␈α Artificial␈α Intelligence␈α Project,␈α Stanford␈α University.
␈↓↓9␈↓Simmons,␈α~R.F.␈α→Some␈α~semantic␈α→structures␈α~for␈α→representing
English␈α&meanings.␈α&Preprint,␈α&1970,␈α'Computer␈α&Science
Department,␈α University␈α of␈α Texas,␈α Austin.
␈↓↓10␈↓Smith,␈α⊂D.C.,␈α⊂MLISP,␈α∂Memo␈α⊂AIM-135,␈α⊂1970,␈α⊂Stanford␈α∂Artificial
Intelligence␈α Project,␈α Stanford␈α University.
␈↓↓11␈↓Weizenbaum,␈α∂J.␈α∂Eliza-␈α∂a␈α∂computer␈α∂program␈α∂for␈α∂the␈α⊂study␈α∂of
natural␈α-communication␈α.between␈α-man␈α.and␈α-machine.
Communications␈α of␈α the␈α ACM,␈α 9,36-45,1966.
␈↓↓12␈↓Wilks,␈α Y.A.␈α Understanding␈α without␈α proofs.␈α (See␈α this␈α volume).
␈↓↓13␈↓Winograd,␈α⊃T.␈α⊂A␈α⊃program␈α⊂for␈α⊃understanding␈α⊃natural␈α⊂language.
Cognitive␈α Psychology,3,1-191,1972.
␈↓↓14␈↓Woods,␈α∨W.A.␈α∨Transition␈α∨network␈α∨grammars␈α∨for␈α≡natural
language␈α≠analysis.␈α≠ Communications␈α≠of␈α≠the␈α≠ACM,13,591-
606,1970.