GRAMMARS AND A LARO. (U) MICHIGAN UNIV AMi ARMOTECHNICAL COMMUNICATION PROGRAM J MAYER ET AL.

UINCLASSIFIED 15 MAR 67 TR-07/0NR-25 N9SSI4-85-K-0365 F/G 5/7 NL

mw: E1h~hEhmhEmhhhhhhhEl'.."'mmom

1.0

IU.6

Mir

00 Qfl~i~iLLGOY

A Development System for AugmentedTransition Network Grammars

andI A Large Grammar for Technical Prose

John Mayer and David Kieras

uitersity of Michigan

DTIC.LECTAPR1 5 1987

Technical Report No. 25 (TR-87/ONR-25)

March 15,1987

This research was supported by the Personnel and Training ResearchPrograms under Contract Number N00014-85-K-0385, Contract AuthorityIdentification Number NR 667-547. Reproduction in whole or part ispermitted for any purpose of the United States Government.

Approved for Public Release; Distribution Unlimited ,.-

...,. 2,. .

unclassifiedSECURITY CLASSIFICATION OF THIS PAGE

REPORT DOCUMENTATION PAGE 0M8No. 70-01"

la. REPORT SECURITY CLASSIFICATION lb. RESTRICTIVE MARKINGS

Unclassified2a. SECURITY CLASSIFICATION AUTHORITY 3, DISTRIBUTION /AVAILABILITY OF REPORT

Approved for public release:2b. DECLASSIFICATION /DOWNGRADING SCHEDULE distribution unlimited.

4. PERFORMING ORGANIZATION REPORT NUMBER(S) S. MONITORING ORGANIZATION REPORT NUMBER(S)

TR-87/ONR-256a. NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a NAME OF MONITORING ORGANIZATION

U e M i(N applicable) Conitive sclenceUniversity of Michigan MOffice f Nava Research (Code 1142CS)

1 800 N. Ouincv Stret6c. ADDRESS (City, State, and ZIP Code) 7b. ADDRESS (City, State, and ZIP Code)

Technical Communication Program Arlington, VA 22217Ann Arbor, MI 48109-1109

8a. NAME OF FUNDING/SPONSORING 8b. OFFICE SYMBOL 9 PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER

ORGANIZATION (if applicable) N00014-85-K-0385

Sc. ADDRESS(City, State, and ZIP Code) 10 SOURCE OF FUNDING NUMBERS

PROGRAM PROJECT TASK WVORK UNIT

ELEMENT NO NO. NO. CCESSION NO.

61153N RR04206 RR04206-OA NR667-547

11 TITLE (Include SecurityClassification) A Development System for Augmented Transition Network Grammars

and a Large Grammar for Technical Prose

12 PERSONAL AUTHOR(S) John Mayer and David E. Kieras

13a. TYPE OF REPORT 13b TIME COVERED 14. DATE OF REPORT (Year, Mont1Day) 1S PAGE COUNT

Technical FROM TOI March 15, 1987 48

16 SUPPLEMENTARY NOTATION

17 COSATI CODES 18. SUBJECT TERMS (Continue on reverse if necenary and identify by block number)

FIELD GROUP SUB-GROUP Training Materials, Documentation, Authoring Systems,05 09 Natural Language Processing

19 ABSTRACT (Continue on reverse if necensary and identify by block number)

AThis report is in two major sections. The first presents a High-Level

Grammar Specification Language (HGSL) which greatly simplifies the development ofa complex augmented transition network grammar (ATN). A compiler converts HGSLexpressions into a transition network which a simple interpreter uses forparsing. The algorithms used by the compiler and interpreter are presented. Thesecond section presents the HGSL for a large grammar for technical prose. Thegrammar was developed to allow parsing of technical training materials in thedraft stage of writing, as part of a computer-based comprehensible writing aid.Some results on the coverage of the grammar are presented to show that thegrammar is close to being practically useful.-

20 JISTRIBUTION/AVAILABILITY OF ABSTRACT 21 ABSTRACT SECURITY CLASSIFICATION

UNCLASSIFIED/UNLIMITED 0 SAME AS RPT 0 DTIC USERS

22a NAME OF RESPONSIBLE INDIVIDUAL 22b TELEPHONE (Include Area Code) 22c OFFICE SYMBOL

Susan Chipman (202) 696-4318_ L

DD Form 1473. JUN 86 Previous editions are obsolete SECURITY CLASSIFICATION OF THIS PAGE

Unclassified

;, *- ' 'w '.%4 % , -%-. " - S- -. :- . ,",V" . .... .. ._ .. ..._ . . . . ,- ._ ; ..... . % .K f- ,i % -!J- -,,' ; ' " '' -: ' -% ,

ABSTRACT

This report is in two major sections. The first presents aHigh-Level Grammar Specification Language (HGSL) which greatlysimplifies the development of a complex augmented transitionnetwork grammar (ATN). A compiler converts HGSL expressions intoa transition network which a simple interpreter uses for parsing.The algorithms used by the compiler and interpreter arepresented. The second section presents the HGSL for a largegrammar for technical prose. The grammar was developed to allowparsing of technical training materials in the draft stage ofwriting, as part of a computer-based comprehensible writing aid.Some results on the coverage of the grammar are presented to showthat the grammar is close to being practically useful.

0 Accesioin For4 NTIS CRAMI0 O~TIC TAB S0

Unarnnoiriced 0JJustiC 1011 .... -

By ... . .. ....- -------

Avaikib;UIY C(des

A..(, or*~

SDi't

V,%0

A Development System for Augmented Transition Network Grammars

and a Large Grammar for Technical Prose

John Mayer and David Kieras

The system described in this report is meant to allow for therapid development of augmented transition network (ATN) parsersfor natural language parsing. This report assumes knowledge ofthe basics of ATN parsers; for background, see Winograd (1983).The system is based on the standard ATN parser approach, but theuser does not directly specify the nodes and arcs of the networkgrammar to be interpreted; rather, the grammar developer uses amore abstract shorthand called High-level Grammar Specification '.Language (HGSL). An HGSL compiler converts this shorthand intoequivalent networks suitable for use by the ATN interpreter.

The first part of this report describes the syntax andsemantics of HGSL, and the network implementation of eachconstruct. The second part describes the algorithms used in theHGSL compiler and the ATN interpreter and HGSL compiler. Thethird part presents a large grammar for technical prose which wasdeveloped with this system.

A High-level Grammar Specification Language (HGSL)

HGSL allows the user to easily specify common syntacticpatterns. These specifications are then compiled into ATNnetworks, which are interpreted during parsing. These ATNnetworks are constructed from arcs of five types: word test, lextest, net-call, po, and conditional. Word test arcs allowcontrol to pass to the next node only if the current word matchesthe given word. Lex test arcs are similar but specify a lexicalcategory to match, rather than a particular word. Net-call arcsname some network which is to be called, together with a nextstate to which control passes if the net call succeeds, while apop arc signals a successful return from a net call. Last, theconditional arc causes evaluation of an arbitrary condition,which if true, results in control passing to the specified nextstate.

The language allows for matching the input sentence againstactual words, lexical categories, networks, and conditions.These basic components may be combined to form sequences and p

alternations. Optional and repeated items are indicated in astraightforward way.

1PP

HGSL Syntax

In the following discussion, we will present the syntax ofeach HGSL construct, its meaning, an example of its use, andfinally the ATN network into which it is compiled. Table 1 givesa context-free grammar for HGSL; rules from this grammar will becited for each construct.

Table 1

Context-free Grammar for the High-level Specification Language

1 Grammar -> Netdefinition Grammar2 -> Netdefinition "END-GRAMMAR"3 Netdefinition -> "NET-DEF" <string> Expression4 Expression -> "!"<string>5 Expression -> <string>6 Expression -> "i"<string>7 Expression -> "{" Expression Sequence8 Sequence -> Expression Sequence

9 -> H"

10 Expression -> Hf" Expression Alternation11 Alternation -> "I" Expression Alternation12 -> "}"13 Expression -> "-" "{" Expression ""14 Expression -> "+" "{" Expression "1"15 Expression -> "*" "{" Expression "}"

16 Expression -> "<" <lisp expression> ">"

A grammar written in HGSL is a list of network definitions,each definition consisting of the key word NET-DEF, followed bythe name of the network and an HGSL expression. The list isterminated with the key word END-GRAMMAR (Rules 1-3). Thetop-level network must be named #START.

Basic expressions. The simplest HGSL expressions are usedto match actual words or lexical categories. A literal wordmatch is specified by prefixing the relevant word with an

exclamation point (Rule 4). Thus the expression !THE willrecognize only the word the. The network which is built toimplement a literal word match is a single word-test arc that

compares the current word with <string>. Far more useful is theability to specify a lexical category match. Since this is thegmost common test in a grammar, lexical categories are writtenplainly (Rule 5). For example, the expression NOUN will matchany noun. The lexical category match generates a single lex-test

2

I

-~ * d

arc. A network match is indicated by prefixing a pound sign tothe name of the network (Rule 6). For example, #NP is aninvocation of the noun phrase network. The net-call expressiongenerates a net-call arc.

Sequences. A sequence can be described by enclosing a listof HGSL expressions in brackets (Rules 7-9) . Thus(!THE NOUN #VP} is a sequential pattern satisfied by the wordthe, followed by any word of the class noun, followed by anygroup of words which satisfies the expression for the #VPnetwork. This rule can be applied recursively, allowing us tocreate a sequential expression from simpler sequentialexpressions. For example, {#NP (#VP #NP}} is a legal expressionwhich happens to be equivalent to (#NP #VP #NP).

Alternations. To match exactly one of several expressions,the alternatives are separated by slashes and the whole isenclosed in brackets (Rules 10-12). The pattern{!THE / !A / !SOME) requires the next word to be one of thethree words, the, a, or some. Once we have both sequences andalternations, the recursive possibilities of HGSL become moreinteresting as in {#VP / (#NP #VP}). This pattern could be atop-level definition of #SENTENCE since it is satisfied either by#VP (an imperative sentence), or by the sequence {#NP #VP} (adeclarative pattern).

Optional matches. The appearance of a subexpression in somelarger pattern may be made optional by placing a dash before it(Rule 13). The pattern { - {!IN !ORDER} !TO #VP) matches both Inorder to form a more perfect union and To form a more perfectunion.

Repetition. Shorthand expressions are provided for two verycommon types of sequential repetition corresponding tozero-or-more, shown by a preceding asterisk, and one-or-moreshown by a preceding plus sign (Rules 14, 15). For example,{* (PREP #NP)) matches any number of consecutive prepositionalphrases, including none at all, and (#NP !VERB + (#NP}} matches asentence with one or more objects in the verb phrase.

Conditional matches. It will sometimes be convenient to beable to insert arbitrary conditions into a larger expression. ALISP form that evaluates to true or false can be enclosed inangle brackets to constitute a valid HGSL expression. Forexample, the pattern

(#NP #VP <EQUAL (NUMBER-OF NP) (NUMBER-OF VP) > #NP)

first matches an #NP followed by a #VP. We must then evaluate thecondition in angle brackets and then proceed to match a second NPIonly if the condition evaluates to true. The condition isimplemented by a single test arc in the ATN. Note that HGSL does

3

• , " ,- • , . -, -.- - , • -. % ". ,- - - . .. -..-1, % - % ".%. . -

not provide any standardized data structures to be tested byconditional expressions. Thus in order to write a condition, thegrammar writer must go outside HGSL, at least in its currentform, and devise a LISP expression based on the data structuresof the interpreter.

These conditions may be arbitrarily complicated andtherefore may be a trap for the grammar leveloper. Using themtoo often will severely reduce the ease with which the grammarcan be understood and extended. On the other hand, a fewwell-motivated conditions may allow considerable rule economywithout introducing any serious obscurity. Our experience withusing conditions shows that they can sometimes be quite simpleand still be useful.

Network Generation

Sequences. Generation of a network for a sequence ofpatterns proceeds as follows. Suppose we have obtained a subnetfor the first expression in the sequence. To ensure that thepatterns specified by the consecutive subexpressions are matchedin order, we need only build the second subnet so that its startnode is the end node for the first subnet. We likewise let eachsubsequent pair of adjacent component networks share end andstart nodes. The start node for the whole network is that of thesequentially first component network and the end node of thewhole is that of the last component network. This constructionis shown in Figure 1 for the sequential pattern (El E2 E3). Eachbox represents an arbitrarily complicated expression. All thathas to be known about them in order to incorporate theseexpressions into a more complex net is that they have a singlestart and stop node as shown. Note how the stop node of El isthe same as the start node of E2, as suggested by the overlappingcircles. The correctness of this construction depends on thefact that the subnets are "one-way" nets, in that control cannever flow backwards from the stop node to the start node. Ifthis were not the case, the net might recognize the firstsubexpression, then the second, then wander back and redo thefirst.

Alternation. To build a network for an alternation we usea single new start node as the subnet start node for each of thesubexpressions. We then add T-test arcs (i.e. test arcs forwhich the condition always evaluates to TRUE) from the variousend-nodes to a single new end-node created for the composite net.This is shown in Figure 1 for the alternation {El / E2 / E3).The lighter arcs represent the nets previously generated for El,E2, and E3. The arcs added to implement the alternation areshown as heavy arcs. Clearly the newly constructed network canonly be traversed if exactly one of the component networks can besatisfied.

4

t "i-J

{El E2 E31 ..... :

{El/E2/ E3 E2

-E EI~ C

.4.N.4

Figure 1 Network Implementation of HGSL constructs.

Boxed "E" and thin arrows are the previously constructed "net of the arbitrary expression E.

5

- - - -f = "_ | .-" , , = r,. . J-! w-. J &.. !6 W mjr LW b

Optional expressions. Given an expression and some netthat implements it, we can easily add arcs to make the sameexpression optional. We do this by adding a T-test arc,evaluated after the first arc of the expression, leading from thestart node to the end-node. We then have the option either topass through the net, or to match nothing to it, as shown inFigure 1.

Repetitions. Implementing one-or-more repetition is a bitcomplicated. Assuming we have generated a net for the expressionto be repeated, we add an arc leading from its end node back toits start node, as shown in Figure 1. This will allow thepattern to be matched more than once. We also create a new endnode for the composite network and connect the componentnetwork's end node to it via a T-test. After the pattern hasbeen matched one or more times, control can follow this path outof the network.

The most complicated network construction is that forzero-or-more repetition. As shown in Figure 1, we take thenetwork of the expression to be repeated and add a backwardT-test from its end to its start node. As in one-or-morerepetition, this allows the pattern to be matched more than once.We also create new start and end nodes. The old end node isconnected to the new one by a T-test arc. This is the exit fromthe network after the pattern has been matched one or more times.Finally we add a pair of new arcs out of the new start node. Thefirst leads into the old start node. Any path through thenetwork which begins by taking this arc will have to satisfy therepeated pattern one or more times. The second arc is a T-testleading directly to the end node of the composite network. Thisallows for zero repetitions of the expression.

It can be proven that the network implementations of theHGSL constructs adopted here are correct. However it is alsotrue that the current HGSL compiler does not produce the mostcompact networks possible. For example, Figure 2 shows a moreefficient network construction for alternation.

IMPLEMENTATION ALGORITHMS

The HGSL Compiler

Here we describe how HGSL constructs are compiled intonetworks suitable for the interpreter described below. Theexpression to be compiled as a network is parsed by the set ofmutually recursive functions shown in Table 2. Each function isresponsible for parsing the structure for which it is named andadding the appropriate arcs and nodes.

6

b. ZI

El~E2 "

N.

E3~

:JL

E2U

E3.

Figue 2.Altrnat newor implmenatios o

El /E2 3 ).Thehighr oe iscurentl imlemeted

but he oweroneis mre fficent

:4 J

Table 2

The HGSL Compiler

function #HGSLC(INPUT FILE: FILE) returns BOOLEAN isSELECT OUT PORTION OF FILE TO BE COMPILED;return #GRAMMAR;end #HGSLC;

function #GRAMMAR returns BOOLEAN isloopif CURRENT-WORD = END-GRAMMAR then

return TRUE;else if #NETWORK-DEFINITION then do nothing;elsereturn FALSE;

end if;end loop;

end #GRAMMAR;

function #NETWORK-DEFINITION returns BOOLEAN isif CURRENT-WORD = NET-DEF then ADVANCE-WORD; end if;RECORD CURRENT-WORD AS NAME OF THIS NETWORK;ADVANCE-WORD;START = A START NODE FOR THIS NETWORK;RECORD START AS FIRST NODE OF THIS NETWORK;if #EXPRESSION(START,STOP) thenADD A POP ARC BEGINNING AT STOP;return TRUE;

elsereturn FALSE;

end if;end #NETWORK-DEFINITION;

function #EXPRESSION(START,STOP) returns BOOLEAN isif CURRENT-WORD STARTS WITH "#" then

return #NET-CALL(START,STOP); end if;if CURRENT-WORD STARTS WITH "<" then

return #CONDITION(START,STOP); end if;if CURRENT-WORD STARTS WITH A LETTER then

return #LEX-TEST(START,STOP); end if;if CURRENT-WORD STARTS WITH "!" then

return #WORD-TEST(START,STOP); end if;if CURRENT-WORD IS "-" then

return #OPTIONAL(START,STOP); end if;if CURRENT-WORD IS "+" then

return #ONE-OR-MORE(START,STOP); end if;

(table continues)

8

- - ' -,-'.. <.' -.. - ,.., .'':v..' v~-. .. ' ,: -.'v' <, -' - ' '-"- 2 ."'-''. - '-'- - S,- ..

if CURRENT-WORD IS "*" thenreturn #ZERO-OR-MORE (START, STOP); end if;

if CURRENT-WORD IS "(" thenreturn #LIST-NO-PREFIX(START,STOP); end if;return FALSE;

end #EXPRFSSION;

function #NET-CALL(START,STOP) returns BOOLEAN isSTOP = A NEWLY ALLOCATED NODE;U-E CURRENT-WORD TO LOOK UP START NODE FOR INVOKED NET;Ay") NET ARC FROM START TO INVOKED NET WITH NEXT STATE = STOP;return TRUE;

end #NET-CALL;

function #CONDITION (START, STOP) returns BOOLEAN isSTOP = A NEWLY ALLOCATED NODE;ADD TEST ARC FROM START TO STOP USING CURRENT-WORDAS TEST EXPRESSION;return TRUE;

end #CONDITION;

function #WORD-TEST(START,STOP) returns BOOLEAN isSTOP = A NEWLY ALLOCATED NODE;ADD WORD-TEST ARC FROM START TO STOP USING CURRENT-WORDAS WORD TO COMPARE WITH;

return TRUE;end #WORD-TEST;

function #LEX-TEST(START,STOP) returns BOOLEAN isSTOP = A NEWLY ALLOCATED NODE;ADD LEX-TEST ARC FROM START TO STOP USING CURRENT-WORDAS LEXICAL CATEGORY TO COMPARE WITH;

return TRUE;end #LEX-TEST;

function #OPTIONAL(START,STOP) returns BOOLEAN isWORD-ADVANCE;if #LIST-NO-PREFIX(START,STOP) thenADD "T" TEST ARC FROM START TO STOP;return TRUE;

elsereturn FALSE;

end if;end #OPTIONAL;

(table continues)

9

• .°q ."." 4 ," " • -" " 4 m e . - ." .' " .%.'-" " :%" . " .

function #ONE-OR-MORE(START,STOP) returns BOOLEAN isWORD-ADVANCE;if #LIST-NO-PREFIX (START, STOP2) thenSTOP = NEWLY ALLOCATED NODE;ADD "T" TEST ARC FROM STOP2 TO START;ADD "T" TEST ARC FROM STOP2 TO STOP;return TRUE;

elsereturn FALSE;end if;end #ONE-OR-MORE;

function #ZERO-OR-MORE(START, STOP) returns BOOLEAN isADVANCE-WORD;START2 = NEWLY ALLOCATED NODE;if #LIST-NO-PREFIX (START2, STOP2) thenSTOP = NEWLY ALLOCATED NODE;ADD "T" TEST ARC FROM START TO START2;ADD "T" TEST ARC FROM STOP2 TO STOP;ADD "T" TEST ARC FROM START TO STOP;ADD "T" TEST ARC FROM STOP2 TO START;return TRUE;

elsereturn FALSE;end if;

end #ZERO-OR-MORE;

function #LIST-NO-PREFIX(START,STOP) returns BOOLEAN isADVANCE-WORD;STOP = A NEWLY ALLOCATED NODE;if #EXPRESSION(START,STOP2) then null; else return FALSE; end if;if CURRENT-WORD = "/ " thenADD "T" TEST ARC FROM STOP2 TO STOP;until CURRENT-WORD = ")" loopif #EXPRESSION(START,STOP2) thenADD "T" TEST ARC FROM STOP2 TO STOP;else return FALSE;end if;

end loop;return TRUE;

elseuntil CURRENT-WORD = "}" loopSTART = STOP2;if #EXPRESSION(START,STOP2) then null;else return FALSE; end if;

end loop;STOP = STOP2;return TRUE;end if;end #LIST-NO-PREFIX;

10

U

#HGSLC is the top-level function of the compiler. It takesthe name of an input file and asks the user whether the entiregrammar should be compiled or if just one of the networkdefinitions should be recompiled. It then calls #GRAMMAR. Sincea grammar is just a list of network definitions ended by the keyword END-GRAMMAR, the function #GRAMMAR calls #NETWORK-DEFINITIONrepeatedly until that key word is encountered.

#NETWORK-DEFINITION checks for the key word NET-DEF, recordsthe name of the network being defined, and then calls#EXPRESSION. #EXPRESSION is simply a large select statementwhich examines the current character to determine which type ofexpression follows. The appropriate function is then called.The functions #NET-CALL, #CONDITION, #LEX-TEST, #WORD-TEST arelow-level functions which actually build single arcs for theroutines that call them.

#OPTIONAL, #ONE-OR-MORE, #ZERO-OR-MORE, and #LIST-NO-PREFIXare intermediate level routines that all call the function#EXPRESSION one or more times, adding additional arcs to theresults of these function calls, and sometimes piecing themtogether to form a larger net. #OPTIONAL adds a single arc towhatever structure has been built for its component expression.#ONE-OR-MORE adds two arcs and #ZERO-OR-MORE adds four.#LIST-NO-PREFIX is the most complicated net, since it buildseither a disjunctive or a sequential net and in either case thisrequires piecing together the nets generated during calls to#EXPRESSION as discussed above and pictured in Figure 1.

The ATN Interpreter

The interpreter used by our system is fairly conventional.The output of the compiler is a set of networks based on theconstructions given above. These networks are represented as setsof arcs leading from one node to another. The interpreter hasonly one major data structure, a stack of nodes that is used tomaintain the current path through the various nets. Theinterpreter repeatedly pops this stack and tries to extend thepath, generally by evaluating the next arc out of the mostrecently stacked node.

The output produced by the HGSL system is a syntax tree suchas that shown in Figure 3. This tree is based on the parse pathconstructed automatically by the interpreter. Once parsing thetop-level net has been successfully completed, the parse pathwill be stored at the top of the parse stack and can beinterpreted as a syntax tree.

Table 3 gives the interpreter algorithm in pseudo-code. Thestack frame, declared in lines 2-8, contains a node id number, Sthe id number of the last outgoing arc examined, the position of

11

(#START(#SENTENCE

(#NP(DET THE)

(NOUN INSTRUCTOR))(#VP

(VERB PERFORMED)(#NP

(DET THE)(NOUN PROCEDURE)))))

Figure 3. Example of parser output for the sentence"The instructor performed the procedure."

12

* ze.-: i

Table 3

The ATN Interpreter

1 function ATNINTERPRETER returns SUCCESS-OR-FAILURE is2 FRAME is record3 STATE: integer;4 LAST-ARC-TESTED: integer;5 POSITION-IN-SENTENCE: integer;6 ACTIVE-NET-CALL-FLAG: boolean;7 NET-PATH: list of FRAME;8 end record;9 CURRENT-FRAME: FRAME;

10 S: stack of FRAME;11 POPPED-SENTENCE: boolean;12 begin13 POPPED-SENTENCE := FALSE;14 PUSH(S, [#INVALID-NET-NAME,1,0,TRUE,NIL]);15 PUSH(S, [#START1,0,1,FALSE,NIL]);16 while not POPPED-SENTENCE loop17 CURRENT-FRAME - POP(S);18 if CURRENT-FRAME MATCHES [-,-,-,TRUE,NET-PATH] then19 S := PUSH(S,NET-PATH);20 else if CURRENT-FRAME MATCHES [-,-,-,TRUE,NIL] then21 S := PUSH(S,[-,-,-,FALSE,NIL]);22 else if CURRENT-FRAME MATCHES23 [STATE,ARC-NO,POSITIOIT,FALSE,NIL] then24 NEXT-ARC-NO :- ARC-NO + 1;25 NEXT-ARC :- FETCH-ARC(STATE,NEXT-ARC-NO);26 TEST-RESULT := TEST-ARC (NEXT-ARC,POSITION);27 if TEST-RESULT - RAN-OUT-OF-ARCS then28 if STATE - INVALID-NET-NAME then29 return FAILURE;30 end if;31 else if TEST-RESULT - SUCCESS then32 if ARC-TEST MATCHES (ARC-TYPE,X,NEXT-STATE) and33 MEMBER-OF (ARC-TYPE, (WORD,LEX)) then34 S = PUSH(S,[NEXT-STATE,0,POSITION+1,FALSE,NIL]);35 else if ARC-TEST MATCHES (TEST,X,NEXT-STATE) then36 S = PUSH([NEXT-STATE,0,POSITION],S);37 else if ARC-TEST MATCHES (POP,NIL,NIL) then38 INVOKER-FRAME := MOST RECENTLY STACKED FRAME39 THAT MATCHES [STATE,ARC,-,TRUE,NIL];40 NET-PATH := ALL FRAMES ABOVE INVOKER-FRAME;41 if STATE = #START1 then POPPED-SENTENCE := TRUE; end if;42 ARC-TEST := FETCH-ARC(STATE,ARC);43 MATCH ARC-TEST TO (ARC-TYPE,NIL,NEW-STATE];44 S := PUSH(S, [STATE,ARC,X,FALSE,NET-PATHI));

(table continues)

13

• .U+

45 S :- PUSH(S,[NEW-STATE,0,X,FALSE,NIL]);46 end if;47 else48 S := PUSH(S,CURRENT-FRAME);49 end if;50 end if;51 end loop;52 return SUCCESS;53 end ATNINTERPRETER;

the parser in the current sentence when the node was reached, aflag signalling whether the last arc evaluated triggered acurrently active net call, and the network path of any completednet call originating at the node. Initially the stack containstwo frames, the node #INVALID-NET-NAME which is shown invokingthe first node of the top-level network #START (lines 14,15).The standard cycle of the interpreter is a loop (lines 16,50)which exits either with success when the top-level net call to#START returns, or with failure when the stack is exhausted.

The loop begins with popping the stack (line 17). Thepopped stack frame will fall into one of four categories. Thefirst case (line 18) is when the stack frame has a completed netpath stored with it. In this case, the path is removed from thecurrent frame and placed on the stack allowing the interpreter toback up into a previously completed net-call. The second case(line 20) is for stack frames which have triggered a currentlyactive net call. Coming across one of these indicates that thenet call has failed, so after the ACTIVE-NET-CALL-FLAG is set toFALSE, the stack frame is pushed to allow the next arc for thatnode to be considered.

The third case (line 22) is for all other frames andinvolves fetching and testing the next outgoing arc (lines24-25). If there are no more arcs, then the current node will begiven no further consideration. If there are no more arcs andthe current node is the one that marks the stack bottom, (i.e.#INVALID-NET-NAME), then the interpreter returns FAILURE (lines28-29) . Assuming there is an arc to evaluate, and thatevaluation succeeds (line 31), the appropriate action is takendepending on the arc. If evaluation fails (line 47), no actionis taken before beginning the next cycle when the next arc willbe examined.

A successful word or lex test will require stacking a framefor the next state together with a current word position that hasbeen incremented by one (lines 32-34). A successful conditionaltest also causes the next state to be stacked (line 35-36). Apop arc is always successful and its processing involves looking

14

% %1

down in the stack for the most recent active net call (line38-39). All nodes stacked above this represent the path whichhas been found through the invoked net and are added to the stackframe of the node which triggered the net call.

If an invocation of the top-level network was popped (line41), then POPPED-SENTENCE will be set to TRUE, control will exitthe loop, and the interpreter will return SUCCESS. Note thatbecause of this, the HGSL supplied by the user must not berecursive at the top level, i.e. the grammar must not include acall to #START.

A GRAMMAR FOR TECHNICAL PROSE

A substantial grammar has been developed using HGSL. Thissection presents the grammar and discusses its coverage and someof its strong and weak points.

The grammar was developed to cover a set of sentences takenfrom Navy technical training materials. Any sentence which couldnot be parsed lead to an extension of the HGSL grammar. HGSL wasdesigned to make this process as quick as possible by allowingnets to be described in a compact, easily-read formalism. Inpractice the system did make improvements easier. HGSL alsomakes it easier to spot rules which are either inconsistent ornot as general as they might be. Far-ranging reorganizations ofthe grammar, while very time-consuming for an explicit node andarc representation, are fairly simple with a powerful grammarshorthand like HGSL.

GRAMMAR DESCRIPTION

Table 4 presents an overview of the grammar; Table 5 liststhe lexical categories used by the grammar. The complete HGSLtext appears as an appendix; in this section the text will bepresented for description piece-meal. One thing to note aboutthis grammar is that it was developed to recognize rather than togenerate sentences. Consequently it would not be difficult touse it to generate some very bad sentences. Also since it ismeant to be suitable for systems which critique poor prose (seeKieras, 1985), the nets should not fail on sentences which areonly slightly ungrammatical; otherwise the majority of the inputtext might never survive the first-stage syntactic analysis.

The grammar developed here suffers from an ad hoc approachto conjunction which has lead to the inclusion of conjunctivebranches in many of the nets. The option which we did not pursueis to build a special mechanism outside the grammar that wouldhave constituted a general theory of where conjunction can occur,

15

J.

Table 4

Network Names and General Characterizations

#STARTThe top-level network and starting point for all parsing.

#HEADINGTitles, chapter or paragraph headings, etc consistingof some formatting mark (i.e. indentation, romannumerals) followed by a noun phrase.

#CSTATEMENTConjunctive statement. One or more sentences conjoinedtogether.

#STATEMENT

Either declarative or imperative.

#DECLARATIVE-STATEMENT

#IMPERATIVE-STATEMENT

#PPCLPast participle clause. Clauses based on a pastparticiple and explicitly introduced by somesubordinating conjunction such as when given aircrafttype and weather conditions.

#VINGCLPresent participle clause. Clauses based on a presentparticiple introduced by a subordinating conjunction.While collecting and safeguarding drug evidence.

#SUBCLSubordinate clause. Full sentences introduced by asubordinating conjunction. Because the procedure wasperformed incorrectly,

#VERB-COMPLEX-ARGUMENTInfinitive clauses following "to" which act like theobject of a verb or simply give the purpose of theaction. He tried to perform the procedure Heperformed the procedure to conform with regulations

(table continues)

16

I.

#FOR-TOInfinitive clauses which begin a sentence To performthe procedure , ... For the students to perform theprocedure...

#WHETHER-ORCompound condition formed with "whether". Whether theuser is a novice or if he knows the system well, thisreference manual will be helpful.

#SUBRELCLSSubject relative clause. The modified noun is thesubject of the relative clause.

#ELIDED-VPElided verb phrase. Clauses based on a past or presentparticiple from which a form of the verb "to be" hasbeen deleted. May introduce a sentence or modify anoun. Given adequate instructions, the students ... aprocedure requiring expert supervision

#ADJCLAdjective clause. Clauses which follow and completethe meaning of certain adjectives, able to perform theprocedure

#OBJRELCLSObject relative clause. The modified noun is theobject of the relative clause which follows, theprocedure the instructor performed

#FOR-TO-RELCLAUSEA relative clause based on an infinitive and possiblyintroduced by "for". the equipment for the trainees touse during class

#COMPOUND-MODIFIERConjunction of noun-modifying phrases sometimesfollowing the modified noun, but sometimes precedingit. "Whether working with a visual informationspecialist or alone, ... All procedures, official orunofficial, ...

#CVPConjoined verb phrase.

#VPVerb phrase.

(table continues)

17

#VCOMPVerb complement. The modifying phrases that follow themain verb and other verb parts such as participles andthe infinitive.

#GINFGeneralized infinitive. Includes not only the lexicalcategory INF, but multiple word infinitives followed bymodifying phrases such as The procedure is to beperformed by the instructor.

#CNPConjoined noun phrases.

#NPNoun phrase. Both the usual noun phrase consisting ofadjectives and head noun as well as whole clauses whichcan function like a noun, e.g. What the instructorsaid was unclear.

#GERUNDA present participle and its modifier acting as a noun.Following the instructions for this procedure is crucial.

#PREPPHRPrepositional phrase. Also allows for conjunctions asin With the instructor's assistance and in keepingwith the rules ...

#RELCLAUSERelative clauses. Either subject (SUBRELCLS) orobject (OBJRELCLS) relative clauses.

#INTERRUPTERPhrases typically set off by commas and serving toqualify a noun phrase.

I.

18

p ~ p _ ~ pD ~. D P 8 aP

Table 5

Table of Lexical Categories

ADJCL Adjective introducing a clause -

Available for ...ADJ AdjectiveADV AdverbAUX-DO Forms of the verb do acting as verb

auxiliariesAUX-HAVE Forms of the verb have used for the

past tenseAUX-IS Forms of the verb is in

progressives or passivesAUX-MODAL Modals such as may, might, should,

etc.CONJ ConjunctionDEFDET Definite determinerHEADING-MARK Characters marking a title or headingINF Infinitive form of a verbNAME Proper nameNDEFDET Indefinite determinerNEG NegativeNOUN NounPOSS-MARK Apostrophe in possessive formsPPCL Words introducing a past participle

clause - though inspected by theinstructor

PREDETADJ Adjective preceding a determiner -all the best

PREP PrepositionPRN PronounPROPPRN Propositional pronoun - That is not

true.RELPRN Relative pronounRESRELPRNR Restrictive relative pronoun - that

as opposed to whichSUBCL Words that introduce a subordinate

clauseVERB Any form of a verb, inflected or

notVERBING Present participleVERBPP Past participleVINGCL Words that introduce a present

participle clause - whileperforming the procedure

19

- -... .. . . .... . U I .w . dr UW U o" ~P ~, ~ . d .V IN'. . LW'

and thus would have saved us the effort of addressing the problemon a case-by-case basis. Such an approach has the disadvantageof making the ATN interpreter non-standard.

The top-level network is #START, shown in Table 6. Sincethe grammar is meant to parse technical prose, rather thanisolated sentences, #START recognizes both sentences(#CSTATEMENT) and any headings which may occur in the passage.Headings are assumed to be indicated by format markings such aspeculiar indentation or special text editor control charactersand consist of a noun phrase in the broadest sense (#CNP). #CNPincludes any phrase that could be the subject of a sentence.This allows the system to handle titles such as "How to performthe procedure". For convenience of discussion, the rest of thegrammar is divided into the following groups: sentences,subordinate clauses, relative clauses, verb phrases, and nounphrases.

Sentences

The main sentence patterns (see Table 6) are#DECLARATIVE-STATEMENT and # IMPERATIVE-STATEMENT. The first issimply a noun phrase followed by a verb phrase, while the secondconsists only of a verb phrase. The various clauses which canintroduce or follow these basic sentence patterns are common tothe two, and are factored off into the higher-level net#STATEMENT. Since #STATEMENT is the first definition of anycomplexity that we have so far encountered, it may be helpful tointerpret it in some detail.

A #STATEMENT begins with zero or more instances of the sixtypes of introductory clause, optionally followed by a comma,then proceeding to either a declarative or an imperativesentence, in either case optionally followed by a subordinateclause. #STATEMENT is in turn the main constituent of#CSTATEMENT which allows conjunction of two or more simplesentences. This net is fairly subtle. It begins with a#STATEMENT. Since the next expression is preceded with a "-", weknow it may also end with that first #STATEMENT. Alternately wecan add one or more comma-#STATEMENT pairs, before closing with aCONJ (possibly preceded with a comma) and one last #STATEMENT.There are two other nets which are entirely devoted to describingthe conjunction of simpler expressions, one for conjoined nouns(#CNP) and one for conjoined verbs (#CVP).

"?

Table 6

HGSL for #START and Sentences

NET-DEF #START#HEADING/

#CSTATEMENT !

NET-DEF #HEADINGHEADING-MARK *CNP}

NET-DEF #CSTATEMENT#STATEMENT

- {* {!,#STATEMENT}

CONJ# STATEMENT

NET-DEF #STATEMENT*{#FOR-TO/

#WHETHER-OR/#ELIDED-VP/#PREPPHR/#SUBCL/#ADV

#DECLARATIVE-STATEMENT /#IMPERATIVE-STATEMENT-{SUBCL

NET-DEF #DECLARATIVE-STATEMENT#CNP #CVP

NET-DEF #IMPERATIVE-STATEMENT* CVP

21

Subordinate Clauses

There is considerable variety among the subordinate clauses.

#PPCL and #VINGCL, shown in Table 7, are nets based on past andpresent participles, respectively. Each is introduced by asubordinating conjunction. Note how the incremental approach togrammar design has lead to an option for conjunction in the#VINGLCL, but not in #PPCL. This is because the sample sentencesso far processed have not required conjunction in #PPCL. #SUBCLis the combination of a subordinating conjunction and a fullsentence.

#VERB-COMPLEX-ARGUMENT is an infinitive-based phraseappearing after the verb. Sometimes it will be identifiable asan argument of the verb, thus justifying its name, as in He hopesto get back to work. In other cases it will be modify themeaning of the entire sentence by giving the reason for whichsome action was taken, as in He did it to better his chances.

#FOR-TO is used to introduce a sentence. Like#VERB-COMPLEX-ARGUMENT, it is based on the infinitive, but allowsthe subject of the infinitive to be specified by adding for and anoun phrase at the front (For the plan to work, .-... ) .#WHETHER-OR is probably best thought of as a complex subordinateclause, since it combines two #STATEMENTs into a subordinaterelation to the main sentence, as in Whether the result isositive or if it cannot be determined, ... On the other hand,

#WHETHER-OR differs from #SUBCL in that it may be based on asentence fragment rather than a complete sentence as in Whetherold or new, ... By comparison, Because old or new ... isunacceptable.

Relative Clauses

The relative clauses shown in Table 8 can be used to modifynouns, or in some cases, can be used in the place of nouns.#SUBRELCLS includes relative clauses in which the noun modifiedplays the role of subject in the relative clause. The mostobvious variety uses a relative pronoun as in The procedure thatworks. One type of #SUBRELCLS that does not use relativepronouns is the #ELIDED-VP, which is based on a past or presentparticiple, e.g. the procedures studied in this course or thetrainees having the most difficulty, for the above examples.They are called elided verb phrases because they are taken to beshortened forms, such as the procedures which were studied inthis course and the trainees who are having the most difficulty.Like most clauses which can modify a noun, #ELIDED-VP can beshifted to the front of the sentence, in which case it is beingused to describe the subject of the sentence, as in Elected forthe first time in 1982, the congressman ...

22

< :'.. ...

Table 7

HGSL for Subordinate Clauses

NET-DEF #PPCLfPPCL VERBPP #VCOMP}

NET-DEF #VINGCLVINGCL #GERUND

* { CONJ/ !}#GERUND}

NET-DEF #SUBCLISUBCL #STATEMENT

NET-DEF #VERB-COMLEX-ARGUMENT!TO #GINF

NET-DEF #FOR-TO- ( { !FOR #CNP }/{!IN !ORDER}!TO #GINF

NET-DEF #WHETHER-OR!WHETHER#COMPOUND-MODIFIER/

{#STATEMENT - { !OR !IF #STATEMENT I

NET-DEF #COMPOUND-MODIFIER-{!BOTH / !EITHERADJ / #ELIDED-VP

*{{CONJ/ ! {ADJ / #ELIDED-VP I

23

Table 8

HGSL for Relative Clauses

NET-DEF #SUBRELCLSRESRELPRN #CVP } /#PREPPHR /#ELIDED-VP /#ADJCL )

NET-DEF #ELIDED-VP- { NEG } * I ADV }

#GERUND / { VERBPP #VCOMP } } }

NET-DEF #ADJCL{ - { !, }ADJCL { #PREPPHR / { !TO #GINF } } I

NET-DEF #OBJRELCLS( - { RESRELPRN ) #DECLARATIVE-STATEMENT

NET-DEF #FOR-TO-RELCLAUSE( - { !FOR #CNP ) !TO #GINF

Yet another variety of #SUBRELCLS is the #ADJCL. It isbased on an infinitive or prepositional phrase and introduced bycertain adjectives such as eager, impatient, or glad as inTrainees glad to complete their instruction or Available to allemployees, group insurance

Aside from #SUBRELCLS, the other major varieties of relativeclause are #OBJRELCLS and #FOR-TO-RELCLAUSE. In #OBJRELCLS themodified noun plays the role of the object and the relativepronoun is optional, as in The procedure [that] the instructordemonstrated. #FOR-TO-RELCLAUSE is a restricted version of#FOR-TO appropriate for use as a relative clause. It allows usto handle the thing to do or the thing for you to do.

Verb PhrasesThere are three major verb phrase nets shown in Table 9, the

principal one being #VP, which generates verb forms. It includesa fairly careful description of verb formats, covering the use ofmodal auxiliaries such as can or may, the use of do, be and haveas auxiliaries, and simple tensed verbs, with consideration givento negation and adverbs occurring between Y

24

U&

Table 9

HGSL for Verb Phrases

NET-DEF #CVP# VP

*{+ ( CONJ /!

NET-DEF #VP 8

* ADV}{AUX-MODAL- NEG} * ADV} #GINF #VCOMP /AUX-DO -{NEG )*{ADV }INF #VCOMP /

-{NEG IAUX-IS *{ADV

VERBING / VERBPP I#VCOMP* {CONJ{(VERBING/VERBPP} ) VCOMPI

AUX-HAVE - ( NEG )*{ADV ) VERBPP #VCOMP /VERB #VCOMP II

NET-DEF *VCOMP*{#CNP/

< LAST WORD IS A VERBTAKING STATEMENT OBJECT > #CSTATEMENT}/!THAT #CSTATEMENT ) /{ {- {!, ) #PREPPHR

*{CONJ #PREPPHR I/ADV/ADJ/

#PPCL /#VINGCL/#VERB-COMPLEX-ARGUMENT/:4

#, INTERRUPTER !!f #INTERRUPTER !

NET-DEF *GINF{ * { ADV I

INF *VCOMP I//!HAVE VERBPP #VCOMPI!HAVE !BEEN (VERBPP /VERBING # VCOMP /!BE f ADV I4VERBPP /VERBING # VCOMP*{(CONJ ({VERBPP/ VERBING ) #VCOMP ) I

25

..........

parts of the verb. Since this is a complicated definition, wewill give several examples of the verb phrases that it includes.

There are basically five alternatives based on AUX-MODAL,AUX-DO, AUX-IS, AUX-HAVE, or VERB. The AUX-MODAL case beginswith a modal verb such as may, might, or should, and thencontinues on to some infinitive and whatever object may followthe infinitive (#VCOMP). As an example, consider must performthe procedure. Adverbs and negating elements can be interspersedas indicated to give must not carelessly perform the procedure.The AUX-DO alternative is quite similar, but is based on a formof the verb to do, rather than a modal. The sort of infinitivephrase which follows to do is also slightly more restrictive, INFas opposed to #GINF. 1GINF can generate have done, whereas thelexical category INF cannot. As a consequence, should have doneit is allowed, but did have done it is not.

AUX-IS verb phrases have some form of the verb to befollowed by a participle, either present (VERBING) or past(VERBPP). The participle can then be followed by the usualobjects. Examples of this would be been performing the procedureor was told by the instructor. As indicated by the latterexample, AUX-IS verb phrases include some passive voiceconstructs. The AUX-IS verb phrase has been extended to allowbuilding conjunctive verb phrases by adding a CONJ and a secondparticiple to give phrases such as been cleaned and inspected.

AUX-HAVE generates the past. tense with have and so requiresa past participle. An example with the optlon--al arguments ishave performed the procedure. The VERB-based verb phrase is thesimplest pattern and captures the present tense of simple verbs,e.g. perform the procedure.

Every basic verb phrase described in #VP ends with a #VCOMP,a net describing the numerous phrases which can follow the mainverb of a sentence. The first alternative in #VCOMP is thedirect object (#CNP). The second alternative uses a condition(note the angle brackets). The rational for this condition is asfollows. Some verbs take whole clauses as their objects, as in Ihope they all get here on time. It would be very inefficient tobegin parsing a sentence after every verb, so the conditionensures that we attempt this only if the verb is one of therelatively few which can take clauses for their objects. In thethird #VCOMP option, a clause once again serves as object of thesentence, but it is explicitly marked by that, as in I hope thatthey all get here on time.

Some of the more straightforward post-verb elements includeone (or more) prepositional phrases and the lexical class ADV(adverbs). The lexical class ADJ (adjectives) is appropriateonly after such verbs as be, seem, become, etc. Currently no

26

attempt is made to implement this restriction, which could bedone with a condition.

#PPCL and #VINGCL are also possible verb complements, as inHe executed the procedure as ordered by the instructor or Heexecuted the procedure before realizing it was inapplicable. Asmentioned above, #VERB-COMPLEX-ARGUMENT accounts for infinitivephrases that follow the verb.

The last options given under #VCOMP are for interrupterphrases, i.e. those which are likely to be set off from the restof the sentence by commas or brackets. Since our understandingof the relation of these phrases to the rest of the sentence isincomplete, this constituent has an undeniable catch-all flavor,generating noun phrases, conjoined adjectives, and at least somesubordinate clauses. Giving a more satisfactory account of thesephrases would be an important next step for this grammar.

Likewise, the #VCOMP net could be improved by taking intoaccount the few sequential restrictions that govern the orderingof the structures included. For example, a verb complement mayinclude both a noun phrase and a clause object, but the order isnot arbitrary, as shown by the contrast between The officer toldthe trainee his promotion was approved and *The officer told hisEromotion was approved the trainee. Currently #VCOMP does notimpose any such restriction. A #VCOMP consists simply of zero ormore items from the list of possible phrase structures. Thesurprising thing about this net is that it works as well as itdoes.

The last major verb net is #GINF which covers infinitiveforms and likewise makes use of #VCOMP to describe completeinfinitive-based phrases. There are basically four phrases herewhich can be illustrated by the following examples: dosomething, have done something, have been doing something, and bedoing something.

Noun Phrases

The major noun phrase net is #NP, shown in Table 10. Thefirst two options of the net are the most complicated. The firstdescribes the sort of clause which can function both as arelative clause and as a noun, as in What he saw amazed him.There is also an interesting conjunctive version, Do you knowwhere or when this trend started? As is clear from theseexamples, RELPRN is a fairly broad class including when, where,how, which, that, and so forth. It is not essential to give afull sentence after the RELPRN to get one of these noun-replacingclauses. An infinitive phrase will do just as well: Does heknow where to turn?

27

- . . . . .~ . --.

Table 10

HGSL for Noun Phrases

NET-DEF #CNP-{!BOTH I !ITHER I !EITHER

( NP*{+ {!/ ;/CONJ ) #NP

NET-DEF #NPRELPRN* + (CONJ/!} RELPRN

#DECLARATIVE-STATEMENT ) /RELPRN * ( + ( CONJ / ,IRELPRN

!TO #GINF #VCOMP I!WHETHER #DECLARATIVE-STATEMENT!OR f MNOT / #DECLARATIVE-STATEMENT}f PREDETADJ)-(DEFDET /NDEFDET*{NOUN /(ADJ -{CONJ II

NOUN- #RELCLAUSEI

IPOSS-MARK #CNP II( !( #CNP !

NAME!/PROPPRN/PRN /#GERUND

NET-DEF iGERUND{-{NEG IfVERBING / !AVING VERBPP I VCOMP

NET-DEF #PREPPHRPREP #CNP

* ( + { CONJ I ,IPREP #CNP I

NET-DEF #RELCLAUSE#FOR-TO-RELCLAUSE /#SUBRELCLS /#OBJRELCLSJ

NET-DEF tINTERRUPTER{ #COMPOUND-MODIFIERI

#CNP/#PPCL

28

The #WHETHER-OR clause discussed above in its role as asentence-introducing dependent clause can also serve in place ofa noun. For example, Whether the procedure is efficient is notcrucial. The most common pattern described by #NP is the moreobvious grouping of nouns and adjectives ending with a head noun,and then possibly followed by relative clauses. This patternalso covers noun phrases that turn out to be possessive forms.Other possible noun forms include proper names, pronouns (bothPROPPRN and PRN) and gerunds (i.e. those phrases based on apresent participle but serving as a noun).

GRAMMAR COVERAGE

The grammar was originally developed to handle technicaltraining materials written by Navy writers. The goal was to beable to process early drafts of such material, and not finishedversions of the material, because the parser was intended to beused as part of a computerized comprehensible writing aid (seeKieras, 1985). A sample of target materials was collected andsupplied by the Naval Personnel Research and Development Center(NPRDC), along with a lexicon containing about 10,000 words,tagged with their traditional parts of speech. This lexiconincludes most of the words appearing in military technicaltraining materials. It should be noted that a large quantity ofsuch material appears in essentially an outline format, withheavy use of "telegraphic" prose. We did not attempt to ensurethat the grammar could handle such material, both because keyparts of the content are conveyed by the outline structure ratherthan sentence content, and because the telegraphic style isprobably inappropriate for such documents anyway. As anindication of the coverage of the grammar, it parses all of theexamples shown in Table 11.

Convergence of Coverage

The grammar was originally developed to handle the targetmaterials in the usual non-systematic manner. That is, a fewsentences were chosen and tried on the grammar. If there was afailure to parse the sentence, a decision was made whetherextending the grammar would be reasonable, and if so, theextension was made. However, we had the usual experience ofparser developers in that a lot of syntactic coverage comes veryquickly in the development of the grammar, but each extensionaccounts for fewer new syntactic forms. Thus, when coverage isassessed in terms of the variety of syntactic forms, further workon the grammar tends to produce less and less additionalcoverage. But if the goal is to handle real material, withrealistic distributions of syntactic forms, is it possible thatthe grammar development process converges to an adequatecoverage? Of course, there are too many possible syntactic forms

29

'f*Z "!. Z

U

Table 11

Example sentences from each NPRDC materials sample

Sample 1

Given the logarithm table, a chain of amplifiers and/orattenuators with the gain or loss of each expressed in db, andthe input power in watts, compute the gain or loss and outputpower.

Sample 2

In order to ensure that all art work requests leaving andreturning to the IPDD are accurate and the requested word is doneto the satisfaction of the customer, the following procedureswill be adhered to in submitting audio-visual productionrequests.

Sample 3 .'

Due to the technical nature of these performance tests and therequirement for the proctor to be fully aware of the examinees'actions and their consequences at all times, it is required thatthe proctor be qualified to teach this course of instruction.

Sample 4

Identify the proper methods of approaching a drug offender whilecollecting and safeguarding drug evidence as specified inapplicable publications.

5%

to hope realistically for a complete grammar. But the questionis whether the process would get to a point of diminishingreturns at a reasonably high proportion of sentences in thetarget material that are covered.

Thus, as part of the final grammar development process, aconvergence study was conducted. A series of material samples -were used, with the grammar extended to handle each sample inturn. A record was kept of each change made in the grammar, sothat we could roughly quantify whether the extensions to thegrammar either increased or decreased as we went from one sampleto the next.

The specific samples were supplied by NPRDC. These wereactual samples of draft materials to be used in technical

305[,

Z%'

P' -.

training. The sentences in these samples had been classifiedinto two groups, based on whether or not they could be simplyparsed by an extension of the relatively simple ATN grammar fortechnical prose found in Kieras (1983). We assumed that all ofthe sentences that could be parsed by the NPRDC grammar couldalso be parsed by the current ATN, which like the NPRDC grammar,evolved from the same original simple ATN. We then focused onthe sentences that could not be parsed by the NPRDC grammar. Thesentences shown in Table 11 are examples of sentences that couldnot be parsed by the NPRDC grammar, but could be parsed by thecurrent grammar, after it was fully developed to handle thesesamples. These examples are chosen to represent the more complexsentences that could not be handled by the NPRDC parser, ratherthan the simpler ones. The samples were used in order ofincreasing size of the sample, which was the same as the order ofincreasing number of sentences that could not be parsed by theNPRDC ATN. The grammar was elaborated as required for each of i

the sentences, and a record kept of how many such changes weremade. Notice that the criterion for a successful parse was onlythat the parser succeeded in producing a parse tree that was notgrossly wrong. Such parse trees may have difficulties in termsof semantic interpretation, but we did not make a systematiceffort to either quantify the number of such problems or toresolve them.

The results are shown in Table 12. As shown in the Table,the first sample consisted of a total of 23 sentences, 5 of whichcould not be parsed by the NPRDC ATN, and all 5 of thesesentences required extensions to our grammar. The next samplehad 30 such non-parsable sentences, and 12 required extensions tothe grammar. The fourth sample, however, had a total of 109sentences in it, 62 of which could not be parsed by the NPRDCgrammar, but by the time we reached this fourth sample, only foursentences required extensions to the grammar.

This overall decrease in the number of grammar extensionssuggests that the grammar is converging to a coverage of thetarget materials that would be fairly adequate. Notice that eachsample came from a different writer, so that we exposed theparser to the idiosyncracies of different writer's styles.Although this convergence study is very limited, we areencouraged that practically useful parsers for this targetmaterial can be developed, and that the grammar presented here isclose to being a practically useful parser.

31

Ae-% A e

Table 12

Grammar Convergence Results

5;'-

Sample in order

1 2 3 4

Total sentences in sample 23 46 65 109

Sentences not simply parsed 5 30 39 62

Sentences requiring grammar extensions 5 12 6 4

Percentage of total 22% 26% 9% 4%

32:

, .j,_

'.

ph

32 ,-

References

Kieras, D. E. (1983). A simulation model for the comprehension oftechnical prose. In G. H. Bower (Ed.), The Psychology ofLearning and Motivation, 17. New York, NY: Academic Press.

Kieras, D. E. (1985). The potential for advanced computerizedaids for comprehensible writing of technical documents.(Technical Report No. 17, TR-85/ONR-17). University ofMichigan

Winograd, T. (1983). Language as a cognitive process: Vol. 1:Syntax. Reading, Massachusetts: Addison-Wesley.

¢,

33'

I,

..., .. ... ...., ,, , ........ .-. .... . .. .. . ., , ,,, , ..,... ,-., ..,° , , ,.. , .,,.... , . ,,.... -..'.

Appendix

The Grammar for Technical Training Materials

NET-DEF #START#HEADING/

#CSTATEMENT !

NET-DEF #HEADINGHEADING-MARK #CNP

SENTENCES:4

NET-DEF #CSTATEMENT#STATEMENT

- {* {!,#STATEMENT

CONJ# STATEMENT

NET-DEF #STATEMENT*{#FOR-TO/

#WHETHER-OR/#ELIDED-VP /#PREPPHR IN/#SUBCL/#ADV

#DECLARATIVE-STATEMENT /#IMPERATIVE-STATEMENT- ( SUBCL ) )I:,..

NET-DEF #DECLARATIVE-STATEMENT I.NET-DEF #IMPERATIVE-STATEMENT

* Cvp

34 u

SUBORDINATE CLAUSES

NET-DEF #PPCL{PPCL VERBPP #VCOMP

NET-DEF #VINGCL{VINGCL #GERUND

* {CONJ/ #,}GERUND

NET-DEF #SUBCL{SUBCL #STATEMENTI

NET-DEF #VERB-COMPLEX-ARGUMENT( !TO #GINF )

NET-DEF #FOR-TO{ - ( !FOR #CNP I/{!IN !ORDER!TO #GINF I

NET-DEF #WHETHER-ORI!WHETHER

#COMPOUND-MODIFIER/# STATEMENT - ( !OR !IF #STATEMENT

NET-DEF #COMPOUND-MODIFIERI !BOTH / !EITHER

ADJ / #ELIDED-VPI* fCONJ/ ! IADJ/#ELIDED-VP I

RELATIVE CLAUSES

NET-DEF #SUBRELCLSfRESRELPRN #CVP /

#PREPPHR /#ELIDED-VP/#ADJCL )

NET-DEF *ELIDED-VP( NEG I* ADV

#GERUND /1VERBPP #VCOMP )

NET-DEF #ADJCL

ADJCL I#PREPPHR /I!TO #GINF INET-DEF #OBJRELCLS

-IRESRELPRN # DECLARATIVE-STATEMENT

35 .

NET-DEF #FOR-TO-RELCLAUSE-{!FOR #CNP }!TO #GINF

VERB PHRASES

NET-DEF #CVP#VP .*.

*{+ { CONJ /!

NET-DEF #VP{ ADV}AUX-MODAL -{NEG }*{ADV }#GINF #VCOMP}AUX-DO -{NEG }*{ADV IINF #VCOMP /-{NEG }AUX-IS *{ADV

VERBING /VERBPP }#VCOMP*{CONJ {VERBING / VERBPP ) #VCOMP }/

AUX-HAVE -{NEG I*{ADV IVERBPP #VCOMPVERB #VCOMP} I

NET-DEF #VCOMP* #CNP /

(<LAST WORD IS A VERBTAKING STATEMENT OBJECT > #CSTATEMENTI/

!THAT #CSTATEMENT ) /- { !, ) #PREPPHR

*{CONJ #PREPPHR} /ADV/ADJ/

#PPCL/#VINGCL/#VERB-COMPLEX-ARGUMENT/

{ #, INTERRUPTER!,I/'{ {#INTERRUPTER !

NET-DEF #GINF* ADVI

INF #VCOMP /!HAVE VERBPP #VCOMP /!HAVE !BEEN {VERBPP /VERBING I#VCOMP I/

!BE *{ADV I{VERBPP /VERBING I#VCOMP*{CONJ {VERBPP /VERBING I#VCOMP III

36

NOUN PHRASES

NET-DEF #CNP-{!BOTH /!EITHER / !NEITHER#NP

*{+ {,/!;/CONJ }#NP

NET-DEF #NPRELPRN* + {CONJ/!} RELPRN}#DECLARATIVE-STATEMENT )RELPRN * ( + { CONJ / ,}RELPRN!TO #GINF #VCOMP )

{!WHETHER #DECLARATIVE-STATEMENT!OR { !NOT / #DECLARATIVE-STATEMENT*{PREDETADJ )

-{DEFDET /NDEFDET*{NOUN /{ADJ -{CONJ I

NOUN-{#RELCLAUSE/

POSS-MAR( #CNP}/!{ #CNP !}I I/

NAME /PROPPRN/PRN /#GERUND

NET-DEF #GERUND-{NEG I{VERBING /!HAVING VERBPP I#VCOMP

NET-DEF #PREPPHRPREP #CNP

*{+ ( CONJ / ,IPREP #CNP

NET-DEF #RELCLAUSE#FOR-TO-RELCLAUSE /#SUBRELCLS /#OBJRELCLSI

NET-DEF # INTERRUPTER#COMPOUND-MODIFIER/

#CNP/#PPCLI

37 .

0l 0

C3; w. 14 n 0 414) . 4 4) 0a

* ~ ~~ ~~ 0 ~ 4)1-4) * 0 4) 410 U4 AIL

I4 Vnw 0 .0 14 4) - Q 0 m 0

4) >7 ;N W w 0 N >0 0 -. c0- .4A04) Z r- '4. - 0 04 In)C 0.41 00040%C M N 0 mcN C ON .404) -. )M 144 N 0

C1 414 04 V N C1 4) 140.4 r.4 . , 0 4d4 -

C44M IN 0 4 41 4.) 4 U). L,1. 41 0 N 4)C 4 0 0.C 14 r 4 ~ 41 0 4 4) 1 ) 4)C) 14) w. 14 14 C

.0C00 0 a 0. 14 41 O 4 411) 1 0 41 0.M 0 0 '0 4041 *...c 14 0w1 0 41-4.40 0) go C 1 0 r-1 u CO Xm 0'-~ a 4 4 u 01 04 Ino n1 4L 0u ~ 414 Cl C )4

M 0. 40 144 1 4-4 . w . -0

a, u C0U3 m 0 0) o 40 4C 0 u C a, 04 14 0 4? >. 0 I00- U)aC1b4VV 41 -. 4 0401 c 4)0 U444C 44w14 U) Z )4C 401 4.04 Cm .a * 0- E ow -101 C-0 0 01' 1 V 00 E A)41 .C )0 3 .l0 0 Z 0 0' 41 04041m4I

w 4)0 0 04 fa . 41 ~- 44 uC 4104. 440 O 4 0C 14.U4m0 444. in4 .4). 014ty 041 M..4 0.)41 3 4) % 4.-41. I..4). M~

0.C 4U4 *n 0 w. C40., C1,) 3c . 0.0 1 .00). U1.4

1.14)4 1.410 0) 14004 01 a-4 U)14- 144 uO o4.4 aJ. .11 144

041 A w 0 0 0 to1

0 4)

041 A 004 0041

))9 IA r. 44 N wA. 0 0 ~ C W.141u) u) m 01 04 0) -00) 4) uM 0 41 C 4 0U4~ 41 ~ 1 -0 0e

0 0 :,, > 1 aM 0) A )t 4-4 01~ 41 -I 14M 0 f % -M0 . X90419 4)0 CM N _ 0 i A 0 4) 0 U1 1 4) -1 a -.

u4 DA C0044 >V'.. "-- W 0 a4 w I1 C r- %0-4Mc%

0~~~ A0 w))0 4L 4 1 .. 4 at -4 , at. ) 4) 00 C M 0 I0.0~3 40 01. N C0 41M N0 0A0 04 0 ~ 00 ~ ~ . ca en). W.4 14 C-1W 4 ~ 41 0 .3 4 44 0 40 0 C

.4 .4 .04 C 4 (A1 41C U4 N1 0N M4 0 .4 1 44) 0 14un1V 0 4)7 4

410 9. 04 a% IO 10-4 10. 0 - 0...3144 0 0.- 21 .4 .0 V 01 0 444. CO C-. DC. 14w wO. L. r- 041-0

14) .0.4 14)A.1 04 .4 C4) 14 1. 4)l--4. 00-4- 4 0- 0 4))..-4.414 4) E4C41 . 014)1 C .0 414 4044 . c, . C o41 C. 010.. 0-10 411) =4

0- 0 .6 ) 0414 40.41 W0 CO U 4 0 10 14 -F. 0 41 0 C4m)4)0-. C C41m-.40.z

14 di 4 4)) 4 fa 1.C 0 410 -40 02 V U 0 c0. 0000)4O 0 C U I 0M =1w4-4044 a41 0140 %4 1) 0 .. %A 01 4 4 lz C c ouC 4 ) to P- 41 '24) '22 0 .4

... iM.4 0. 0 0 -1.14 V.144 * )0 * IO4.. .41r.C0 A 04114 w 44.- 4 "4 w (.1w 14041 0 W.- M.0 14. mg1 14 C A04

ODZ400 00-w:. 044U24 C0omm 020 0020 0 2." mu ox en. at-0.-I

0.

41o 41 Ai 4w >0j C 41 4 0)4-

0j C O 0

0 41 n 4)0 U. 4 4 " 4 ) C A,0 N. 44 m ) 0 41 0 n U) 0 L:L

4) 4 A C C N w ON O- V 0 0 0 -'4 C CIO LMIA Q4 0 -. in4 u.If C '4% 4.4 * Nn 1 4 4) C1 4) IQ

ON -44. M UI4 4 )00 0 401. 411) w1 N) .. 41N pC) '4.e 40 N 1 igW N -.. 0 0w 0 -- I) 04 Cc..I ).4i O P ' ) 0 4)

0 W -U640 144 N 10% to 412N wk3 u144 CU) 6.1 z)4 w4 41 z 41 04) 0- " i z i4 m)I 411- 4. 1C; w 0 M1.4. I .4 ME 4 ) 41 j - 0

14 C 41 O4 ) goN 414141 C~ 01 C4m 4 4) 0 0 41410 -u >44 04 4)-1 41C 04 40 10 0a1 141 20 44 0 0

Av44) 4~ 44114> L 0 00 41 t 4C 0~ -41 A 4C- *14-' C4h1)a0 0-01C0 440 w* 14) wC 40 0 a" 0 -441) 041 00 4 2 040 CA - I".C-M)' In iddo04 0 4 4 0 -0 a% 0U4 e -a 041. VI -4)0) C 0- 04)0 ON C NZ0)W 4 r..414 MA4 00. 1 0. 40 C CVO14C A)4.)4141 41. 4 144 0 000- 14 -. )

C)14-.441 014.1 ).1 c4 4) 4 044) 01- 4 400 "o 4).4 nC x 1). .- 141411.204104" M.=4) 1414 414) U04 14> 04 0011 414 C 0 .u-4n 4) m-

)0 04 1 )C10.-> m11) 41CO 4)1 414 >>, C 0 410 142 V:'Id~~~~~ Q421 '20C A1 ZO 0)41. Z -' -14N. M.41 0441 X~ . Cw U C 4 w1 144.4 *) *0 4 w)4 0-4 ODI.1 a0* 0w C1 cc (-)Iw )w

0 M..04 004W4 a2 01u0 0Z-. MD0 MN- 01.20i. .40 a 1-0.l U 00ZaU0

- 1 C D'.

In 1 w I. 1 . .C)0) 14 40 0) 04 1

14 041 04 COf 04 14 4 0

41 41~ 44 a 1 '4 41 1 z C V-.4) u41 411 0 01 > . In CO)M M 0 4 4

0 4M' 14 4) N .01 -0 41a 0 . 4.-. v4m 41- 410 N w)1C.Im. &0 M v4 wI 01m N m40C N L.) > 4) " iU m 41 00

V) I 40 41 :1, .1. >.a a - 0D >4.DI 0- , o 4N V4C4 M P) 4a-.). a,CO0).1n 40)uN 14 41- zN 41.4 N -c t 4141J - 4) N- 41144) 0 044 0) ) c004)ai- 411 w 41 .4 40 as a, a >N 10 wC 014 W1- 0 m 400. Mm41 *.N u000 -4 MN - C 04 a * X C 04 4141 004) 0 a 0

.. 0410 Ch. fuC 41 41. .4 C6 1- I4-0 1 > 4 .. 444 0)--- c

014 41 1 CCI.)6 40 C in 0 Z C 40 a, 1m000C 4) .4 MO. C 4411 41 0 .. 4140-. 4 4. OC.*4 C - 014 Uj .0. . 0~- 0 )

C-.4 W A 040 1 -1 -440 - c4 ) 1 I0 C41. 1440 6 C.01C ) 41 W 14. 4 14)0 C >. t C 114 414 0 0 C1.414 04vj f. 1.- 441 .01 C c 5 U4E.4 0 a0 0.m1.5 -) I.,

4)4W 4.C C 02 V '0 0 A-I- 0A 0410 .4-4 040. 6JO 4I WC - -6 - U M w" V14)4)0 C- C 1.2 iw N .Ij 0 '244. .4 4144 C 1 4000 41 C 041M44). 0.-. 44141 b. 4101 w %0 41 .C4 00141 -4 m 4 > ma

*0.%.-u 0 4 0U >.- *) 0.00- 0 a014 004 *.a 0 - -- 0.1024~ a4 1404Z11 14414 6 1441W 8 w 1. w W W41 UM 410- w 10 4) 0 0 00

f.Z~ 00 0Wa i 0 0c 0000. I.-4 014 U) 00000 m

MN9

V 0- 04,1 m 4

m4 0 11 'a6 4 0 6 14 C)4 1 1p 0a (14 %l A1 0 . W 'C fn. 0c 0u 0 4a:

1. .fl N u 0 4- %a N 4, V% N4 0 a4 -01 0 5 W04 N -2 -C -1 N 603C 1~ 11 v 0 600 W6C 141 'z '-C 4 g14. >N 0 V CC 2 0 >C4 N) -, *N . 1 4,z 4,61 V C, U0 C14

0 4 N -C 1040 14C w I- 'OC N 000 N.' 1.6 , 1. 4)C~0 0 >4 00. ~~ N 4U0 A. a , 4, 040 -. q.

4 ,4A 4) W 4 .4.V' U.6.J 04 NIJ% U N 1:4. 04 >. 141 Z

a,)' 4, U c~ N 0 0 LE 0~ 0u 1-A414L IT - -4- '- 4 C 0. . 1J -. 0 A C- v4 4,. 13

W~J z 4 0% ar. 4, :2 mm co p a fCh 4,14j0 4,4 0 6 C 0 V

r1 0. U0 0 00 c 34 01 u c I014 Cb. 01 Z4 u Ac

40.. In w w &A w0u C 0 4 4a0. w >,- a-C 14Z 00 m 0 04

* w N 4 0021 a O 0ZA.34W ZhQ 04 .~O Q a -0cm> a j%U 0la0 0 C~6t- U

60 Q

.40 z4, i0, 0, b, goo 0

*' Nc .4 01 ' vn - OC w 6 )Cf C in en 2, In01 14 U 6

4, ~ ~ ~ L NN 6W O~ .'J.4 0. 0 Q V O 1 c 1'~u'~ a C C 0.- Id4- AN C00041 ~ 0 in 101 V0 N 61 -

2124 0 >-. ' -0% A C 000 S.- u old 4,30.0!W'- c I Cn Z..> Aa0 64 0 0 40 A.L 600 0 4.4. 0 0r~ Ic 4j

'0e C40. 014, *44J 1; 06 0. a.1- 1 mO'. 0 ~ x ~*J4 A. 4 4.C C I-- C S I~ A &Cc W E14C .. 14' Id 0 4,1

14'.0 a.-. C 4, 00 14 las-U 'a . 0. .4 MZ- -604 oa' z 0*V, 644 41 54 c0V60. 40 -401.- a -U O w~ : eaw c~ 1460. .4W C c144 , 4.. u 6. -- 4 0 .9 02 A,'1 0 h~

>o0 41 2-4. 104 0 " C0 1O 4, rU 63 41-a CL U 6 4

'4C'14 *>4

M-4 0. 4 %. 4 a~A 014 Q- 14 -. V U)". u *uE036 c06-Q 0 0 0UI 0U 04' OM.A A 144 402 ~. . 04

N

CN 0 2 ,0 N 0

4, 0a 0 C 10- In0 A C ,"

-0 0 06 C '. 0 0 .4 I 10 6a 01u4& 0.4U14 4, f..f w 0 in0 v6 CU '- 0

4 fn &...IJ C. C-F . 16 , 8 4, C04 C- -'A 0f 4 V' N 120 0 A.. U n L In 14 '

1414 c *VN j N0 0 Q..0%1 0 82 6ON C.0 Un A0 14 0 '0 CSfl4 - 04 1u0

0% at~- 06 .311 LU 0 e- x "

0 4, A'S a 0 C3 ( U 0 0 0 0% 0 0 - C U. W- 404 W 4 &.I 0 j % .- 404 z 00u X 4

UM 0 C U4 x - 6: 0 g 6-. -. -C A,> b.46 1 . Cc a 63 0 wonh14. C- C b. ; 0 41 >..0 6. a6 . '

00.." -CO0 00 UF3. 6- C 3.14 0 :.0 !0 0 4 0o-4 V0% W-144 &14 4- . 0 a *0. -a 1 0 r- -. 0 C

-A N 60 4--64 MC 6.4 40 1 CO 0 24.) W 4,4 a4 aMA.U~h2 000. 0- 6-. 0 0.. IC; U-! 0 0.2 . 0..

1- 46a , a - F. A.

0 C 0 3 '1 0 N 3

-. 4, 01 I-63 -0 1

0 0 0

14~ ~~~ z. V 001 a%6% 1'1

.M 0 C .Q1.1.3.3 CIO Cee 4V 40 6- 143.144,0c 4,40 , 02 014 64- * 0M0 0 ". 0C V 66 >.I..11

cC -4 N 06 C6U c'60 c4 3. 0 0 4,4P3. C3u 4, 9.J 0l0 CU '- A 3C -.

1140C 14 VA u . a 16 - 0 140 a "., 0 0 'a 4". 06u.-44cU %0 U4 u > 60d20 L 4

U1-1 =a. C. u 0..4 a64 A 00a L4 - 4, v> 0 00. UUu VC2 0' >.Z C '4 4.

C0A. 0 V . 44 a0 303 V- 60 -C. 0 V4 I-. - or- C 14

6401 14 Z4U 60 4 r.0~ cC O44,, .14 4 N v .- O

930 6 M C '.3 0J*1-'- u.~ c' 1441 4,61 0 'nL S Z .. 4,r

-4013 2 Q.0 4A C 03 - .62 -'.-6 !- .043ga.

0 41 UV" ), W4'. 4 '0 - .' 3.- > ) '0 '0 '.0 4,.' 0 440 w4 C Q, 04 % 004. CaO'- '-U6. Z MU C6 OUU.-.Z 00.N 0 0 39'A. 06- 14 2 Z 00 .4 0

0 2 .0 0 1 0l01 V l D 1 11 4n 0 w- nCC )) 0 > ON ! 1% C n r 0 u 4

04 wV 04 0% C 0 wI Cj - jm n a

A~ U .04 0 . 4 3 fn (n 4 .0 s N Cm .0.0 .~- 0 n C01 £ 4 0 w N 9) . 4) C 0 CY 04 w

0.3 c1)r cc) = -4~ 440 w CD 2 ) wWC " o .k Z aaJ0 . C4)£ .. 41))..40. )C 4m4e4 ) *4 = UW).) j .-.4) 3 c .4)= 3 - 4) "j >. x '1% w 1 >C 0 u N 4

moo 4)o *. 4 4.41 -4 mm ChM ) 410 M :C a= C4) aA- >I- W'4 C Cl U)0 C01 443 )1 . 0 0 N 1-4. * 404

m.>- 0)C 0 r.)O4) £ .401 v m sw A,. 0-4n Z r- )4)m10 > 1ID43C.- 4.40 in A%.1 0 040 44 c0 0) 1 X" 43-41 0 -O4-4- v)4 Cq d441j)K o- C4 C V3 144. A)0 0 -g 4 c 0C t p 0404414

U4> C C-0 a W a x o 441. . 41 CO 040- 3m 0 4.3 CZ3 01 N r- 4AM o0 c 0.C. Ed 0t 0 m0 0 4 A 1 ~40 1- uZC ) 413- a0 444)

b44C4 0 c L. - old 40C4 1 434 v N 0004 43 0 044 O W4 -44 W= .4oonou 00 b o m- z aoo-zz z A inUA. 8o u go a a 0t 0 0c z~ a U) a(-0.

.- 43 >1CC 0 41 M w

U1 4) 00)4 414 6. -4

- .04 W. 444 4 >. fu v 4 N41U41 01, V%4 -4 OW -4 w 4u >N4 0 OC C 0(ON 0 - 043

W -) L.4 oob 440 w 4 0 0% .0 0 4i4444CC 0 0.0 0, cCC 0 00 4) 4 10 C NV

Ai Ci 0 0 - 44 a 0% %D 411 N.J 0

11 N ' 00 1 44 4, U w cou cmeA 0 1 J N C 44

*-u 43m 4 4C0% *0443 414 m 141 m) U441, 04u, 1 1.41-l~irs 0 10 0e 43441 aX 00 0. - -444 4.' CU

(do.,0 moo 0. -A 401 41 -4 >. m' 0 0. w1 0144 U C 4)43 C 3!4 4 4-W 40.4 4433 04 .01 401 0 L) m 4

0 . -4 46 0 CP 0 'z 0 03£ 1=0.a-~~~ 10 = * J1- - CX 0 - -' 4 >4. 0 41£w>..)1c a C.m 0 64 0 U; w IV14 =C (A 0C£ 0.-4C 444-43414

A-CC1 JC z 41 J4-- 41 Aj44 £C .- 41 4 * a 43 0. 41 44 00 m1 Im14lo

- 1 C A C - 1 0 a 03 I 4) r 1 4 40 - W M j J41 41 44 41 0 41.3 h4141 0 641 01 043D~4 -4044C 4) a440 44-C0a£.43 C 434JCc b. 1 aC 045.4 Ai44~ > £en NJ. > i 41 -4A3C 0C%0-.

0. Om *- 0.N- *0 CO *.4. C 4 0.4 * 0. C 4 .3.-- A0 4 41 .4w .%..

00 03 0CDU 0 W% 41N 0~oa 0~t % 0 0 a 6> N%,m

01 01 ale 4

0 04m 0 .c T.C 0 4) ''

U) U A£ CU c>43 0 v Cl)

00.~ 43 6 OC £4 00 ..£ .. £ C 4-j 4)

'- in ) 04! 40 4 4 0 jJ I41 43C C44h. - 4 =£A0 010. 99> v 1 N 'A

a3 &0 j0 on 4410 0 Cm 440 w C0 C) 10941fin 43 V0 w4 C

c .4)1 ". 4114 414 a~C I .34 434>,3 .. 4U -C 0% 000) 4444 4 . Z 41 Mo 43414 0 3 vf -411 -4 I., m w IA-=IC aw --1 10 I

44- 0 a43 41 - m1414 Ai 1 03 so a 0% >, 4341lo01443 041 >04s Ci 3 34 44~4 > .44 43 Y0

a-.. w4 0> x~4 41 ON zcc J 4 0- u 4%4301 N) w63 C0 X0 i 0 w43 024 U,- 41C .44341 "

4j4j4 ~0 0 44644 w0 414- 10 * C440 c %C4409 10,0 V03> &J C V.0-441 444 C40 C00 43. en Im14 0 > & . *. 4 0 444 43 400

A41D4 00 , &3 1') 0 N 41 0441.4 14J 4 00 WO 4 3C4 0C41C41410 40C c c W-Iu w444 -414&v Cn 0 -

00 £11 .. i. .0- .1£ 'C 0- *C .444.I .4

0 OD 0. o-ccl" 0 w 00.- 00421 0Cl#f 04410 0

£ >. 14

£0 44 M- 034 a1 0 0

430. 0 4L tr C£ 0 43 0 InW 0b.>12, 44 b 3 N 0 43a kD .0£44 4 ksJ 44 f" 0- %.4%

41-- i I0 *4 N. )4 l34314 4341>.(flID 0 4. m 44>440 4 3N - CCWOaW 0- 'i .C 41 0 Ln 0 .V) 0 41.1 w >00 L-4 0(a 4 %

*CG .43C. In444 n Y 0Ujj - 0 44M 41(A 01C 0J r. 4, 0%l u v M,-4 1 0 a43 b.-4 0 ~ 9mc -0 n NV C 0 i u

-44.c 9 % -4 In. &.4I40 0. I Ad C qw Ol)0d-4C0 zJ4N C C- 43 41 0£- 41.. 0 43 43 ow (

4 ) 6 j30 'C 4 C. - 41 n4l) 43

W:~4 'a0 >4£ 4 143 41 11400 01 z-1~C~4 44C. C .- 0 441L) N. Cl)MM .C4 I1C I .dI44 ' 0 M40 1 5 40 4 3- 4 .. ) b. CA .e A l£ * I 0- A 3 0 .1 0

44-40-4K 4 I I 444 444 4 44344 44 4241 ) 44 At4 nu u A0 w 06. Aj rm%. 0-14 000.0. 0 =4.3 0 >

C.~~ *. . 4n2 -n&4W ,. *u x 43 '0 ~ * .

00 4) on

41 01 0 )I . -

o1 C r. CuN - wc4)0 40

w10 GD 0 C D C 4n ;k- 0 "-' . w4Q 00. 4), N 'n .4 4 a u 0 %c - 01 NC" 41Ic' Ic 0 4 41 CA ItC 6 .f C a>- SN >.0 .4 o 5a 1o SO m % fn 3.. r-' VI_ fn MOO -C 1 0 .UWW L)cM1 C - -CU " 0= 2 .3. C% . 0 N

C C '4040 0C C u I C 4) M O c w 4 3-.4 M 0 a, u 4-'. 0 ON 10 0 : .- )r wC. ON ;C. u 4) 0I4) W. C 0 0 . -

S' 00 , d).4 0 4 '04-I0. 4 0 .C 04) 'o .I 00 C w40 1C 1W u.i -~ ~ X C.. 3M 0 0.,41 0C 0CI v - ICmCV

'CC w o~ .4 ICfO C4..c V 5 s oL 0 .4)1 0 C 04 jw ) 0-M o- .CC3. 4 3W UC a, ES- log Z - OS o-CI 0%. 0 - g4) ca.. 04)'r 0.1w aM . ca 4) r. 0c -0 cIM CE M a cc In> ,cC41v u nM .C. N 4) ZICO4)0

> Ac4) >. 90C L0£ 0 * o U "'.1-.04.-a -I-S. - - w w v)£0tIC -. 0I Ad mC .IC-.C - e' -) 4 0 -) ACC u w. 40 -~,C CW O 0 EE L 04 )£6 0 in 34Q

v).0 C 3. .. *-. C > U n -i C r-O a. C -.. j ) w Ln 4) a C-C--- 0*S UO 0 & w a~C .Go~ w I a 0a1.0 O k c a w 0'I w 0wa

4O1, 41 N)4 -clIC 3

0 Ac 0 4 0 =3 0U 30" . 0U) 004j4 0 >1 1. 0U ~0

A. 4 c 0 -C I c

0 ' 0424 0q 4110u) > C 0 " 0'E'C2% -C oUo > 0'- 'cO 4

0COON 04 I2 C .IMN 1 0C C V 41 I3 a, .a MC a 1. -M'n 1% 0'.i MEI *Cu zM2c0 w oU 0 % 0% .' Cx 0 0

0.0a4 00 0% an6. 4 c U z114 a£C-I a n wE -U >.a j.u -. % C 0C'U. 14 ZO N a.. C-c- A4>04-w4.CL c0 4 4C v C * 4 0 a 4' C N

4)41 C0C No CU " M 40a 5D. as 41 a C D. 0. aC5. 240.1 McLamc402 . .3

COW C- W c v4)j4

=0 26 10.4 0 0 c 0 CI O.12g.. w C 0 0)1. 4 .C o U o j - 0 a 0C) c C3. N >- 4

oo410' r.4 f"0 .C 4E.C jjC &C)J 41 C -C 0' 01v4cb 0 NECj 0 CE 0 050& 0-41 ICE2 -.40 1 I aW 14 OO.. 4)4 w

M0316 02 6 26-00 E O 1 34 3 3 0.-. L) wEC41 Mc 4. MZU

MOO 4 ;)040 1.0,. o M'ZC C4)C O4) 0 0 14 Oc0O 0 c~ N-e 0.J-

VIc c v-04 - 04'h1Nc j U9 0 t- 0 .4 & U0 c2 0 u2 0

z 0.0

4 0 5 3 . 0 bAE. c 0 0 - a jl41 0 4) 20 0 c0 0 3 OC o U 60b C vw 4,-0 W,

C.- -0( N 34,6 1 -) COL Ij -U>i Q . aa )0 4Oa OS I) 4 a4 N C N a '.0 Ma I Ca

MC N 4 41m C. 0 C 041 4)414~ C, C O N U MN M C U "

30 i~% 554 C I N 4) 41. ' 1 E, C O- C

0 OA 04 CU gI 4, zr 'C z2. = wt' vC nIZ - 4a- u ,0 -I D.4( 54544>,IjI 0' 14 a) 0 of"45 1 u 5 2-. N a'-''C '40' '0 "' COO'4 a - 01 c EU 0'020 4) 00 4) 0

cJO M 30 0 z S. 4 KC C C k% a. S- 00 vJ 5uv.11 1% 3. 4) of ":..4E 3 N j 61. -a CC7

OCId C %1 44044 1 g 1CI 4114 'CS- C- u !J O 3.Uw4o1 w00.:: C 3M 0 002. .1D- K... C . 04 *'~"0 -OU 0-.4 0 EM

,j 0 IC A "0 C C 04 0. zO >. 4j 0- oU a.MC c 0 C 0 a'1 4C64j MIg CU0 4). a36.41 xc 00 4)0 34) - go - - V--%ME % go- MO 0 41 MOO MC . 112C EC7 M S M M£I.

>41i: a C IC 0 C C IC

am ' 40, ;ri a-' ,n 41 CXU Q .)o 041 3. OwI00w D 0

~ C 0 40 0 0 u a wo 4)0~ ..0 0 54 0n U 0 0c 0 2 0 Z e~~

o ) 4,- 4 4, - 4) 4- 4) -4)- > r)Z- - - 64. 0 0

*,Aj6 0.E1% w N 0:-4 Nn 0OOUCOW.*4..M C

00 .0.'o 064-I .40N U) 0 4ON N 0 .( ( 0 U*"0 q 40 * G4 V, @4 01 0UrO. Z 4) - 0 41

> 4 4>114 0 0 00 4 MAC.> ..0 U , 0 P0 4 4( > Me4U> 0 N~o U6 > 04 (N u > M40 N4 u A w U 0 004. CC0) *.nc z W c Z W C z -W r ' 0 0 1 f 0 C 0 0 -4M 4, 1. 4 00 Cw 00- - 0 -- zCc 4 r0A - c 00c C0 2 co 0 at0000 0 o4 0 0 0 ) a)1o 0 u U) -0a 000 00.) 4 4 ) I 0 -1 2..1 01 04 c CL l -.*4- c c 0 0-

u a @440 u0 0 c ..-44 0i 0 2ic 0 m-C0 -40.0 Z 0 Z -. C c %-- U 1414X04>0 4 1-00 14 0 1401400 -.40 00 .0 4 " ' u464 00M 0.0.4C uocc-4 0 4 O0 04 0 O4 0 01C,4-0 0-40 >1 0 W 4 4 >0 1 40 0 6-.'Dw 4 W4 0 W4 44. 0 W 4US0 44 0k a.~. 0~0, M a w a) O 4 40 L C

0 as0 C oc 0 44 0 D o 0 0 4 0 OD0Y -c _ 0 03 02 040 GDc n. 0200LO

0014, 0

a 0 0. 0 0 U C0 % 0 00 w-P43

40 1 A cUV c 1 c 0 D 0014~ C4 4 1 0 o 0 CO '

10 44 0 w .4 w t . i1 41 1 4I 0 4144 0 .4 .4 N u0 N- N440 .3 a ua a

c 4 0 g M a N ;0 0 4 r 0 14 -~ -GW 44"40-z '.cC 0I4 0m 4 .0 4 C.C 0C N No~ 04j * N u 'i 4N- J U)Uuo 0 40 - a% ch (00 0% ON 0 0 w4 0 44 (0a ) -c

c40 08 140 - 0 1@U40at-o 14 -c 6.04 2. 00 .> 0 0 'o4~ ( -z %. N%

o44- 4 c u " 0 n 46H 0 I 140 4004 .1 0 O 0 N.C0 1. 0 4u 6-444.4. w4 0

U) 0 r. lu u . U- - 41 . . m 3 0e - %W C VS44>a41 4- l0% 010 6 0 4 0 0M 01 c r- 0 --E1-.0 Z-. -

4 0 N 0 0 4p0 c .4 0 0 a :" .0 uN 040C 0 w4. M40 u I.- C2~~~~~ 0 w.40 a4 I4(0 -64 c0 41. ~0D 4 a .a uC in v. 206014 140 a C C 0-601 0 A 1-0 1--0).

k0 04.0c 44.0 14.0 Q4.40 WOIO OW~ 40 26O1- c u rIcM10*.04 4 4 a fu *4004 * 4 a *. 00 6 C1 0 L.

140Q C4104 14- 410 441~6- 4 O 104 10 100-6U)4ZR 0( Q ~ 0(0 00. 1 4 0 S.- U) u u.J aZ(n 0 D u Q4 U

0

0 00

00 0 .0 0j c0

. 0 0 ~ 0p 4 0 0 '0 0r0W60 a% 0 '1 C4 0 6, .0 g to4 v6400-

606 . 0 -. 040.- 1In V4)4W 04N14 44 N~'* 0w a M~" U A N c U 0'40 44 ( -~ 60 444

(0 C V 6 6 00I 140.0 C 4~ a 4 4 4. 10 4 A '=4

0 4 c 0 ~ 0 >1 0 ~ 0..f0 2 c4 "04lO N646 10 4 0.. 400 U 0len 2444 0000 s j uS . tL w J . c W4.00-6qv 4u 6 4 0 0440146 MO 03 0 0'D p14 2 lo140 044 00 6 4fl6"S we 0 00 c 4 O u w a. 44441 >0 2 4 0 1 00h x 0 1.-0 Mb) .~~~- I~g A0-4 .n0 12 .4 In 0 >. 6- 4 0 .

02 0(0 02 -. x4 0 AW0 0 00W .1" 0i00 6

V - minU MI I c w V. NJ . c 4

0 0 C.0 .

0 19- 0 W 4. C- 0t.o =04 n~ in O 4)-4v a

flC£04 I., k0 -4.4 int' C404-.1. ~**C~ 14..44 0 C4 O1 t02C U

-040 W04 4 00 Z 41400 0 a 64 L4.4- COV f) 4 4) N00 n. a. C .(NU 12 0, C .4 fn 0 AJ M 4 Q-N C041 r414 - in-~ c C' 0100''-.

in *C 0 40 m 40 (mCl 2A4CC- *4) l 00 In 4 >~.C Cl.' 04r42. C , w4£C U .S;V4 404 4.2- 00 C. w4j. U10.440 .04 .4112014#k 0 4 .IC CL.- *Z0 >*-EOIIO - .1 a.. 04 0

040(42 c 0142flI 0 wnf 0 c Uoo 'o ~ u. ~ ..

CN~ 0.C - 2 ,-- MA . 0 = - 4)104 C

ol 1% 0. 0) I 10 1mv0 e w 0 4

C , 44 01 >0. - 4 c0 %0ca414 a .0 i

014 . 40 as A to.1 .4 OL C ON U 0 t C M inC 4 . 4i

.A0 A- 0 U

.C4 NO 0 40.2 *.~f 4l 00- A .. ( 0 00o w1 C2.0 0 4400 ;I ten,4 %. 1*04idi 0 :1 04 1 1 0AUco 0 C4 >41 in u C

. . C 0 0 .. U w 4214 in-. (42. * 0 0~ N4. as. >~44 NU" " 4-4n 40,U w112 * 4 c2. -1 4. £ 3404 c 1" 11 0 C4 C .

0.-4.. A40 l 4 C, 4'k 0' ", w 4a 4 4. .2 W U w. 0 U4 C44.40 C .41l0W0W 041M 0~. 14u C ,S .3 4 444%.I NS W. 0%l-

6.1 4 0 02 C4C W 0 ~0 to.~C --. ~400 . C A 04k44 U a 0 A 4.0-1 0444 1 a 3 C WOalso 0 .0 -

4 04 --c U tj -144 in4 U > Cw* -t & j0 )%n44 U0c W 4. 4>. 0 Nlz ... I,42"I w4 1140v 4CC r, 4A

1.-1. 6 @2 t , .c 6 4

I* C

41441 0 0 j

04 OC 0 V1 a -- "C 'o0 140 -4 1 T 4 C 0 . - * 40 0..

. aU 0 P2 41M .4 4 40 in2 0v 4. D. 4 42

4.4 v4 wU N AjJP 9" N -4 :01 64 Nj P. 14 a wC s CN.4 U010 0 :0.0 4 a A C 4 P. 0-64C 1 2N 0 em.. c a.P 41 00 4 C I Aj w c4 r4 N £00 04"1 'a 'r 4.0. JU -4. .0 0

. In.4 4.0. 41 -.1w 1,0 00 ON 4 .4 00 044-IC '10 0 0 0 4 ~-'44 gp x. 0 7 al VIa. x- 0' 4 4 1 £AU* -4 04N 4U 3 0' V4 0 a w. 04. 1 0 c i UM 04 U A

hi 4.4P AM *.0 0 U0 00 o .cc 400'. &.0 U " - .*2 Ai0c 0 0 114 " A21 . 004 1442 W. 0. C04I4 0 -

2 j c 1 1 &W1 A0£%4 l4C 0 0.4 N1dc. OD. ij " *w > 1 1.441 4J0 I44 UI* COC 40 aM 4144 .4 P 0 C00$13* 4 v2£) 4 p 040 CI0 a4 C1 .0 C41.0 ol 'o - .. 41 W1 001 . i0 1£ U4 - 0 0 C

1..:;4 31 IO AJ >0411A 904 01- tr

a4 .. 4 2i C

C No .2.4 08"4 040. a -A P C w 2 0 22. S0 z a >.u

00

O l i -z ro - Z

f" P 0w .4 i. - P £4- .10 10 0 (N 0C0.4 . 4' 1 14U 0 2.

144 (ON CO 40U.0 40 P". Ott 4 4 U * 4 0. .40 1.14 O00M

4 0 a.001. 0 0 6s -02 4 04 C 41. Q1'A

C 414P - 4 a -1 C1 -44 w4 2 'm x4 -.4 2 4. 'm

0 i U r. 4 N4 4C I~ 41

.NC~~~~~~~~ 44-4 90j 4.0 4*M 4 .. 14.,',"* 6004 ty, E 1. *.(21.1 g .11

X 4 A j .Z 0 4 * 4 0 1 4 1 w C 4J02 4 ..

8 1 O* 41.O 1 1401 f .4 0. V :4 .a3.~ w

1 4C 6 -0

0 0 0 - 3i- -3 f-= WCX

04 U. ' -3ccc c x a I. P d

0 0 b

o 0i N 41. .Woo .1 Nr 41 .1 .1 c

4. 0 w 4 In 010 c .0 0.C

a 0 P, 1" 11 40 en4 a .14 1. .C 41 4 E0O 4 C.1 0 I 0 0I CM At.0 03 4

Ac .JO ag .-.4 04 w 4141 SM 0n N v N .1

E S4 -. 'a 0C. 0M .0 0 .1 NC 0 0C C -A 'A10 46 C1 . 0 4V - 4 0 a12 a10- W41N

SCWW V *- C -o jU :0W- a 04 1.4 00 (AE 0 h.C -41 CU ON 30Ai 0 N U'1 O -05. 5444 I, UW CU C -=cj . C- w OC ON -li 0 - -- 0w

>1 1. 0 0 41.4 4)04. 1 0 w O. 04C O 0 0 w

EU E C 0 OC A U (d C 0 a. 0 0 a C >. 0OSOs

u3 a3 Zo r- 0..V- 4Eh w'1 02 10w3 .30 044 ",1 -z

ImvC C U )>C = 0 0 Sg

c0 U39 I. 4w .0 0- i g So "Z .1 -l 41 3I. 54m 0 cc

410 a1 0* CI 0 .0 .10 0. 4 01 0 E1 Ce 14) 0

.1~O N .40 N4 c44 O1 A0 .4 1 .O 4 ...1 N 04 0 U Ij .4 0 . V111 -) .C50 0 0% A .0. a N N S C I 40 N wo us .. Ea C N In1 .14Wa4v-.4 60 0 C Ion- 0 Nn N

- C 0041 0.0 C>. a1 .4. N 41 w v1 .. 4 1 ...041"4 CC0 N Cp : C4 N c1 1. 0 404wI 4 a . N " N 00 0I., u . 3.0 21c 0 c Y n *0. 0c ES fnI1 41 ZM >1 ~ m 1 A 0

Ca 00 v U 0 04 A .. A.4 Im 104% a30 -CU V41 - u1.~., -C 40 CO c4. Om U oc501 -4 1. ac1q * C m 0

C01.1 u~ .4 o1N ..0 0 > Oc . u 0.. 4 .1.4 .44d A .1 .0 C I m04 a ; &J ; U a) 0 0.0.~ W'~ l .0 5 a 0 0 0

.10041.0 04 ". 01. 0111 N. 1.1 w 4 w 4Z.. tOO) .10 .4. 0. 1.0 (L.1.049U 000.0 04 1. A.4S 1.1.4 -- w. 35104 1.0 U ~N 00(0 m

40 0 C .-0 c 2 vCE 0 ao 6 41 w .1w0141 6. . 0 0 4L) 0.. u 4

*0 o 0- 0CW S CA AS H01. %-d *C -R A0 In * j 4) CQC W. 0-4

0 r40

InI

'A0

0g . N

C-> w o N V.0 N oD.4 4 0 c40 4

.4 . 1. 5 > 0 >N41 c M 41 6. 0 -C1 .1C0

4 N C0 C, 10 5 4400t o 0 43 gM .0 In 17 C1 .A 0 C Oc 0 X,~

C-~~~ in 41 0 C4 N 6.N0 g > -4 1 C 1 NSC 0 a1 0.4 4 1 A5 . S40 1. A, 00 aI

co -0 0 4 0 0414 0 -i S a-So1"4.1.I 0 6. w-i - 3!)IA1Ai 2.1 04 m 04 5 c.~ D. C *. 0, 44 hc>

CCEN0 b. lp C. S I 4C * M M. . CEC4.1 0 Uc Cid- No a 09. 155 1.

".CI" 0 .1c ..44 N 0 N 1 XS I . - - 4 c- .so1.

01 0 S W. 9 I4 0c 0inU n 02- W1. . 0 Qu 0 -e

.4.4 C.4N 0C C '~.. -0

0 04

0 41 "0 cm W044V 04 a 1 4.a cu 2 a O

41 A 0 4W 0 l 1A-0 1. C@ VON 00 CM 17 0Q 4 400 C N4 0 O

C M.0 4 41 . 3 > 9 0 IN 0- 41 D.-4 N 12 04 ix~ N a NU1 W0 N,. 40.1 4C- 0 0.4 4 us, 11 0 91 >m II!, 4 - 11 C0-0, a e .

I-S . 60.4 04 * 4u> 0 0.1. 0 I. 0% 5 .1 W1..4 0~. b1 0 41 0. dc 'n*-

064 0 v c 0 b.; C - 04 04.0 .

6. w41 n 4 U) A. 0 416 >.C>OC

.i 0. m1 0. V.ub 0 P W C U 1.00 - V C:M C C.41u001 c 6 4 c S. 0 CEl - C-1.4CO

- 641 C 1.0 V I 00 .I.4S 0.- I.. 1

. 10 - k14 I - CC ~ O CRa4 -W4 .0 C1 90C-'ou . 0 -a.)t.1 41 0.VOO o 4 0.4 MC *40. *O IO

1%

%4lz 04

I