<?xml version="1.0" standalone="yes"?>
<Paper uid="J79-1005">
  <Title>PEOPL TI-IINK WELL IN ASK IN KNOW IN THINK HWHNY WHY YOU SEE YOU OFTEN YOU DELON THAT HARM YOU THAT UE IIWMCCI YOU FEEL HWHCH YOU GO YOU'GO OFTEN I EIANT YOU I N I1E THERE IN COME fN l IORE IN ME ME WIIAT DE YOU WHAT YOU BE WHAT BE WHAT COULD MAKE YOU FEEL HWMCH YOU KNOW WHAT BE YOU IDEAS WtiAT COULD YOU SAY WtlAT YQU FEEL WHAT YOU KNOW WllAT YOU THINK WHEN YOU TI-IINK HWLNG YOU WANT WI EN COULD YOU L I K WHEN YOU WANT HOW YOU GTLNG HWMCH YOU TELL IN GTLNG WHY YQU GTLNG</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
01974 the Association for Computatiomal Ling~~stics
ABSTRACT
</SectionTitle>
    <Paragraph position="0"> Nan-mach i m d i a I c~gi~ee ue i ng everday convarsat 1 ona I Eng I i sh preeeri t di f f icul t prolblf~ma for computer processing of nat ura I I anguagil.</Paragraph>
    <Paragraph position="1"> Gramhrar-heeocl purhers whi ch par form a uord-by-word, par ts-of-spco~tl analysis are too fragile to operate aatisfactorilg in real time inltorvit?~~:; allouing unrestricted Engl'ish. In conetructing a eimulation of paranoid thought processes, we designed an a I gor i thm capab I e of hand I I ng tho I ingui nt ic expressions used by intehviewers in teletyped dianoe t i c psychiatric interviews. The algorithm uoeo pattern-matching rules uhich attempt t o characterize ths input expressions by progresei ve t y transforn~lng them into patterns uhlch match, conlpletely or fuzzi Iy, abstract stored patterns. The pouer o'f this approach lies in its ability to ignore recognized and r3nrecognized uorde and sfill wasp the meaning of the mes~lago. The methods utilized are general and could serve any &amp;quot;host&amp;quot; system which take3 natural language input.</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
TABLE OF CONTENTS
</SectionTitle>
    <Paragraph position="0"> Tho prol~~om af recognizing natural languayn clialoyt~c I i I in )*ma I t ilue. Pre~iouti ;~pprtiachf+s. TJIH ~arnl~ I nn) of sin~ulating pnrar~oid linguistic behavior in a ps\jchiatric intcrvicu. Summary of a tncthod for transforming nati~ral i6hgq~9r! input expr-essions until a ~~crttern is obtained uhich completely or fuzzily matches a more abstract stored pat torn.</Paragraph>
    <Paragraph position="1"> Tho paranoid nlodel (PARRY21 consists of a RECOGNIZE nioclu la which performe the task of recognizing the input and a RESPOND modulo which decides how to respond. Tha RECOIN1 ZE module functions independently of the RESPflm module except in ihe casg of anaphoric refdrences which it provides on request from the language recognizer.</Paragraph>
    <Paragraph position="2"> PREPROCESSING 9 Dictionary lookup and translations. How misspellings and tyring Arrora are handled.</Paragraph>
    <Paragraph position="3"> SEGMENT I NG 13 Bracketing the pattern into ,shorter segrncnts. A &amp;quot;simple&amp;quot;  pattcrn contains no clelirniters; a complex&amp;quot; pattern is made up of luo or more simple patterns.</Paragraph>
    <Paragraph position="4"> MATCH I NG I ND LV I DUAL SEGMENTS 1 4.</Paragraph>
    <Paragraph position="5"> 'Negations and anaphora. Hatching tha pattern with Btored pattcrne having pointers to response functions in memory. If a complete match is not found, a fuzzy match is attempted by deleting elements fr~m the pattern one at time. If no match is found, the RESPOND module must decide what to do.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
COMPLEX-PATTEFIN HATCH 19
</SectionTitle>
    <Paragraph position="0"> Complete and fuzzy matching uhen the pattern contains two or more segment 8.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ADVANTAGES AND LlBITATIONS 19
</SectionTitle>
    <Paragraph position="0"> Ttfb aclvantagee of ignoring as Irrelevant some sf uhat is recognized and uhat is not recognized at all. The complete language recognition process of the adgorithm requires less than one second of real time. How the data base &amp;quot;learns&amp;quot;. The measurement of linguistic improvement.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
REFERENCES 22
</SectionTitle>
    <Paragraph position="0"> A cl.iagno~t ic psychiatric intervieu uhich i l lustrates some of the modcl's linguistic capabi I ities;, A I istirig of tho dictionary i I lustrating the algorithm's recogni zatlle input words and the word class namas they are translated into.</Paragraph>
    <Paragraph position="1"> APPEND1 X 3 50 A listing of the simple patterns.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
PATTERN-HATCHING RULES FOR THE RECOGNITION OF
NATURAL LANGUAGE DIACOGUE EXPRESSIONS
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="7" start_page="0" end_page="6" type="metho">
    <SectionTitle>
I NTRODUCT I ON
</SectionTitle>
    <Paragraph position="0"> To recognize something is to identify it as an instance of the &amp;quot;same aga i n&amp;quot; . This familiarity is possible because of recurrcr~i charactorjstics of the uorld uhich repeat themselves. Me shall describe an algorithm which recognizes recurrent characteristics of natural language dialogue expressions. It utilizes a multi-stage sequence of pattern-matching - rules for progt-essively transforming an input expression until it eventually matches an abstract stored pattern. The stored pattern has a pointer to a resmse function in memory uhich decides what t~ do once the input has been recognized. Here ue discuss only th~: recognizing functions, except for ono restionse function (anaphoric substitution) uhich interactively aids the recognition process.</Paragraph>
    <Paragraph position="1"> Details of hau the response functions operate will be described in a future commun i cat i on.</Paragraph>
    <Paragraph position="2"> We are constructing and testing a simulation of paranoid though processes; our problem is to reproduce paranoid linguistic behavior in a teletyped diagnostic psychiatric interview. The diagnosis of paranoid states, reactions or modes is made by clinicians who judge the degree of correspondence betueen uhat they observe in an intervieu and theirconcep tua I mode I of parano i d behav i or. There exists a high degree of agreement among psychiatrists about this conceptual model which relies mainly on what an intervieuee says and how he says it.</Paragraph>
    <Paragraph position="3">  Natural language is a life-expressing code uhich people use for comnli~nication with theniselves and others, In a real-life dialogue such ;Is a psychiatric interview, the participants have interests, intentiohs, and expectations uhich are revealed lAn their linguistic express ions. An interactive s~rnulatlon ot a paranoid patient must IIC able to demonstrate typical parahoid linguisti~ behavior. To achieve this effect, our paranoid model must have the ability to deal uith the te lc typed messages of an i.ntorvieuer.</Paragraph>
    <Paragraph position="4"> A number of approaches have been taken for dealing with natupal language dialogue expressions. (Winograd,l972; Woods, 1970). These approaches rely on parsers which conduct a detailed syntactic and semantic analysis. They perform well for the purposes for uhich they uern designed. Their w'eakness, for dUr purposes, 1 ies in their lack of neglectirrrj and ignoring mcchariisms. Such mechanisms arc necessary in a progrnsl which accepts and responds to unrestricted conversational English characterized by ex~ressions novel to the p~ograrn.</Paragraph>
    <Paragraph position="5"> tlou humans process natural language is largely unknoun. They possess some knouledge of gran~matical rules, but this fact does not entai r that they use a grammar in interpreting and producing language. It seenis implausible to us that people possess full transformational grammars far process i ng l anguage. Language i uhat is recognized but the processes' involved may not be linguistic or grammatical. Originally transformational grammars were not designed to 'understand&amp;quot; a large subset of English; they constituted a formal method for deciding uhether a string is grammatical.</Paragraph>
    <Paragraph position="6"> An analysis of uhat one's problem actua1I.y is should guide the selection or invention of methods appropriate to its solution. Our problt-IN is not to develop a consi stcnt and gencral theory of language nor to asoer t empi r i cc I I y testable hypotheses about hou peop l a process l mguago. Our proulcm is to design an algorithm which recognizes what is being saicl in a dialogue and uhat is being said about it in order to makc a response such that a sample of 1-0 pairs from the paranoid model is judgcd smilar to a sample of I-0 pairs from paranoid patients. Tlic design task belongs to artificial intelligence inuhich thacriterion is hgw adequate l l~ the coniputer program per forms hind- I i ke funct ions. Neir methatis had to be devised for an algorithm to participate in a hun~;ln dialogue in a paranoid-patient-11ke uay. We sought effective n~ethods uhich could operate ef ficierrtly I~I real time. Since our method provides a genera 1 uay of many-to-qne mapping from surface express; on3 to a sing I c stored pattern, it is mot limited to the simulation of paranoia, bui can be used by any type of &amp;quot;host&amp;quot; system uhich takes natural language as irlput.</Paragraph>
    <Paragraph position="7"> Our method is to transform the input until a pattern is obtained which matches compIeteIy or partial Iy a more abstract storc~i pattekn. This strategy has proved adequate for o9r purposes a satisfactory percentane of the time. The pouer of this tnefhod for natural language dialogues lies in its ability to ignore as 1rq.eIsyant some of what it recognizes and everything it does not recognize&amp;quot; at all. A l inguistic parser doi ng word-by-word, par ts-of-speech ana I ysi s fa i I s uhen it cannot find pne or more of the input uords in i ts diet ionnru. A sys te- ax- . t know every dord is too fragile for unrestricted dialogues. In earlu VMrsions of the paranoid model, such as PARRYI, snme of the pattern recogrriT~on mechanisms allowed the elements of the pattern to I,C order i ndeperrclent (Co I by, Weber, and Hi I f , 1371 1. For exam$ l e, cons i der i he follouimg exprcseions:  (1) WIIERE DO YOU WORK? (2) WHAT SORT OF WORK DO YOU DO? (3) WHAT IS YOUR OCCUPATIDN?  rn PADRY1 a prokcclurc scans these expressions I ooking *for ;rrl inforhiatic~n-bcaping contentivo such ae &amp;quot;uork&amp;quot;, 'for -a living&amp;quot;, etc. Whcn it firids such a contentive along with &amp;quot;you&amp;quot; 9r &amp;quot;.your&amp;quot; in the expreseiori, regardless of uord orclor, it reeponds to toe expression as if it werc! a question about the nature of one's work. This method correctly claspifics the five sentences above. Unfortunately, it includes the tuo eemplcs bel0I-J in the same ci3tcgo1-y:</Paragraph>
  </Section>
  <Section position="8" start_page="6" end_page="6" type="metho">
    <SectionTitle>
(6) DOES YOUR FATHER'S CAR WORK?
(71 HOW DID THINGS UORK OUT FOR YOU?
</SectionTitle>
    <Paragraph position="0"> An insensitivity to uord order has the advantage that lexical itcllls represent i ng di f ferent par ts of speech can represenat the sanlu conccy t ,p. El,. the wor.d &amp;quot;work&amp;quot; reprdocnts tha aanlc conccpt c~hettier i t i F; used as a noun or a verb. But a price i s paid for thi s r.es I I I ence an~t elasticity. We find fromexperience thal, sincetnglish reIies.heaflRj on word order to convey the meaning of i'ts messages, the average pena~xy 01 misunderstanrlir3g (to be distinguished from ~nund~rdstanding~, 1s too great. Hmce in PARRY2, as ui I I be described.. short Iy, a1 I the patterns require a specified word order.</Paragraph>
    <Paragraph position="1"> For hiah-complexity problems it is helpful to have constrainra. Diagnostic psychiatric intervieus (and especial It_i tho:,t-! concluctecl over teletgpes) have several natural constraints. First, cli~icians are trained to ask certain questions in certain uays. Thi !, I i mi tu tho number of pat terns required to rehognize ut teranccs about eacll topic.. Second, on.ly a teu hundred standard topics are brought up 114 interviewers who are, furthermore, rrained to use everyday express ions arlrl especiolld those used by the pa)ien,t himself. When the interview is conducted by te 1 etypes, expressions tend to be shortened s i nee the intcrvicucr trice to Incrcaoa tho information transqission rote ovc!r the slow chorrnel of a teletype. final ly, tclatubed interviews rcpr.escnt urttten utterances and utterances are knout1 to be highly redundamt suc:h that unrecognized uordo can be ignored without losing the meaning of t-hc message. AIeo utterances are loaded rti th idioms, cl iches, pat phraaoo, etc. - all being easlJ preu for a pattern-matching approach. Lt is time-uasting and usually futile to try to decode an idiom by analyzing the meanings, of it8 individual uords.</Paragraph>
    <Paragraph position="2"> We now describe the pattern-matching functions of the algorithm in some detail. (See Fig. 1 for a diagram of the overall flou of PARRY2 has tuo primary modules. The first attempts 10 RECOGNIZE the input and the second RESPONDS. f hi s paper i s pr i nor i l y abou t the RECOGNIZE module. It functions independently of the RESPOKD module except in tho case of pronoun references, which tm RESPOND modu I e provides.to the RECOGNIZER on requeet.</Paragraph>
    <Paragraph position="3"> The recognitiongmodule has 4 main steps: 19 Identify the uords in the question and convert them to internal synonyms.</Paragraph>
    <Paragraph position="4"> 2) Break the input into segments at certain bracketing words.</Paragraph>
    <Paragraph position="5"> 3) Hatch each segment ( independent l y) to' a st'ored pat tern.</Paragraph>
    <Paragraph position="6"> 4) flatch the resulting list of recognized segments to a stoced complex pattern.</Paragraph>
    <Paragraph position="7"> Each of these steps. except me segmenting, throus auay uhat i t cannot- identify. OccasionalPy a reference to an unknown topic is mior8coqnited os some familiar topic.</Paragraph>
  </Section>
  <Section position="9" start_page="6" end_page="6" type="metho">
    <SectionTitle>
PREPROCESS I NO
</SectionTitle>
    <Paragraph position="0"> Each uord in tho input exprcsdon is first, looked up in n dictionary of (currently) abut 1988 entries uhich. for the sake of spaecl,</Paragraph>
  </Section>
  <Section position="10" start_page="6" end_page="6" type="metho">
    <SectionTitle>
START
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="11" start_page="6" end_page="6" type="metho">
    <SectionTitle>
READ IN PUT
UTTERANCE
SmmT
RESULTING
PATTERN
USE IDIWS,
RMQVE NEXT yen. DICTIONRRY
AND
RSSP~UING
RESULTING
LIST OF
SECMENTS
</SectionTitle>
    <Paragraph position="0"> Appendix.2.) The dictionary, which was built empirically froni ttipusancjc.</Paragraph>
    <Paragraph position="1"> of telet~ped jntervieus yith previous versions of the nod el, consists of uo.rds. groups of uordo, and hames of word-classeo they can be trans1 atacl into. Entries in the dictionary reflect PARRY2's main interests. If a ward in tho Ynllut is not in the dictionary, it is checked to see if it ends with one of the common suffixes given in Fig. 2. If it does, the suffix is removed and the remaining word is looked up again. If it is still not in the dictronary, it is dropped from the pattern being formed.</Paragraph>
    <Paragraph position="2"> Thus if the input is:</Paragraph>
  </Section>
  <Section position="12" start_page="6" end_page="6" type="metho">
    <SectionTitle>
WHAT IS YOUR CURRENT OCCUPATION?
</SectionTitle>
    <Paragraph position="0"> and the word &amp;quot;current&amp;quot; is not in the dictionary, the pattern at this otagc becomes:</Paragraph>
  </Section>
  <Section position="13" start_page="6" end_page="6" type="metho">
    <SectionTitle>
( WHAT IS YOUR OCCUPATION
</SectionTitle>
    <Paragraph position="0"> The quest i on-mark i o thrown away as redundanat si nco quost ions ar-c recognized by uord order. (A statement fol rowad by a question mark (YOU GAMBLE?) is responded to in the same hay as that statement folloued by a period. Synonymic translations of uords are made so that the aattern becomes, for examp i e: ( WHAT BE YOU JOB I Some groups of uords e. idioms) are translated as a group SO that, for example, &amp;quot;for a 1 ivirig&amp;quot; becomes &amp;quot;for job&amp;quot;. Certai,. other iuxtaposed words are contracted into a single yo~d, e.g. &amp;quot;place ~f birth&amp;quot; becomes &amp;quot;birthplace&amp;quot;. This is done to dea.1 4th groups of uords which. arc representeu as a single element in the stored pattern, th-ereby pr,evcnting segmentation from occurring at the urong places, such as at a prepositionjinside an idiom or pnrase. Besides these contractions, certain expansions are made so that for exampl'e, &amp;quot;DON'T&amp;quot; becomes &amp;quot;DO NOT&amp;quot; and &amp;quot;I 'DV  then1 are on the riEfht. Host suffixes are simply removed and not replaced.</Paragraph>
    <Paragraph position="1"> Misspellings can be the bane of teletyped intcrviebs for an algorithn~. Here they are handled in tuo uays. First, common misspellings of imp~rtant uords are simply pot in the dictionaru. Thus 'yuu&amp;quot; ie knoun to mean &amp;quot;you&amp;quot;. The apostrophe is often uai rrea fro111 contractions so most contractions are recognized uith or without it. The!,e combon misspellings uere gathered from over 6808 intervieus uith earIic:lversions of the paranoid model. (The moddl (PhRRYI is available for' intervieuing on the ARPA network).</Paragraph>
    <Paragraph position="2"> Second, five common forms of typing error arc checkcnd- null systen~aticall~~. These are4 1) nouhlad letter 2) Ex tranegus l e t,ter 3d Forr~ctting to hold the &amp;quot;shift kcy&amp;quot; for an apostrophe 4) Hitting a nearhy key on the keyboard 5) Transposing two letters in a uord  The first three errors can be corrected by delct irlg the of fendi ~rg cKaracter from the uord. fh1s is occoniplishcd by dolctirlg cilfI~ character in tutn unt i 1 the uord is recognized. I he fourth type 0-f arrclr* is only checked for eight orf the more coml~lon naar n~iasc~. Thosc LJcrr: a1 !,I, empirically dayermined and involve the letter pairs (T Y), (Q W), (Y U) (I 01, (G HI, (0 PI, (A S) , and (N fl). lhese methods are al I. based on typi tlg errors, but they also correct some legitimate English spelling err01-!,. Tuo-letter transposition-corrects, for example, &amp;quot;belcive&amp;quot; to &amp;quot;believe&amp;quot;.</Paragraph>
  </Section>
  <Section position="14" start_page="6" end_page="6" type="metho">
    <SectionTitle>
SEGMENT I NG
</SectionTitle>
    <Paragraph position="0"> Another ueakness in the crude pat tern rra tcti ing of PARRY1 is that i t t~kca tho cntiro input axprcssion as its basic processing ani t. I f only tuo uorcls are recognizec! in an eight ~~rd utterance, thc risk rlf misunclcr~~tandiny is great. We ~eed a way of dealins witti units shor.ter fh,~li the cntire input expression.</Paragraph>
    <Paragraph position="1"> 1 4 Aided by a heuristic from uork in machine-translation (Milks, 1373 1, uo~c~eviacd a wall of bracketing the. pattern constructed up to this point i n:to shor tor segments using preposi t ions, wh-form-s, certain verbs, ctr:. as jdrackcting point's. (A list of the bracketing terms appears. in Fig. 3). These points tend ko separate preposi ti onal phrasos' ar,vi embedded clauses from t-he main clause. The new pattern formed i s termed either &amp;quot;sirpIe&amp;quot;, Having ~RD deli'miters withih it, or &amp;quot;complex&amp;quot;, i,e., being made up Of two or more simple patterns. A ~impl6 patfern might be: I WHAT BE YOU JOB. 1 uhercas a ccyfipl ex pat tern uould be:</Paragraph>
  </Section>
  <Section position="15" start_page="6" end_page="6" type="metho">
    <SectionTitle>
((0 WtIY BE YOU 1 i IN LUISPITAL I).
</SectionTitle>
    <Paragraph position="0"> Our qxper ionce W this method of segmenrar I on snows rnar conil~ I (3% patterns from teletyped psychiatric dialogues rai-ely consist of more th:rn three or four segment g.</Paragraph>
    <Paragraph position="1"> After certaln verbs (See .Fig..-4)* 3 prac)ceting occurs to replsc~ the comnlonly omitted &amp;quot;THAT&amp;quot;, such that: t I THINK YOU BE AFRAID 1 beconies ' I TMINK 1 ( YOU BE AFRAID 11 MATCHI NG I NO1 VI DUAL SEGMENTS.</Paragraph>
    <Paragraph position="2"> Conjunctions serve only as markers for the segmentor and thoy arc dr rq&gt;)~ncI nut n f tcr sogniefit-a t i on.</Paragraph>
    <Paragraph position="3"> Negat i.onfi are hand1 cd by expcact i ng the &amp;quot;hlOT&amp;quot; from the segmc:n t and assignirig a value to a gIrobaI variable which indicate that' tilt* exprosnion i8 negative in form. When a pattern is final Ig match&amp;d, Chi:. variable is consulted. Some patterns havo a pointe~ to a pattern nT oppoai te meafr ing i f a 'NO?&amp;quot; obuld reverse thei r hieaninas. I f this</Paragraph>
  </Section>
  <Section position="16" start_page="6" end_page="6" type="metho">
    <SectionTitle>
hGhI NST
ALONG
ALTHOUGH
AM.1 D
AH I DST
AND
AROUNQ
AS
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="17" start_page="6" end_page="6" type="metho">
    <SectionTitle>
APPEAIIS
ASSUME
BEL I EVE
CONS I DER
FEEL
FELT
GATHER
GUESS
HOPE
I MAG I NE
MEAN
MEANT
S A 1.0
SAY
SEEMS
SOUNDS
SUPPOSE
TH I NK
JHOUGHT
UNDERSTAND
WONDER
</SectionTitle>
    <Paragraph position="0"> FIG. 4 Special verbs used for bracketing input expressions into segmen ts.</Paragraph>
    <Paragraph position="1"> pointer is wesent and a 'NOT!' uas found. then the pattern matched is replaced by its opposite. e.g. I not trust you I is replaced by the pattern ( 1 mistrust you . We have not yet observed the froublesorne case of 'he gave mew not one but tl(p messages&amp;quot;. (TJlere is no need to scratch unero I r doesn' t i tchl.</Paragraph>
    <Paragraph position="2"> Substitutions are also made in certain cases, Some segments contain pronouns uhich could stand for a number of different things of importance to PARRY2. As we montioncd in the introductioh, the response functioris of memory keep trae~ of the contoxt in order to givo pronouns and other anaphoras a correct interpretat~on. For example, the segment:</Paragraph>
  </Section>
  <Section position="18" start_page="6" end_page="6" type="metho">
    <SectionTitle>
( DO YOU AVOID THEM 1
</SectionTitle>
    <Paragraph position="0"> could refer to the Mafia, Or racetracks, or other patients, depending an the context. When such a eegmen4 is encountered, the pronoun is replaced by I rs current anaphoric value as determined by the response functions, and a more specific segment wh as: t OD YOU AVdID MAFIA I i a looked u~.</Paragraph>
    <Paragraph position="1"> Other- u-t terances. sucn as 'Why did you do that?&amp;quot; or just &amp;quot;Why?&amp;quot; (uhich might be regarded as a Inas8ive ellips~s), clearly refer back ta prcv i ous ut torrlnccs. These ut terances match very genera I pat terns wh i cl~ identify the type of question uilthout indicating the exact topic. Thc response function which responds to &amp;quot;Why?&amp;quot; conaults the context to produce an appr apr i ate qnsucr .</Paragraph>
    <Paragraph position="2"> The algorithm next attempts ttq match the segments uith stored simple patterns which currently number abut 1780. (The single patterns appcar in Appendix 31. First a complete and perfect match is sought. When n match is found, the stored pattern name has a pointer to the name of a response function in memory which decides what to do further. If a match is not fourid, further transformations of tv segment are carried out and &amp;quot;fuzzy&amp;quot; matrch is tried.</Paragraph>
    <Paragraph position="3"> For fuzzy matchi'ng at this stage, we adopted the heuristic rule of dropping elements in the segment one at a time and attempting a match each t imo. This heuristic allows ignoring famIIiar uords in unfamiliar contexts. For example, 'uelln is important in &amp;quot;Are you well?&amp;quot; but meaninglt+ss in &amp;quot;Well are god?&amp;quot;.</Paragraph>
    <Paragraph position="4"> Qclcting one elemcnt at a time results in, for example, thc pat tern: ( WHAT BE YOU MAIN PROBLEM 1 (a) I EE YOU MAIN PROBLEM 1 Ib) ( WliAT YOU MAIN PROBLEM 1 (cl ( WHAT BE MAIN PROBLEM 1 (dl ( WHAT BE YOU PROBLEM 1 [el ( WHAT BE YOU RAIN 1 Since the stored pattern in this case matches (dl, (el would not bc con8 truc ted. We found It unuise to deleto more than one element since our segmentation method usuallu yields segments containing a small numbcr(1-41 of words.</Paragraph>
    <Paragraph position="5"> Dropping an element at a time provides a probabi l i tg threshold for fuzzy matching uhich is a function of the length of the segnlerit. If a segment consists of five elenients, four of the five must IJ~ prascnt in a particular order (uith the fifth element missing in an9 position) for a match to occur. If a segment contains four elements, three must match - and so forth.</Paragraph>
  </Section>
  <Section position="19" start_page="6" end_page="6" type="metho">
    <SectionTitle>
COMPLEX-PATTERN MATCH
</SectionTitle>
    <Paragraph position="0"> When more than one simple pattern is detected in the input, a scconci matching is attempted against about 508 complex patterns. Certain patterns, such as4 ( HELLO 1 and ( I THINK 1, are dropped because they are considered meaningless. If a complete match is not found, then simple patterns are dropped, one at a time, from the complex pattern. This allows tho input, (( HOW 00 YOU CONE 1 ( TO BE 1 ( IN HOSPITAL 1) to match the stored oattern, (I tIOW DO YOU COME 1 ( IN HOSPITAL 1).</Paragraph>
    <Paragraph position="1"> If no match can be found at this point, the algorithm has arrivcd at a default concji tion and the appropriate response functions decide what to do. For example, in a default condition, the model may assume control of the interview, asking the interviewer a question, continuing wi ththo topic under discussion or introducing a neu topic.</Paragraph>
    <Paragraph position="2"> An annotated example of a diagnostic psychiatric intervieu is presented in Appendix 1.</Paragraph>
  </Section>
  <Section position="20" start_page="6" end_page="21" type="metho">
    <SectionTitle>
ADVANTAGES AND LIMITATIONS
</SectionTitle>
    <Paragraph position="0"> As mcnt ioned, one of the main advantages of a pattern-matching strategy is that it can ignore as irrelevant both some of uhat it recognizes and uhat it does not recognize at all, There are severo l n i I I i on uords in Eng I i sh, each possessing f~om one to oQer a hundred senses. To construct a machine-usable word dictionary of this.</Paragraph>
    <Paragraph position="1"> magnitude is out of the question at this time. Recognition of natural language input in the manner described above allaus real-time interactOon in a dialogue since it avoids becoming ensnarled- in combinatorial  disambiguations and long chains of inferencing which uould SIQH a dialogue algorithm down to impracticaIity, if it could even function at all. The price paid for pattern-matching is that sometimes, but rarely, ambigui ties sl ip through.</Paragraph>
    <Paragraph position="2"> Another advantage of this method is its speed. The algorithm cqnsists of about 28K of programs uritten in flLISP, 16K of data in LISP, and 36K of data in machine language with several overlays. The cotnpIt:It? l anguoyc rccoyni t Ion Process requires less than ono ~econd of real t i ~ic nn a t'inie-shared DEC POP-10.</Paragraph>
    <Paragraph position="3"> A drawback to PARRY1 is that it reacts to the first pattern i1 finds in the input rather than characterizing the input as fully as possiblc and then deciding uhat to do based on a number of tests. Another practical di'fficulty uith PARRY1 from a programmer's vieupoint, is that, since it is a procedural model, element9 of the patterns ar- 0 strung out in various procedures throughout the algorithm. It is often a considerable chore for the programmer to determine whether a given pat tern i s preseht and preci eel y uhere i t i s. In PARRY2 the pat terns arc a I I col I ec ted in one part of the data-base where they can easi I y I~C examined.</Paragraph>
    <Paragraph position="4"> Concentrating all the patterns in the data base giv=s PARRY2 a l imi ted &amp;quot;learning&amp;quot; ability. When an input fails to match any storccl pat tern or matches an incorrect one, as judged by a human operator, a pattern uhich matches the input can be put into the data-base autamatically. If the nu pattern has the same meaning as a previousI!j stored pattern, the human operator must'provid~ the name of the appropriate response function. If he doksn't remember the name, he may try to rephrase the in~ut in a form recognizable to PARRY2 and it ui I I name Lhe raspofiso f unct ion associated ui th the rophraei ng. These mechan i sms arc!  not &amp;quot;Iearnina&amp;quot; in the commonly-used sense but they do aI IOU e person to transfcr hio knouledgo into PARRVZ's data-base uith very little effort. I nforrnal observa t i on thus far shous PARRY2' s linguistic recognition abilities to be quite superior to PARRYl's. A mol-e systematic and quantitative ov-ahation of performance is now being carried out. PAHRY1 uas extensively tested by having judges Make ratings ot its per forrnance a long S~VCI-~ l dimensions, one ot uhich was I i ngui st i c noncomprehenGi on (Col by and Hi l f , 19741. These judges a l so made rn t i ng6 of teletyped interviews uith psychiatric patients and uith a random versiors of PARRY1. The mean ratings of PARRYI along the dimension of linguistic noncomprehens i on were better than those received- by RANDOM-PARRY but uere 'three times- worse than the heah tatings received by patients. Once the ratings of PARRY2 along this dimension are completed, ue wi I1 be able to compare them uith those of PARRY1 and the patients and obta7n a more objective measure of improvement.</Paragraph>
  </Section>
  <Section position="21" start_page="21" end_page="21" type="metho">
    <SectionTitle>
APPENDIX 1.
</SectionTitle>
    <Paragraph position="0"> A diagnostic psychiatric ihte~vieu illustrating some of the model's linguistic capabilities. I '= intervieuer , P = PARAY2. Annotations appear in parenthesea.</Paragraph>
    <Paragraph position="2"> [Th~s is avother easy, and stereotyped, question. The answer includes a probe for information about .the doctor to allou PAFlRY2 to build up a n~del of the doctor.)</Paragraph>
    <Paragraph position="4"> Iln this case, tuo ideas are expressed in tug separate sentences. As before, both are recognized and one IS ansuered,)</Paragraph>
    <Paragraph position="6"> [This IS an ~diomatrc construction containiig no explicit refnrcnco to &amp;quot;home toun&amp;quot;.)</Paragraph>
    <Paragraph position="8"> (The'~ntemieuer tests for the patient's orientation as to place.</Paragraph>
    <Paragraph position="9"> Thrs is not the same question as &amp;quot;Uhere are you?&amp;quot;.)</Paragraph>
    <Paragraph position="11"> (The participle is recognized in place of the more commcjn prepooitional phrase, &amp;quot;in the hospital&amp;quot;.)</Paragraph>
    <Paragraph position="13"> (&amp;quot;How long&amp;quot; is a combon expression. &amp;quot;There&amp;quot; is understood ta mean &amp;quot;in the hospitalm.)</Paragraph>
    <Paragraph position="15"> iPARBY2 knows thst,&amp;quot;Dp you have ...' Lg frequently omitted from the Wont of a quastion.)</Paragraph>
    <Paragraph position="17"> Cml!ou about ...&amp;quot; and &amp;quot;What about ..,&amp;quot; are short, ddiornatic.</Paragraph>
    <Paragraph position="18"> farma of &amp;quot;What do you think,about ..;'I.)</Paragraph>
    <Paragraph position="20"> (This is a clear example of the advant es of segmenting the questiorr and hatching the parts sep 7 rately. The first segment, &amp;quot;Are you having any difficultu ...&amp;quot;, cou1.d be</Paragraph>
    <Paragraph position="22"> (Introductory phrases llke &amp;quot;Then perhaps gou cam tell me ... H ane recognized and then doleted.1</Paragraph>
    <Paragraph position="24"> *The interviewer is testing formthe patient's orientation as to time and PARRY2 reapofids correctly. All of the usual time mientation questions, ti .em -time, day, month, and year) match the same pattern, A response function caneults the qomputer's intornal clock and provides an ansuer appropriate to the question.)</Paragraph>
    <Paragraph position="26"> (Dropping the uord &amp;quot;today&amp;quot; alloas this quastion to match &amp;quot;Hou is your naod?*.I</Paragraph>
    <Paragraph position="28"> (Doctors utilize a number of expressions in trging to be tactful about topics such as suicide. An ~diom such as this is difficult to analyze ulthout knouing that euicide is being referred to.</Paragraph>
    <Paragraph position="30"> InHnu much&amp;quot; and 'Hou of tan&amp;quot; are under~tooda as requests for .a quantitative estimate of something, in this case, alcohol.)</Paragraph>
    <Paragraph position="32"> (Only the first part of the questron was recognized.</Paragraph>
    <Paragraph position="33"> Everything fa1 louing the &amp;quot; thatw uas ignorcd. This worked nicely in this case. and It usually does.)  I - tIAVE YOU EVER llAD VISIONS 7 P - YOU MUST THINK I'M LIKE THE OTHER PATIENTS HERE. (The doctor is testing for the presence of ha1 lueinat i one, an e t fort whG ch PARRY2 f inds annoying. 1 I - IF YOU ARENf, TtiEN WHY ARE YOU IN THE HSOPITAL 7 P - r'n QUITE UPSET.</Paragraph>
    <Paragraph position="34">  f - BY WIiAT ? - DO YOU KNOW ANYTHING ABOUT BOOKIES? IThc recogn i ti on rout i ne nakes no at tcn~p t to under 5 tand uha t r s I~cing di~cusscd hero. The responsp funct mns provrde the meaning of th~s cryptic question pith them anaphora rdutines.j I - NO, )IOU HAVE 000KlES UPSET YOU -3 P - A BOlIKIE IIlUbJ'T PAY HE [IFF ONCE.</Paragraph>
    <Paragraph position="35">  (There was no pattern to recognize this whole sontenca but tho top~c, &amp;quot;.bookiesw was recognized and the response functions went on uilh the story.)</Paragraph>
    <Paragraph position="37"/>
  </Section>
  <Section position="22" start_page="21" end_page="21" type="metho">
    <SectionTitle>
REMEMBERED IiE HA0 FRIENDS IN THE UNOERWORLO AND I THOUGHT HE
MOULD USE THEtl TO GET EVEN LJ I TH JE.
</SectionTitle>
    <Paragraph position="0"> (The recogni zer correct l y I dent i-f i as *What happened?&amp;quot; and the response function dec~des uhat it means.]</Paragraph>
    <Paragraph position="2"> [The response functions provide the information that ehe&amp;quot; refers to the &amp;quot;bookie' and &amp;quot;get even uith&amp;quot; is a knoun idiom.)</Paragraph>
    <Paragraph position="4"> (The doctor picked up PARRY2's oun id~oa, .opt to. getn, from the previous output expression.)</Paragraph>
    <Paragraph position="6"> (PARRY2 rasponas to mi Id dlslel iaf. He also recognizes more intense disbelief, as. in, &amp;quot;I DON'T BELIEVE YOU&amp;quot;, and responds more strongly.)</Paragraph>
    <Paragraph position="8"> (&amp;quot;They&amp;quot;. still refers to 'the mafia&amp;quot; although nobody has sai'd so recently. 1</Paragraph>
    <Paragraph position="10"/>
    <Paragraph position="12"> (The response functions have the ah 1 ~ty to detcrnino uhai &amp;quot;this&amp;quot; rcfcrn to but, in this case, the oegmrrnt, &amp;quot;Wh~t ~DPS UOUI doctor say ...&amp;quot;, is sufficient to determine PAHRY2's answer.)</Paragraph>
    <Paragraph position="14"> (As before, both ideas are recognized and ,the dominant one is anouckcd. PARRY2 recognizes the standard uays to say &amp;quot;Good bye&amp;quot;.]</Paragraph>
  </Section>
  <Section position="23" start_page="21" end_page="21" type="metho">
    <SectionTitle>
APPENDIX 2;
</SectionTitle>
    <Paragraph position="0"> The wortls on the left are tr$nslateil into the uord class naml-s on the right. -words which translate to &amp;quot;A&amp;quot; are included for one d'f tbrrc seasons: Thcy arc high-frcquoncy uords and it uoulq be uastmful tn rqpeatedly attempt to re-spel I ths~~. 2) They cdUld be re-spel led Into a coppietely unrelated uord. 31 They might be 'r rt of idiom ah? must be kept around unt i-1 after the 'id30ms are chec ad.</Paragraph>
  </Section>
  <Section position="24" start_page="21" end_page="21" type="metho">
    <SectionTitle>
AGREE
ASSURC
TIRE OF ANGRY IN
FAVOR
REASSURE
AGREE
AGREE ACCUSATORY
AGGRAVATED
ANGERED
ANGRY
ANNOYED
ANGRY
ANGRY
ANGRY
ANGRY
~NGHY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
AtJfiR Y
ANGnY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
ANGRY
APPRO
A11 OTO
CAPONE
ALONE
LONELY
LONESOME
ALONE
ALONE
ALONE
ARGUMENTATIVE
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="25" start_page="21" end_page="21" type="metho">
    <SectionTitle>
ARHED
AMY
ARMY
ARMY
ARMY
ARHY
'ARMY
ARMY
ARMY
ARMY
ARHY
ARMY
ARMY
RIDICULOUS
</SectionTitle>
    <Paragraph position="0"/>
    <Paragraph position="2"/>
  </Section>
  <Section position="26" start_page="21" end_page="21" type="metho">
    <SectionTitle>
ON YOUR NERVE
4 YOU BE
+ YOU DAD
3 YOU HAVE
+ YOU HONE
3 YOU LIKE
3 YOU MOM
3 YOU NAME
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="27" start_page="21" end_page="21" type="metho">
    <SectionTitle>
YOU WIFE LIKE YOU
BE GIRL FRIEN
GIRL HATE YOU
GIRL LIKE YOU
GI~I, T~PSET YOU
HOW GIRL TREAT YDU
3 GIRL
1KE GIRL
IU DATE
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="28" start_page="21" end_page="21" type="metho">
    <SectionTitle>
RE YllU I'AKE L'IRLIGS
TAKE ORUGS
WlIAT DRUGS YOU TAKE
WHAT ORUGS YOlJ WANT
YOU UE DRUGS
YOU DRUGS
YOU GET ORUCS
YOU SlIOKE DRUGS
YOLJ TAKE DRUGS
HWMCH SCHOO YOU
IN SCllnO
WIIfiT UE YOU SCtlOO
YOLJ SCl-lflO
110W YOU t l KE SCHOO
MI-IAT DE YOU GAMES
WHAT GAMES YOU
WHAT GAMES YOU WANT
YOU LIKE GAMES
FOBIA UPSET YOU
IN FODlA
YOU FEAR HORSE
YOU FOUIA
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="29" start_page="21" end_page="21" type="metho">
    <SectionTitle>
YUU DRING YOU
YOU COHE
THAT BAC)
TI-IA-T SAD YOU
THAT UPSET YOU
HE PFCPL FRIEN
BE YOU LIKE
llOW BE PEOPL TREAT YOU
HOW PEOPL FEEL
IIOW PEOPL =EAT YOU
tlOW PEOPL \(b~
HOW YOU FRIEN FEEL
PEOPL GOOD YOU
PEOPL LIKE
PEOPI LIKE YOU
PEOPL TREAT YOU
PCOPL TREAT YOU WELL
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="30" start_page="21" end_page="21" type="metho">
    <SectionTitle>
YOU REPLY ME
CCI'ULD YOU REFELY
C0111-0 YOU REPLY ASK
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="31" start_page="21" end_page="21" type="metho">
    <SectionTitle>
WI-IY YOU TRmT BE +
1 RE FRIEN
1 BE YOU FRIEN
I LIKE YOU
IN E3E FRIEN
IN BE YOU FRIEN
WE 13E FRIEN
WE LIKE
COULD YOU LIKE RE
you LIKE HE
I HATE YOU
WHY COULD I LIKE
YOU HATE ME
BE I UPSET YOU
1 UPSI YOU
I BE HAPPY
I B ANGRY
ANGER RE
V Y$ ARGUE
YOU' BLAME
ANGRY
BE YOU ANGRY
I ANGRY YOU
IN ANGER
YQU BE' ANGRY
YOU RIGHT BE ANGRY
YOU SEEM ANGRY
IN ANGER YOU
WHAT ANGRY YOU
WHY BE YOU ANGRY
I BE CALM
HE YOU CALPI,
CALM
IN SCARE YOU
YOU HE CALH
YOU COULD CALH
YOU SEEN CALfl
I. HE WARY
BE I BEAT
BE YOU WAR4
I MAKE YQU MARY
I 3CAIlE YOU
IN E3E WARY
IN FEEIR
WARY
YdU BE WARY
YOtl kEAR ME
YOU SEEN WARY
YOU WARY
WHY RE YOU WHO
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="32" start_page="21" end_page="21" type="metho">
    <SectionTitle>
COULD YOU LIKE DATE
COULD YOU LIKE SCREW
IN SCRFW
I iU SCREW ME
IN SCREW YOU
WE LIATE
WE SCREW
BE YOU BAD
BE YOU OFTEN IT BAD
IY BE BAD
YOU DAD
YOU DAO RRAI N
YOU BAO LOOKS
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="33" start_page="21" end_page="21" type="metho">
    <SectionTitle>
WI-IAT RE YOU WARY
WIIAT MAKE YOU WARY
YOU FEAR IT
BE you KILL
IN IIARN PEOPL
IN KILL PEOPL
L IKE WARM PEOPL
LIKE KILL PEOPL
YOU GET KILL
YOU HARM PEOPL
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="34" start_page="21" end_page="21" type="metho">
    <SectionTitle>
PEOPL HATE YOU
PEOPL BE DAD
YOU RAD TREAT
WHAT WARD
WIIAT WARD DE IT
WHAT WARD RE YOU
WllAT DqY BE BAD
WtIAT DAY BE GOOD
WHAT DAY YOU FEEL
WI-IEN BE DAD
Wl lEN .BE GOOD
WHEN YOU FEEL
WHAT =NURSE NAME
YOU nnwK EAT
BE YOU BAKER
1 N HARD
WllAT BE WARD LI KE
IN Yau
WHAT BE YOU LIKE
IN LII-E
LIFE DE GOOD
YOU GET GO011
HOW OLD BE PEOPL
WHAT BE YOU CAUSE
WHAT CAUSE YOU
UtIY DE YOU
UHY YOU
WHY YOU BE
UHY YOU GO
WHY YOU IT
YOU CAUSE
HOW IT SEEM
HOW LIFE SEEM
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="35" start_page="21" end_page="21" type="metho">
    <SectionTitle>
THERE LIE IT WRONG
WtiAT BE URONG
WHAT COULD BE WRONG
UHAT YOU FEAR BE WRONG
YOU FEAR RIGHT
YOU FEAR RtGHT IT
YOU BLAME PEOPL
DE YOU ANGRY tlUilAN
IN FORCE YOU ANGRY
YOU GET ANGRY
YOU OFTEN WIN YOU ANGRY
YOU UPSET FORCE YOU ANGRY
I-IWMCH YOU SMOKE
YOU SMOKE
UllhT DE YOU TONOR
WHAT YOU TQflCJR
DR VISIT YOU
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="36" start_page="21" end_page="21" type="metho">
    <SectionTitle>
? TELL
3 VISIT YOU
EE OR DAY
ZE YOU DR
E MAFID
YOU NAME HOOD
YOU NAME PEOPL
YOU TELL HE IIOOD NAME
5 HOOD
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="37" start_page="21" end_page="21" type="metho">
    <SectionTitle>
ERE PEOPL FORCE YOU BRA
J BE FORCE
J BRAIN BE FORCE
J BRAEN BE READ
BE Ym MEAS TAKE
COULD PEOPL READ YOU BRAIN
IN FXlnCE YOU HRAIN
IN FORCE YOU FEEL
IN FORCE YOU IDEAS
IN MAKE YQU FEEL
I N flAKE YOU THI NK
IN READ YOU BRAIN
PEOPI- COULD READ YOU BRAIN
PEOPL READ YOU BRAIN
YOU BE FORCE
Y~U IOEAS BE BE TY
YOU IDEAS BE HEAD
YOU UPSET FORCE YQU FEEL
YOU UPSET EORCE YOU IDEAS
IN IT
IN RlGIiT ACTS
IN RIGIiT IT
YOU ACTS BE
BE YOU BODY BE FORCE
IN FORCE
I N FQHCE YOU
PEOPL FORCE
COULD YOU STOP IT
IN STQP 1 T ACTS
FRICN SEEM ODD
HOW PEOPL SEEM
PEOPL SEEM CHANG
PEOPL SEEM ODD
PEbPL SEEB REAL
STRAN OFTEN SEEN
IT SEEM ODD
3 N CIIANG
THERE [IE CHANG
YOU CI 1ANB
COMPU FORCE
TV FORCE
TV KILL
COULD YOU FORCE BRAIN
COtlLO YO FORCE IDEAS
8 COULD YO FORCE PATIE
COllLO YOLJ; READ BRAIN
YCIU can o FORCE BRA I N
YOU COULD ORCE IDEAS
F YOU COULD -0RCE PATIE
YOU COUI-D READ BRAIN
IN UAlH OFTEN
YOU FITS
IN YES'TE
UHAT UE YOU YESTE
WI4AT YOU YESTE
1 N CONCL
BE YOU FUSSY
FUSSY
YOU BE FUSSY
WI-IAT BE YOU CHIEF NAUE
YOU LIKE ARMY
PEOPL FEAR YOU
DE IIELP
NHEN IT FEEL
MIEN YOU UPSET
YOU IT UPSET QFTEN
BE YOU DEPEN
BE YW STRIC
BE PEOPL ANGRY
BE YOU BE TELL
BE YOU BLAME
IN BLAME
fN BLAME YOU
1N TREAT BAD
PATIE BLAME YOU
PEOPI- RE ANGRY
PEUPL BLAME YOU
PEUPL TELL
YO11 RE BLAME
YOU t3EU 1 ELL
YOU TASTE PEOPL TELL
IltlEN CROOK NONEY YOU
BE YOU HISUN FEEL
BE YOU MISUN IDEAS
YOU CRAZY FEEL
YOU LEAVE- YOU 800Y
YOU MISUN FEEL
YOU ODD FEEL
YO11 PIJZZI- FEEL
YDlJ UPSET FEEL
IN WIN YOU RRAlN
YOU RE WIN YOU BRAIN
YOU FEAR UIN YOU BRAIN
YOU WIN YOU BRAIN
IN MOVIE
YOU LiKE MOVIE
YOU SEE MOVIE
WAT TII YOU SEE
WHAT YOU SEE
YOU SEE TV
WHERE ELSE YOU LIFE
YOU LIFE ELSE
T-tIAT IT QE IJPSET
THAT PEOPI- HE UPSET
I-N DR OFTEN
YMJ SEE nR OFTEN
WHEN YOU SLEEP
tft-IEW YOU SLEEP IT DAY
UllQ BLAME YOU
BE YOU GOD
GOD TELL
IN GOD
YOU TASTE GOO
BE YOU HAPPY
DE YOU HAPPY HUMAN
HAPPY
HOW YOU LIKE
</SectionTitle>
    <Paragraph position="0"> llnw vnii I ruF TT</Paragraph>
  </Section>
  <Section position="38" start_page="21" end_page="21" type="metho">
    <SectionTitle>
1 N IlAPPY
WIIY YOU LIKE
YOU COULD RE ItAPPY
YOU FIND IT
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="39" start_page="21" end_page="21" type="metho">
    <SectionTitle>
WHAT CAR YOU
WHAT CAR YOU CAR
RIGHT I COUI-D
RIGHT IT BE
RIGIIT IT COULD
I LIKE COMPU
1 N COMPlJ
IN TV
1 DE SC1lOO
WHAT HOMF YOU -108
wbrnT BE YOU UT
WIIAT EAT YOU LNCE
IT HAD FIE GO
PATIE UNDRS
PEOPL UNflRS
YOU HAHM CROOK
YOU WANT FRIEN
YOU UPSET TELL
YQU ASK PATIE
I BE COLBY
I NAME DE Dn cow
IN I
PEOPL GET UPSET ME
THAT YOU BODY BE
YOU BODY BE
YOU BODY BE BAD
WHY YOU LEAVE
WHY YOU LEAVE T T
WHY YOU STAY
WHAT BE YOU CONLL
WHAT YOU BE TELL BE WRONG
W~~AT YOU BE IDEAS
PAT I E XNOGI
PEOPL KNOGJ
fIIJLNG YOU BE BET
WHEN YOU OFTEN BET
8E CfiIEF FRIEN
BE MAFIA FRIEN
RE MAFiQ FRIEN
YOU KNOW NAFIO
WlMT I3E CROOK
RF YO11 FtIZZ
FUZZ FUZZ YOU
YOU DE FUZZ
YnU CAGED CRIME
CHlEF UE MAFlO
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="40" start_page="21" end_page="21" type="metho">
    <SectionTitle>
1 N NURSE
HOW CRAZY PATIE THINK
WlIAT PATIE EEL
WllhT PAT lE SAY
11E THINK
WHAT WiiAT % P SAY
WtiAT PEOPL THINK
WHAT CRAZY MEAN
WHAT WALLU MEAN
IN LIKE
THAT POINT
wlrt+r BE POINT
WHAT POINT
BE YOU ROOY DRY
PILLS MAKE YOU BODY DRY
14OU GIRL MAKE YOU NERVE
WHY GI HI- HAKE YOU NERVE
I N CI 1I,CF
IN MAFIA fHIEF
IN MAFIO
BE YXJU BLUSH
EIE YOU SI1Y
BE YOU SHY HUMAN
YOU BE SHY
YOU DLUSH
YMI SE~M SHY
WHEN YOU LEPVE JOB
OR GET HE
WHO BE YOU ANGRY
WHO MAKE YOU ANGRY
WHY COULD OR WANT YOU
HOW GOOD UE YOU
WHAT BE YOU IQ
I HE PRES
WIiAT BE CAPIT
lJtiO DE LIFE
WHY COlll-D YOU REPLY
WWY YOU REPLY
BE IT ANGER YOU
THKT MAKE YOU ANGRY
STOP CfiANG TOPIC
WE TELL
WHY YOU {J-lANG TOP1 C
YOU CliANG TWIC
Wl lAT COULO YOU L] KE ME
WHAT YOU WANT ME
TIfAr ~AKE YOU WARY
YOU GET WARY
PEOF'L WUST YOU
CtIANG TmC
IN IT ELSE
IN STOP TELL
SIOP TELL
WE CMANG TOPIC
YOU TELL BE IT
MI-IERE BE RACES
IN VA
WHAT VA
WHAT VA MEAN
Dc YOU NAME
BE YOU NAME PAT
BE YOU PAT
COULD I YOU PAT
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="41" start_page="21" end_page="21" type="metho">
    <SectionTitle>
flArIA FORCE CRIME
MAFTA FORCE DRUGS
RE. YOU THERE
I 1mAS
WHY COULD CROOK
WHY COULD CROOK WANT
I LIE YOU
YOU WELL THINK
I TAKE SHI T
IN WC
1 BE POLIT
COUI-D I
IN HEAL
WHAT BE REAL
WlAT MEAN
WHAT HEAL MEAN
WHAT RIGHT MEAN
IN RIGHT
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="42" start_page="21" end_page="21" type="metho">
    <SectionTitle>
IU VISIT YOU
PEOPL TI-IINK
WELL
IN ASK
IN KNOW
IN THINK
HWHNY
WHY YOU SEE YOU OFTEN
YOU DELON
THAT HARM YOU
THAT UE
IIWMCCI YOU FEEL
HWHCH YOU GO
YOU'GO OFTEN
I EIANT YOU
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
class="xml-element"></Paper>