Parsing in Parallel

(1)

P a r s i n g i n P a r a l l e l Xiuming Huang ~ Louise Guthrie

Computing Research Laboratory New Mexico State University Las Cruces, NMI USA 88003

A B S T R A C T

The paper is a description of a parallel model for n a t u r a l language parsing, and a design for its imple- m e n t a t i o n on the Hype,'eube multiproeessor. T h e pm'allel model is based on the Semantic Definite Clause G r a m m a r formalism and integrates s y n t a x and semantics t h r o u g h the commmfication of processes. The main proeessess, of which there are six, contain either purely s y n t a c t i c or purely semantic information, giving the a d v a n t a g e of simple, t r a n s p a r e n t algorithms dedicated to only one aspect of parsing. C o m m u n i c a - tion between processes is used to hnpose semantic eon- st,'aints on the syntactic processes.

1. I n t r o d u c t i o n

This paper describes parallel model for natural language parsing and gives a design for its implementa- tion. W i t h the a d v e n t of parallel machines, it may be possible to view the relation of syntax and semantics in n a t u r a l language parsing in a wholly new way. The approach is moving towards an application environ- m e a t which is similar to the heterarchical sytem proposed by W i n o g r a d [Winograd

721. Processes which

control the syntactic aspects can be s e p a r a t e d from those which control the semantic aspects in t h a t the syntactic processes contain no semantic information themselves, b u t receive it by c o m m u n i c a t i o n w i t h the semantic processes, a n d vice versa. The a d v a n t a g e of this a p p r o a c h is t h a t t r a n s p a r e n t algorithms can be w r i t t e n t h a t are dedicated to only one aspect of parsing, while the desirable effects of integrating s y n t a x and s e m a n t i c s can be achieved t h r o u g h the c o m m u n i c a t i o n of processes. In our model we use this c o m m u n i c a t i o n to enforce semantic c o n s t r a i n t s on the syntactic processors in order to avoid the combinatorial explosion of producing all legal s y n t a c t i c possibilities.

C o m m u n i c a t i o n between the two c o m p o n e n t s is t h e n our focus in designing a parallel parser. There seem to be three obvious levels at which communica- tion between s y n t a x and semantics can take place: the word level, tim phrase level, and the sentence level. We have chosen to consider c o m m u n i c a t i o n at the phrase level r a t h e r t h a n at either of the other two because it would be too early for the syntactic and semantic c o m p o n e n t s to c o m m u n i c a t e at the word level (too lit- tle information is available at this level to help the parsing), and too late for t h e m to c o m m u n i c a t e at the sentence level (too m a n y syntactic parses might have already been "produced). How the c o m m u n i c a t i o n between the syntactic and semantic c o m p o n e n t s takes place at the phrase level will be described in Section 3.

In Section 4, we design an h n p l e m e n t a t i o n of this pa,'allel model for a 6-Hypereube lintel 85] multlpro- cessing system, which we will have available shortly.

The 6-Hypercube has

sixty-four

identical processors

(Intel 80286's with 512K bytes of m e m o r y ) a n d no shared memory. A l t h o u g h each node in the Hypereube can eventually c o m m u n i c a t e with any other node, each processor can directly c o m m u n i c a t e w i t h only six immediately neighboring nodes. W e therefore i n t e n d to

limit our message passing among processors to immedi- ate neighbors w h e n e v e r possible.

Like the work of Eiselt [Eiselt 85] on parallel inference processing, we have a pers]~icuous assignment of n a t u r a l language processing moaules to processors in the machine, a l t h o u g h we are suggesting a parallel i m p l e m e n t a t i o n of a parser w i t h m u c h more parallelism and w i t h a clearer s e p a r a t i o n of s y n t a x and semantics. The work on "massively parallel parsing" by Waltz and Pollack [Waltz 85 models various c o m p o n e n t s of eompreheilsion by a~tivation and inhibition of nodes of a network. A practical applieation of t h e i r a p p r o a c h requires massively parallel processing, c u r r e n t l y beyond the s t a t e of l, he art in multiprocessing.

We base our parallel model on the Semantic Definite Clause G r a m m a r s (SDCG) formalism of Huang Illuang 8 5 . The SDCG evolve'd from the I)efinite Clause G r a , n m a r s of P e r e i r a [Pereira et al 80 a n d is described below.

2. S e m a n t i c D e f i n i t e C l a u s e G r a m m a r s

The SDCG is em'rently i m p l e m e n t e d on a single processor machine where it is the parser for the X T R A (English Chinese Sentence

Translator)

machine translation system [Huang 85]. The X T R A is a p r o t o t y p e system now r u n n i n g nnd&" a C-prolog i n t e r p r e t e r and fias a wide coverage of English p h e n o m e n a , even t h o u g h its v o c a b u l a r y is r a t h e r small (1000 entries). The SDCG uses the s e m a n t i c s of words and phrases to restrict the n u m b e r of syntactic pm~es of a sentence to those which arc semantically compatible.

A simplified vemion of the SDCG used in the X T R A system is as follows:

(1) sentence(s(Subj Np,

v p ( v ( V e r b sense),Obj Np)) -- >

noun_phrase(S ubj Np), is v e r b ( V ~ r b ) , -

subject verb m a t c h ( S u b j Np,Verb,

- VeTb sense), - n o u n _ p h r a s e ( O h ] Np~,

verb obj ect__matell(Verb sense,Obj_Np). The g r a m i n a r says t h a t an i n p u t string is a sentence w i t h the s t r u c t u r e s(Sub3: Np, vp-(vYVerb sense)

Obj Np)).

if it is composed of Su(~j NP which is-a noun{ phrase, followed by Verb (a ve,'b) whose one sense

Verb sense is semantically compat:ble w i t h Subj Np,

followed by Obj NP (a n o u n phrase) which is semantically c o m p a t i b l e - w i t h Verb_sense.

The s u b - g r a m m a r for pa,'sing a n o u n phrase is as follows:

(2) n o u n phrase[np(det(DeQ, adj(Adj sense),

n ( N o n n ~ e n s e ) ) ) - - > -

determiner(Det), adjective(Adjective), noun(Noun),

(2)

The last predicate ill the n o u n phl'~Lse sub- g r a m m a r , ' a d j _ j i o u n A I : a t e h ' , tries to m a t c h Adjective a n d N o u n to find a compatible pail" of senses for tile given A d j e c t i v e a n d Noun to be eombined. The predi- c a t e s ' s u b ] verb. m a t c h ' a n d 'verb object m a t e h ' i n i h e sentence g r a l n m a r accomplish shn-ilar ta~k. All those m a t c h e s are based on the system of seleetional restric- tions proposed by [Katz & F o d o r 63] and their eod- ings are o m i t t e d here to save spaee, l,ater we will see how t h e y function.

There is a syntactic lexicon ia the SDCG of the following form:

determiner(the).

noun(coach,[eoacht,eoaeh2]).

n o u n ( s t a r , [ s t a r l , s t a r 2 D.

adjective(tough,[tough1 ,tough2,toug h3,tough4]).

verb(marry,[mar,'yl

,inarry2]).

For instance, the syntactic entry for " c o a c h " is a noun h a v i n g two senses, labeled " e o a c b l " and " c o a c h 2 " .

For each word sense, a semantie i n t e r p r e t a t i o n is given in the semantic dictionary:

sem(coaehl,[head(thing)])*.

(eg.

'a

passenger

coaeh')

sem(coach2,[hoad(man)]).

(a

~rah,er)

sem(starl,[head(thing)]). (a celestial object)

Seyl/(star2,[hcad(l-/lan)D, (¢'a sil]g]llg S~al"', etc)

sem(toughl [poss(tlfing)]). (,nodilles

'thing', as h,

% tough m a t e r a : l " )

sem(tough2,[poss0;:,an ) ). (modines ' m a n ' , as i,, " a t o u g h m o u n t a i n e e r )

,

seln(ll:_{-irry

]

,lsubj{man),o ,j(m an),head(do)]).

( J o h n m re'led Mary. )

senl(marry2, snbj (,~1 an),obj(t hing),head(do)l), tog. in "IIe mm'ried money. )

For example, "coach1 labels the sense of " c o a c h " whereby it refms to a "tlfing . in pa~sing (3),

(3) The I, ough coach married a star*.

according to the g r a m m a r in (1) the system s t a r t s w i t h the predicate 'nonnA)lu'~me', which is presented in (2). After it i n s t a n t i a t e s tile variables

Det, Adjective

and

Noun

i n s t a n t i a t e d to t h e ' t o u g h " and coach" it a t t e m p t s to apply the predleate 'adj noun m a t c h ' , whose task it is to find tile first pair of senses for the words " t o u g h " and " e o a c h " , respectively, which are compatible w i t h each o t h e r aeeording to our seleetional restrietions. Here I, he first pair found would be ' l o a t h 1 + coach1', beeause the s e m a n t i c category of " c o a c h 1 " ( ' t h i n g ' ) fibs into bile 'poss(thing)' slot of the word sense % o u g h l " (meaning t h a t his adjectival sense is for

modifying sometlfing whose semant, ic category is

'thing').

Now tile parser is at tile predicate 'is verb' where it finds the verb " m a r r y " . It, t, hen tr:es to m a t e h

Subj ,Np

(%otigh] -I- c o a c h l ' ) witt: a some sense of the " m a m ' y " b u t fails because b o t h " m a r r y l " and " m a r r y 2 " prefer the subject to be of the semantic category ' m a n ' , which " c o a c h l " c a n n o t satisfy. The system b~ektracks, trying 'adj noun m a t c h ' again and producing the next matclfing-pah" of senses for " t h e

t o u g h coach" ('tough2 -I- coach2'). YVhei}

'subj v e r b m a t e h is tried again and it selects 'marry1 as tI:~ appTopriate verb sense. The parser proceeds to

analyse the rest of tim sentence, employing

" n o u n _ 0 h r a s e " to find the object noun p h r ~ e sense

* The s e m a n t i c primitives s n e h a.s ' t h i n g ' , ' m a n ' , ere, are based ell the primitive set suggested in [Wilks 75].

* Modified version of the " s e m a n t i e garden p a t h " sentence by [Chm'niak 83] ( " T h e a s t r o n o m e r married the

s t a r , " )

and " v e r b _ o b j _ m a t c h " to see w h e t h e r this noun phrase sense lits the partieular verb sense. 'S~arl' (a eelest:al object) is t h u s tried and rejected, and 'Stal'2' (a cele- brity) is accepted ( ' m a r r y 1 ' requires the object t'o be of the s e m a n t i c category m a n ). A plausible zeadlng of the sentence is t h u s gained (' "File strict ~ralner married a celebril, y.")

It is clear from t, he above description t h a t in the SI)CG s y n t a x and semantics closely interact: s y n t a x - s e m a n t i c s -. syntax, ere. One class of predicate waits for the o t h e r to make a decision, t h e n makes its own decision. ]low much baektraeking must be done is unpredictable; the pm'se m i g h t only be completed after several routes have been tried and rejeel, ed.

3 . P a r a l l e l P a r s i n g

The model consists of six processes which con> m u n l c a t e to produce all the semantically compatible parses of a given sentence. Each process will be hnple- m e n t e d as a tree of processors. The root node of t, he tree eonl~ains a queue of requests and allocates processors to the elements of l, he qneue as they become awdl- able. For the pnrpose of this model it is sulIicent to note f, hat each process itself has {,he capability of processing several requests in parallel. We identify below each of the processes and describe the c o m m u n i c a t i o n between t h e m .

1) Sentenee maste," - Controlling process which operates as a modified top down syntactic processor (modified in the sense thaC infer- [nation fi'om el, her processes influences its decisions).

2) Noun-phrase m~ster ( N P - m a s t e r ) - Given an a r b i t r a r y string, it identifies syhtactieally all possible initial noun phr~mes in the

string. Ttn'ough eommuilication with the

A N - m a s t e r , it del, ernfines which of I, hese are senmntieally acceptable.

3) Semantic dictionary lna.ster - C o n t a i n s the semantic dictionary and provides appropriate entries for the current i n p u t sentence to the o t h e r semantic processes.

4) A d j e c t i v e - N o u n

master

(AN-master) - Given an adjective and a noun, Iiads all possible pairs (adjective word sense, neat] we "d sense) t h a i are compatible.

5) Subject-Verb m a s t e r ( S V - m a s t e r ) - Given a word sense for a

nou6

and a verb, finds all possible word senses for the verb t h a t are compatible.

6) Verb-Objeel, m a s t e r (VO-lm~ster) - Given a word sense of a verb and a word sense of a noun, determines w h e t h e r or not t h a t verb sense-object noun sense pair ]s compatible. The following diagram illustrates tile processes and the e o m m n n i e a t i o n between them.

r . . . input . . . :

I AN-lnaSLel'-*~ -- ~ N P - m a B L c r I

Ii // "\ , I

S e m a n t i c S V - m a s t e r

] ) i c t i o n a r y " m j - ~ ~ - - - _ S e n t e n c e lllas[,el" - - ~ l]'l t~.q ~ c 1-

(3)

I n p u t is read simultaneously by the semantic dic.~ tionary, and the sentence master. The sentence m a s t e r contains the s,vutaetic dictionary and tlegins a top- down parse o f tim sentence guided by the definite clause g r a m m a r . Wimnever a noun piu'ase is searched for, the noun phrase m a s t e r is invoked to produce all possible initial noun phrases in the r e m a i n d e r (tllc u n p a r s e d portion) of the i n p u t string. After the main verb of any clause imps been identified by tim sentence master, the SV-master is invoked to produce all possible verb senses which are m e a n i n g f u l a t this point in the parse. Ill tim case t h a t a t r a n s i t i v e verb is found and a possible word sense fox" the object noun is deter- mined, the V O - m a s t e r is consulted as to wheti~er or not the given verb word sense and object noun word sense are acceptable as a verb-object pail'.

In c o m m u n i c a t i n g witl~ the N P - m a s t e r or SV- master, several possibilities m a y be r e t u r n e d to the sentence master, and the parse is continued fox" each of these possibilities in parallel.

Tile N P - m a s t e r , which is also a syntactic process, finds all possible initial n o u n phrases which are meaningful by using its own syntactic information (in a top down m a n n e r ) a n d by c o m m u n i c a t i n g with tile AN- m a s t e r for semantic information. This c o m m u n i c a t i o n is similar to t h a t of the sentence m a s t e r witll tile SV- master. After d e t e r m i n i n g an adjective which is followed by a noun, ti~e N P - m a s t e r invokes the AN- m a s t e r to tind all meaningful adjective-noun word sense pairs. Multiple adjectives which modify a noun are considered in parallel by the AN-master, which in ti~is ease, r e t u r n s pairs which consist of a list of adjective word senses and a noun word sense. W h e n e v e r the N P - m a s t e r reeives a pair from the A N - m a s t e r , it continues any work t h a t it m i g h t lmve (such as finding prepositional pin'ascs which modify tile noun, e.g. 'the big boy in the park'). If several pairs are returned by the A N - m a s t e r , the r e m a i n d e r of the parse is h a n d l e d by the N P - m a s t e r and is done in parallel when possible. The sentence m a s t e r produces all the parses of the sentence t h a t have not been blocked. A parse may be blocked rot any one of the following three reasons:

1) The syntactic category needed by tim sentence m a s t e r is not satisfied by any initial segment of the unparsed portion of the input.

2) The S V - m a s t e r r e t u r n s a negative response. 3) The V O - m a s t c r returns a negative response. We use the example in Section 2 ( " T h e t o u g h coach married a s t a r . " ) to illustrate the above commun- ication of processes and to exhibit a p a t h whiei~ is blocked.

F o r shnplieity, we write the SDCG used previ- ously, w i t h o u t the a r g u m e n t s for the predicates

involved. We also add an additional rule for

n o u n p h r a s e and a n o t h e r e n t r y in the semantic dictionary for the noun sense of 'tough',

tougi13 (as

in " t h e t o u g h never suffer"), to m a k e the example interesting.

sentence --:> n o u n p h r a s e , verb, subject verb m a t c h ,

noun p h r a s e , v e r b _ o b j e e t _ m a t oh.

n o u n A ) l n ' a s e - - : > determiner, adjective, noun, adj n o u n match.

n o u n _ p h r a s e -- > determiner, noun.

d e t e r m i n e r - - > [the].

d e t e r m i n e r -- > [].

The sentence m a s t e r receives the i n p u t and in this ease, immediately passes it to the N P - m a s t e r and waits. The N P - m a s t e r finds " T h e t o u g h " and " T h e t o u g h coach" as possible initial n o u n phrases in the string it was given. " T h e t o u g h " (tough_3) is r e t u r n e d immediately to the sentence m a s t e r who begins searching fox" u

verb. Sinlultaneously, ti~e N P m a s t e r sends the adjective noun pair, (tough, coach) to the A N - m a s t e r . The A N - m a s t e r r e t u r n s ( [ o u g h l , coachl) ("rugged vehicle") and (tough2, coacil2) ( " s t r i c t t r a i n e r " ) . Note t h a t these are the same possibilities considered by back- tracking in the example in Section 2. The N P - m a s t e r r e t u r n s these to the sentence m~ster, who initiates the c o n t i n u a t i o n of the parse fox" each of timse possibilities. The sentence master, in tim interim, found a verb (coach) for its f r s t n o u n - p h r a s e (the tou{~!13) and request a s u b j e c t - v e r b m a t c h from the S v - m a s t e r . The S V - m a s t e r r e t u r n s coacil3 (the verb sense of coach) and the sentence m a s t e r continues with the r e m a i n d e r of the i n p u t string " m a r r i e d a s t a r " . Here, a noun A0hrase is needed, and so once again the NP- m'mter is invoked, and asked to find an initial n o u n phrase in the string. Since no noun phrase is

found,

this p a t h is blocked. The p a t h containing

( t o u g M , c o a c h l ) will be blocked exactly as the description in Section 2. The p a t h containing (tough2,eoaeb2) will succeed and produce the correct parse fox" the sentence.

We now consider the function of the Semantic Dic- tionary master. While the sentence m a s t e r is receiving its .input a n d begins the processing described above, the semantic dictionary m a s t e r simultaneously finds all possible word senses for cacti i n p u t word. The semantic dictionary contains an e n t r y m r each sense of a word. The s t r u c t u r e of each e n t r y reveals its syntactic category. W o r d senses corresponding to nouns contain only the s e m a n t i c class to whici~ the word sense belongs. For example, the semantic dictionary e n t r y fox" tile n o u n " n a m e '~ (as in the gh'l's name) is given by:

sere(name1, [>ad(sign)l ).

Adjective word senses contain the semantic class of the n o u n t h a t it prefers to modify. The adjective "specific" has the following entry:

sere(specific1, [poss(sign)]).

W o r d senses corresponding to verbs m'e described witi~ a structm'e which contains the class of the subject t h a t is prefered by this verb, the class of' the object prefered, and the semantic class of tim verb itself. The verb " n a m e " ( " t o n a m e a dog") is represented as:

sere(name2, [subj(man), obj(man), head(make)l ). After finding all possible word senses for words in the i n p u t sentence, the s e m a n t i c dictionary m a s t e r sends these dictionary entries to the appropriate semantic processes. Verb entries are sent to the SV- and VO-masters, adjectives are sent to the A N - m a s t e r , and nouns arc sent to all three. These three process masters t h e n contain a " c a c h e " of the semantic dictionary entries relevant to the parsing of the present i n p u t sentence. T h e purpose of the " c a c h e " is so t h a t the s e m a n t i c dictionary e n t r y fox" any input word can be quickly found by the processes which use these entries.

4. T h e D e s i g n o f t h e s y s t e m

>k

We describe the design of the i m p l e m e n t a t i o n of the parallel parsing model. E a c h of the six processes consists of a tree of processors. We label the root of each process tree w i t h the n a m e of the process t h a t it represents. The design of the semantic processors a n d the n o u n - p h r a s e m a s t e r is i n d e p e n d a n t of tim imple- m e n t a t i o n of the SDCG which is used. The design of the sentence master, however, is heavily d e p e n d a n t on the formal g r a m m a r used for the SDCG i m p l e m e n t a - tion as the parser fox' X T R A . The two syntactic processes above, the N P - m a s t e r and the sentence master, have a significantly more complex design t h a n those of the semantic processes so t h a t different possible s y n t a c t i c alternatives m a y be considered in parallel.

(4)

4 . 1 . T h e s e n t e n c e m a s t e r

T h e desi~;n of tile s e n t e n c e m a s t e r is b a s e d o n t h e following p r o d u c t i o n rules o f t h e S D C G :

s e n t e n c e - - > s e n t e n c e 1)ody.

s e n t e n c e - - > s e n t e n c e h e a d , s c n t e n c e _ f l o d y .

I n t u i t i v e l y , we (:an c o n s i d e r t h e s e n t e n c e h e a d to be w b a t e v e , ' a p p e a r s before the s c n t e n e e s n b j e ~

0t

c a n be an e m p t y s t r i n g ) , a n d t h e s e n t e n c e b o d y Lo be t h e r e m a i n d e r of t h e s e n t e n c e .

T h e s e n t e n c e m a s t e r , as i l l u s t r a t e d below, c a n be t h o u g h t of as t h e root of a t r e e w h i c h h~s t w o c h i l d r e n w h i e h we will refer to as t h e s e n t e n c e m o n i t o r s : t h e s e n t e n c e h e a d m o n i t o r ( S H - m o n i t o r ) a n d t h e s e n t e n c e b o d y m o n i t o r (SB.-monitor). E a c h s e n t e n c e n l o n i t o r is tile root of a s u b ] r e d of ehild p r o c e s s o r s ( S i M l a n d l e r s a n d S B - h a a d l e r s ) a n d a c t s as ,'t m o n i t o r foi' t h e s e child p,'ocessors. W e l a t e r describe t h e s e n t e n e c h a n d l e r s in m o r e detail.

S e n t e n c e m a s t e r

SH-molntor SB-monitor

' - , • - ~ , ' SBhandleil SI h m d l e l 2 S l / h m d h l / SII handlml SII hmdlct2 ~ ' - , ' .." -. a '" ' - ; ".'1

T h e s e n t e n c e minster is t h e p r o c e s s w h i c h deter- m i n e s w h e t h e r or ]lot a st,'ing is a s e n t e n c e . A n y i n p u t to t h e s e n t e n c e m a s t e r is i m m e d i a t e l y g i v e n I,o b o t h t h e SliI-monitor a n d t h e SB-nlonlto," to e x a m i n e in parMlel t h e possibilities t h a t t h e s e n t e n c e does trod does n o t h a v e a s e n t e n c e h e a d . T h e S H - m o n i t o r s a n d t h e SB- ] n o n i t o r s e a c h p u t i n c o m i n g r e q u e s t s f r o m t h e s e n t e n e c I n ~ t e r in a q u e u e a n d allocate t h e first a v a i l a b l e child p r o c e s s o r t,) b e g i n its work. In t h e ease of a SH- h a n d l e r , t h i s w o r k is to i d e n t i f y a possible s e n t e n c e b e a d , a n d in t h e case of a n S B - h a n d l e r , it is to see if t h e i n p u t s t r i n g is a s e n t e n c e b o d y . T h e S I I - h a n d l e r s a n d S B - h a n d l e r s m o n i t o r child p r o c e s s e s w h i c h o p e r a t e in parMlel.

in t h e egse t h a t a s e n t e n c e h e a d is f o u n d b y one of t h e S I I - h a n d l e r s , the. r e s u l t is r e t n r n e d to tile s e n t e n c e m a s t e r v i a t h e S H - m o n i t o r . T h e r e m a i n d e r of t h e i n p u t is t h e n g i v e n to t h e S B - m o n i t o r w h i c h a l l o c a t e s a fl'ee SB-handle," to c o n t i n u e t h e p a r s e of t h e r e m a i n d e r of t h e s e n t e n c e . F o r e x a m p l e , c o n s i d e r t h e s e n t e n c e :

(4) W r i t i n g to ,Iohn w a s dill]cult.

T h e s e n t e n c e m a s t e r g i v e s t h e s e n t e n c e t o b o t h t h e S I I - h a n d l e r a n d tile S B - m o n i t o r w h i c h in t u r n give it to one o f t h e i r c h i l d r e n , s a y S I I - h a n d l e r l a n d S B - h a n d l e r l . Since t h e g r a m m a r for tile S D C G i n d i c a t e s t h a t a n lag--clause is a possible s e n t e n c e h e a d , S H - h a n d l e r l will i d e n t i f y " w r i t i n g to . l o h n " as a c a n d i d a t e s e n t e n c e h e a d . [ h e , e m a m d e r o f t h e s e n t e n c e " w a s difficult" is g i v e n to a n e w S B - h a n d l e r , s a y S B - h a n d l e r 2 v i a t h e S H - m o n i t o r a n d t h e S B q n o n i t o r , to see if t h i s is a possible s e n t e n c e b o d y . S B ° h a n d l e r 2 fails a n d notifies SH- h a n d l e r 1 (via tile SIt- a n d S B - m o n i t o r s ) . S i t - h a n d l e , ' 1 a n d S B - h a n d l e l ' 2 b e c o m e a v a i l a b l e for' o t h e r p r o c e s s i n g a n d S B - h a n d l e r l s u c c e e d s in s h o w i n g t h a t "WrilAng to ,John w a s dilrieult" is a legal s e n t e n c e b o d y .

T h e S]-l-handlers a n d tile S B - h a n d l e r s are a r r a y s of

~

r o e e s s o r s w h i c h i m p l e m e n t t h e o r - p a r a l l e l i s m of P r o l o g )r t h e p r e d i c a t e s s e n t e n c e _ h e a d a n d s e n t e n c e b o d y

r e s p e c t i v e l y . Below is a s i m p l i f i e d v e r s i o n of t h e g r a m - m a r r u l e s u s e d in t h e S D C G for s e n t e n c e h e a d .

s e n t e n c e h e a d - - > i n g - e l a u s e .

s e n t e n c e h e a d - - > p r e p o s i t i o n a l p h r a s e . s e n t e n c e - h e a d - - > a d v e r b i a l p h r a s e .

B a s e d o n t h e s e rules, e a c h Sil-handle," m o n i t o r s t h r e e child p r o c e s s o r s :

S H - h a n d l e r

pl e p o m t r o n a l plu ase m~ clause par e n t h e t i d a l _ p l u ase

T i m S B - h a n d l e r s m o n i t o r five p,'ocessors w h i c h are a g a i n b a s e d o n t h e S D C G . T h e f u n c t i o n of t h e s e five child p r o c e s s e s will v a r y d e p e n d i n g o n t h e t y p e of t h e i n p u t s e n t e n c e ( d e c l a r a t i v e , i n t e r r o g a t i v e or i m p e r a - tive). W e g i v e he,'e a s i m p l i f i e d v e r s i o n of t h e s e n l ; e n e e _ b o d y p r o d u c t i o n s in t h e S D C G for a deelara.- t i r e s e n t e n c e .

s e n t e n c e _ b o d y --:> s u b j e c t n p , v p l . s e n t e n c e b o d y - - > s u b j e c t np, vp2. s e n t e n c e b o d y -- > i n v e r t e d _ s e n t e n c e .

s u b j e c t n p - - > n o u n I)hrase. s u b j e c t _ n p -.- > ing_.elause.

Here v p l r e p r e s e n t s a c o m p l e t e v e r b p h r a s e , like t h a t in t h e s e n t e n c e

(5) J o h n d i d n ' t go to t h e p a r k y e s t e r d a y .

A n d v p 2 , ' e p r e s e n t s a n elliptical v e r b phra.se, like " d i d n ' t " in

(6) No, J o h n d i d n ' t .

A n i l l u s t , ' a t i o n of t h e S B - h a n d l e r s in t h i s case is g i v e n below.

Sl:l-handler

noun phrase mg clause nounA)hrase lag_clause inverted~sentenee

vpl vpl vp2 vp2

i n S e c t i o n 3 we i n d i c a t e d t h a t t h e s e n t e n c e m a s t e r c o m m u n i e a t e r s w i t h t h e N P - m a s t e r . A c t u a l l y , e a c h of t h e child p r o c e s s o r s of t h e s e n t e n c e h a n d l e r s s e n d s a m e s s a g e to t h e N P - m a s t e r , v i a t h e s e n t e n c e m a s t e r , w h e n e v e r tile D C G d i c t a t e s t h a t a n o u n p h r a s e s h o u l d be f o u n d n e x t in tile i n p u t s t r i n g . T h e N P - m a s t e r r e t u r n s all s e m a n t i c a l l y c o m p a t i b l e n o u n p h r a s e s . W h e r e t h e r e is m o r e t h a n o n e a c c e p t a b l e n o u n p h r a s e , a mess.age m s e n t to tile r e q u e s t i n g s e n t e n c e h a n d l e r w h o a l l o c a t e s one possible n o u n p h r a s e to t h e w a i t i n g child p r o c e s s o r a n d d i s t r i b u t e s t h e o t h e r s to a v a i l a b l e child p r o c e s s o r s . E a c h child p r o c e s s of tile s e n t e n c e h a n d l e r s c o m m u n i c a t e s w i t h t h e N P - , SV-, a n d VO- ma:sters v i a t h e s e n t e n c e m a s t e r .

(5)

Should each of the sentence handlers have a waiting child processor and the sentence master a request, we invoke a special processor, called t h e black-sheep processor, to g r a n t the request, so t h a t the requesting processes m a y continue. The black-sheep processor, functions precisely as the current single processor i m p l e m e n t a t i o n of the SDCG and will only be used to avoid deadlock*.

4.2. The Noun-phrase master

Since noun phrases are the m a j o r building block of m a n y s u b s t r u c t u r e s of a sentence, and since a m b i g u i t y often arises t h r o u g h d e t e r m i n a t i o n of different n o u n phrases (eg. " T h e tough coach the y o u n g " and " T h e prime n u m b e r consecutively"), the identification of noun phrases is an i m p o r t a n t place tbr parallelism in the parser. The N P - m a s t e r can be t h o u g h t of as the root of a tree of processors. It functions similarly to the sentence master. The noun-phrase m a s t e r contains a queue of noun-phrase requests and allocates t h e m to available n o u n - p h r a s e handlers.

N

o

u

n

-

~

Each n o u n - p h r a s e handler monitors three child processors. The child processors try to parse the next i n p u t phrase as a noun phrase w i t h no adjectives, one adjective, and two or more adjectives respectively. For example, in parsing the phrase " t h e t o u g h coach," two of the child processors would succeed (no adjectives and one adjective), these results are reported to the p a r e n t noun-~phrase handler, and t h e n sent to the sentenee m a s t e r via the N P - m a s t e r . A t this point, the waiting sentence processor (child of either one of the SB-handlers or one of the SH-hundlers) continues w i t h one of the possibilities a n d an available sibling processor is allocated b y the sentence handler to continue the parse of the sentence using the other possible n o u n phrase.

In the ease of a t r u l y ambiguous sentence, all legal parses are eventually produced. The above example would produce two parses in the case of " T h e t o u g h coach married people"?, b u t not in the case of " T h e tough coach the young.

Each of the child processes of the noun-phrase handler c o m m u n i c a t e s with the A N - m a s t e r via the noun-phrase handler.

4.3. The Semantic Processors

The semantic dictionary m a s t e r and the AN-, SV- and V O - m a s t e r processor trees have a m u c h simpler s t r u c t u r e in t h a t t h e y have only two levels. The root node is the master; children of the root are handlcm.

Semantic A N - m a s t e r SV-master V O - m a s t e r Dictionary

m a s t e r

The Semantic dictionary entries are divided

among the semantic dictionary handlers. The Seman-

tic dictionary m a s t e r reads the input and passes the relevant semantic entries, which it o b t a i n s from its child processors, to the AN-, SV-, and V O - m a s t e r s as described in Section 3.

T h e A N - m a s t e r receives i n p u t which is in general a list of adjectives and u noun, from the noun-phrase handlers. It forms all possible pairs (adjective word sense, n o u n word sense) and allocates child processors to d e t e r m i n e w h e t h e r or not there is a semantic m a t c h . The pail'S consisting of a list of adjective word senses, and a n o u n word sense which matches each of the adjective word senses in the list, are r e t u r n e d to the N P - m a s t e r .

The S V - m a s t e r a n d the V O - m a s t e r receive i n p u t directly from the sentence processors. The i n p u t and o u t p u t of these processes is exactly as d e s e r i b e d i n Sec- tion 3. In b o t h cases, the semantically compatible word sense pairs are d e t e r m i n e d in parallel.

5. Future work

The C o m p u t i n g Research L a b o r a t o r y (CRL) has the use of L o n g m a n ' s L D O C E English dictionary, which is realistic in size, prov]des comprehensive syntactic i n f o r m a t i o n and also has its semantic entries b o t h syntactically and semantically restricted, and lim- ited to a 2000 word vocabulary. We plan to i m p l e m e n t the Semantic Dictionary m a s t e r by providing each of the s e m a n t i c dictionary handlers with a portion of LDOCE.

A f t e r the initial i m p l e m e n t a t i o n of the designed parallel parser, we would like to see how W]lksian Preference Semantics [Wilks 75, Wilks et al 85] can be realized in our parser in the sense t h a t one or more readings (in the case of genuine ambiguity) can be selected by weighting the competing interpretations. We are also investigating a parallel parsing model which is driven b y semantics, r a t h e r t h a n syntax. We have in m i n d t h a t the role of the sentence m a s t e r in this case is purely semantic and t h a t s y n t a x is used only to help the s e g m e n t a t i o n of the i n p u t string. C o m p a r i s o n of the two systems would be of great interest to us. Eventually, we also w a n t to consider the incorporation of p r a g m a t l c s into the system.

6. Acknowledgements

We would like to t h a n k the N a t u r a l Language group at the CRL, namely Yorick Wilks, Jerry Ball, S h a r o n Dorfman, David Furwell, Dan Fuss, C h e n g m i n g Guo, and Sylvia Candeluria de Rain, for their com- m e n t s and suggestions. We also t h a n k Ted D u n n i n g for his m a n y helpful discussions.

7.

References

Charniak,

E . (1983) " P a s s i n g markers: a theory of

contextual influence in language comprehension," Cog- nition Science, 1983, 7, 171-190.

E i s e l t , K . P . (1985) " A parallel-process model of on- line inference processing," Proceedings of International Joint Conference on Artificial Intelligence, Los Angles, CA.

Huang,

X - M . (1985) " M a c h i n e t r a n s l a t i o n in the

SDCG formalism," Proceedings of the Conf. on

Theoretical and Methodological Issues in Machine

Translation of Natural Languages, Colgate University,

New York.

*The system will never invoke the black-sheep proces- SOl" unless the sentence contains four or more conjoined sentences or a p a r e n t h e t i c a l sentence with three or more eonjunets m it. In these eases the black-sheep

processor

is used only for the recursive calls which can-

(6)

Intel, C o r p . (1985)

iPSC Data Sheet,

Intel Scientific Coinputers~ Oregon.

K a t z , J. & Fodor~ J. (1.963) "The structure or a semantic bheolT,"

Language

39, pp.170-210.

P e r e i r a , F. & W a r r e n , D. (1980) "Definite elause grammars for language analysis - a survey of the for-

malism and a comparison with augmented transition

networlcs,"

ArtificialIntelligence,

13:231-278.

W a l t z , D a v i d L. Sz Pollack, J o r d a n B. (1985) M,'~ssively parallel parsing: a sl, rongly interactive

model of natural language inberpretaiton," mimeo. Wilks, Y.A. (1975) "Preference semantics," Keenan

(ed),

Formal Semantics o] Natural Language,

Cam-

bridge University Press, London. ~. ,,

Wilks, Y., Huang, X-M. and Fass, D. (198o) Syn-

tax,

preference and right attachment,'

Proceedings o]

IJCAI85,

UCLA, Los Angeles.

Winograd~ T. (1972)

Understanding

Natural

Language,

Academic Press, New York.