L H I P : E x t e n d e d
D C G s
for Confignrable R o b u s t Parsing*
A f z a l B a ] l i m G r a h a m R u s s e l l
ISSCO, University of Geneva, 54 R o u t e des Acacias, Geneva, CH-1227 Switzerland e m a i h [email protected], [email protected]
A b s t r a c t
We present LHIP, a s y s t e m for i n c r e m e n t a l gram- m a r d e v e l o p m e n t using an e x t e n d e d D C G for- m a l i s m . 'rite s y s t e m uses a r o b u s t i s l a n d - b a s e d p a r s i n g m e t h o d controlled by user-defined perfor- mance thresholds.
Keywords: I ) C G , head, island parsing, robust parsing, P r o l o g
1 L H I P O v e r v i e w
This p a p e r describes L I I I I ' ( L e f t - I l e a d corner I s l a n d Parser), a p a r s e r designed for b r o a d - coverage handling of l m r e s t r i c t e d text. The sys- t e m i n t e r p r e t s an e x t e n d e d D C G f o r m a l i s m to p r o d u c e a r o b u s t a n a l y s e r t h a t finds parses of the i n p u t m a d e from ' i s l a n d s ' of t e r m i n a l s (cor- r e s p o n d i n g to t e r m i n a l s consumed by success- ful g r a m m a r rules). It is currently in use for processing dialogue tr,'mscripts from the t I C R C M a p Task Corpus ( A n d e r s o n et al., 1991), al- t h o u g h we expect its eventual a p p l i c a t i o n s to he much wider. 1 T r a n s c r i b e d n a t u r a l speech con- tains a n u m b e r of frequent c h a r a c t e r i s t i c 'un- g r m n m a t i c a l ' phenomena: filled pauses, repeti- tions, r e s t a r t s , etc. (as in e.g. Right I'll have . . . y o u know, like I'll have to . . . s o I ' m going between the picket fence and the mill, right.). ~ W h i l e a full analysis of a conversation might well take these into account, for m a n y purposes they represent a significmlt obstacle to analysis. L I I I P provides a processing m e t h o d wlfich allows se- lected p o r t i o n s of the i n p u t to be ignored or han- dled differently.
The chief modifications to the s t a n d a r d Prolog *grammar r u l e ' f o r m a t are of two types: one or m o r e r l g h t - h a n d side ( R t I S ) items may be m a r k e d
*This work was carried out under grants nos. 20- 33903.92 and 12-36505.92 from the Swiss National Fund.
tNote that the input consists of wr/tten texts within the Map Task Corpus; LtIIP is not intended for use in speech processing.
2This example is taken fronl the Map Task Corpus.
as ' h e a d s ' , and one or more RHS items m a y be m a r k e d as ' i g n o r a b l e ' . We e x p a n d on these points a n d i n t r o d u c e other differences below.
The behaviorlr of LHIP can best he u n d e r s t o o d in terms of the notions of i s l a n d , s p a n , c o v e r and t h r e s h o l d :
I s l a n d : W i t h i n an i n p u t string consisting of the t e r m i n a l s ( t l , t 2 , . . . t n ) , ~ island is a sub- sequence (ti, t i + l , . . , ti+,,), whose l e n g t h is m + 1.
S p a n : The span of a g r a m m a r rule R is the length of the longest island ( t l , . . . t j ) such t h a t ternfinals tl and t i are b o t h consumed (directly or indirectly) by R.
C o w . ' r : A rule R is said to cover m items if rn terminals are consumed within the island de- scribed by R. The c o v e r a g e of R is then rn. T h r e s h o l d : The threshold of a rule is the mini- m u m ~;alue for the r a t i o of its coverage c to its span s which must hold in order for the rule to succeed. Note t h a t c <_ s, aud t h a t if c -- s the rule has completely covered the span, consuming all terminals.
As implied here, rules need not cover all of the input in order to succeed. More specifically, the constraints applied in creating islands are such t h a t ishmds do not have to be adjacent, but may be s e p a r a t e d by non-covered input. Moreover, an island m a y itself contain input which is unac- counted for by the g r a m m a r . Islands do not over- lap, a l t h o u g h when multiple anMyses exist they will in general involve different segmentations of the i n p u t into islands.
consmned. These ignore rules can be invoked ino d i v i d u a l l y or as a class. It is this l a t t e r capabil- i t y which distinguishes ignore rules from r e g u l a r rules, as t h e y are f u n c t i o n a l l y equivalent other- wise, m a i n l y serving as a n o t a t i o n a l a i d for the g r a m m a r writer.
Strict a d j a c e n c y b e t w e e n RHS clauses can be specified in the g r a m m a r . It is possible to define g l o b a l a n d local t h r e s h o l d s for the p r o p o r t i o n of the s p a n n e d i n p u t t h a t m u s t be covered by rules; in this way, the user of an L H I P g r a m m a r can exercise quite fine control over the required accu- racy a n d completeness of the analysis.
A chart is k e p t of successes a n d failures of rules, b o t h to i m p r o v e efficiency a n d to p r o v i d e a means of identifying u n a t t a c h e d constituents. In addi- tion, feedback is given to the g r a m m a r w r i t e r on the degree to which the g r a m m a r is able to cope w i t h the given i n p u t ; in a context of g r a m m a r de- velopment, this m a y serve as notification of areas to which the coverage of the g r a m m a r might next be extended.
T h e n o t i o n of ~head' employed here is con- nected m o r e closely w i t h processing control t h a n linguistics. I n p a r t i c u l a r , n o t h i n g requires t h a t a head of a rule should share any i n f o r m a t i o n w i t h the LItS i t e m , a l t h o u g h in p r a c t i c e it often will. Heads serve as a n c h o r - p o i n t s in t h e i n p u t string a r o u n d which islands m a y be formed, and are accordingly t r e a t e d before n o n - h e a d items (RHS items are r e - o r d e r e d d u r i n g c o m p i l a t i o n - s e e be- low). I n the c e n t r a l role of heads, L t I I P resem- bles p a r s e r s devised b y K a y (1989) and van N o o r d (1991); in o t h e r respects, including the use which is m a d e of heads, the a p p r o a c h e s are r a t h e r dif- ferent, however.
2 T h e L H I P S y s t e m
In this section we describe the LHIP system. F i r s t , we define w h a t c o n s t i t u t e s an a c c e p t a b l e L H I P g r a m m a r , second, we describe the process of converting such a g r a m m a r into P r o l o g code, a n d t h i r d , we describe the analysis of i n p u t w i t h such a g r a m m a r .
L H I P graxnmars are an e x t e n d e d f o r m of Pro- log D C G graznmars. The extensions can be sum- m a r i z e d as follows: a
1. one or m o r e [tHS clauses m a y be n o m i n a t e d as heads;
ZA version of LHIP exists which permits a form of negation on RHS clauses. That version is not de- scribed here.
2. one or more P~tlS clauses m a y be m a r k e d as optional;
3. 'ignore' rules m a y be invoked;
4. a d j a c e n c y constraints m a y be i m p o s e d be- tween l~tIS clauses;
5. a global t h r e s h o l d level m a y be set to deter- mine the m i n i m u m fraction of spanned i n p u t t h a t m a y be covered in a parse, and
6. a local threshold level m a y be set in a rule : to override the global threshold witlfin t h a t
" r u l e .
We provide a s y n t a c t i c definition (below) of a LHIP g r a m m a r rule, using a n o t a t i o n w i t h syn- t a c t i c rules of the form C -~ F1 I i f 2 - - - I Fn wtfich indicates t h a t the category C m a y t a k e any of the forms F1 to F,~. A n o p t i o n a l i t e m in a form is d e n o t e d by surrounding it w i t h square brackets ' [ . . . ] ' . Syntactic categories are
italieised,
while terminMs are underlined: ' . . . ' .A L t l I P g r a n u n a r rule has the form:
lhiVrute ~ [ - ] t e r m [ # T ] ~~__~> U~i~body
where T is a value between zero and one. If present, this value defines the local threshold fraction for t h a t rule. This local t h r e s h o l d value overrules the global threshold. The s y m b o l ' - ' before tile n a m e of a rule m a r k s it as being an ' i g n o r e ' rule. 0 n l y a rule defined this way can be invoked as an ignore rule in an RHS clause.
lhipbody => lh.ipclause
I Ihipclause z lhipbody
f lhipclause ; lhipbody
I lhipclause - lhipbody
I (~_" lhipbody ?_))
lhipclause ~ temn • te~nn ~. string •
._~ string
- term
[]
~rdoaeod~ }__
The s y m b o l ' * ' is used to i n d i c a t e a head clause. A rule n a m e is a P r o l o g t e r m , and only rules a n d t e r m i n a l items m a y act as heads within a rule body. The s y m b o l ' @ ' introduces a ter- minM string. As previously said, the purpose of ignore rules is simply to consume input ter- minals, a n d their i n t e n d e d use is in facilitat- ing repMrs in analysing i n p u t t h a t contains the false s t a r t s , r e s t a r t s , fdled pauses, etc. m e n t i o n e d above. These rules are referred to i n d i v i d u a l l y hy preceding their n a m e b y the ' - ' symbol. T h e y can also be referred to as a class in a rule b o d y hy the speciM I~.tIS clause '[]'. If used in a rule body, they i n d i c a t e t h a t i n p u t is potentially i g n o r e d - the p r o b l e m s t h a t ignore rules are intended to re- pair will not always occur, in which case tile rules succeed w i t h o u t conslmfing any input. There is a semantic r e s t r i c t i o n on the b o d y of a rule which is t h a t it m u s t contain at least one clause which necessarily covers i n p u t ( o p t i o n a l clauses and ig- nore rules do not necessarily cover i n p u t ) .
The following is an e x a m p l e of a L t I I P rule. Here, the sub-rule 'conjunction(Con j ) ' is m a r k e d as a h e a d a n d is therefore e v a l u a t e d before either
of,s(s))' or 's(S0':
s(conjunct(Conj, SI, Sr)) ~ >
s(St),
* conjunctlon(Conj),
s(S~).
tIow is such a rule converted into Prolog code by the L H I P system? F i r s t , the rule is read a n d the RHS clauses are p a r t i t i o n e d into those m a r k e d as heads, and those not. A record is kept of their original ordering, a n d this record allows each clause to be constrMned with respect to the clause t h a t precedes it, as well as with re- spect to the n e x t h e a d clause wMch follows it. A d d i t i o n a l code is a d d e d to m a i n t a i n a chart of known successes a n d failures of each rule. Each rule n a m e is t u r n e d into the n a m e of a Prolog clause, a n d a d d l t i o n M a r g u m e n t s are a d d e d to it. These a r g u m e n t s are used for the i n p u t , the start a n d end p o i n t s of the a r e a of tlm i n p u t in which
the rule m a y succeed, tile s t a r t a n d end points of the a c t u a l p a r t of the input over which it in fact succeeds, the n u m b e r of t e r m i n a l items cov- ered w i t h i n t h a t island, a reference to the point in the chart where the result is stored, and a list of pointers to sub-results. The converted form of tile above rule is given below (rMnus the code for chart m a i n t e n a n c e ) :
s ( c o n j u n c t ( H , I , J ) , A, B, C, D, E, F,
ELIK]-K, G) :-
lhip_threshold valuo(M),
conjunction(H, A, B, C, O, P, Q,
R - S , _ ) ,
s(l, A, B, fl, D, _, T, C-R, _).
s ( J , A. P, C, ~, E, U, s - E l ,
_),
F is U+Q+T,
F/(E-D)>=M.
The i m p o r t a n t points to note a b o u t this con- v e t t e d form are the following:
1. the c o n j u n c t i o n clause is searched for be- fore either of the two s clauses;
2. the region of the input to be searched for the c o n j u n c t i o n clause is the stone as t h a t for the rule's LIIS (B-C): its island extends from 0 to p and covers Q items;
3. the search region for tile first s clause is B-0 (i.e. from tile s t a r t of tile LHS search region to tile s t a r t of the c o n j u n c t i o n island), its island s t a r t s at D and covers T items; 4. the search region for tile second s clause is
P-C (i.e. from the end of the c o n j u n c t i o n island to the end of the LIIS search region), its island ends at E and covers II items; 5. the island associated with the rule as a whole
extends from D to E and covers F items, w h e r e F i s U + Q + T;
6. lhip_throshold_value/l unifies its argu- m e n t M with the current global threshold
value.
to be 'conjunction'. N a t u r a l l y , c o m p i l a t i o n m a y induce left-recursion as well as e l i m i n a t i n g it, in which case L I I I P will suffer f r o m the s a m e ter- m i n a t i o n p r o b l e m s as an o r d i n a r y D C G formal- ism i n t e r p r e t e d in this way. A n d as w i t h an or- d i n a r y D C G formalism, it is possible to a p p l y different p a r s i n g m e t h o d s to L H I P in order to circumvent these problems (see e.g. P e r e i r a a n d Shieber, 1987). A r e l a t e d issue concerns the in- t e r p r e t a t i o n of e m b e d d e d P r o l o g code. Reorder- ing of lZHS clauses will result in code which pre- cedes a h e a d w i t h i n a L t H P rule being e v a l u a t e d after it; judicious freezing of goals a n d avoidance o f unsafe cuts are therefore required.
L H I P provides a n u m b e r of ways of a p p l y i n g a g r a m m a r to i n p u t . The simplest allows one to e n u m e r a t e the possible analyses of the i n p u t w i t h the g r a m m a r . T h e order in which the results are p r o d u c e d wiU reflect the lexical ordering of the rules as t h e y are converted by L H I P . W i t h the t h r e s h o l d level set to 0, all analyses possible with the g r a m m a r b y deletion of i n p u t t e r m i n a l s can be generated. Thus, supposing a suitable gram- m a r , for the sentence John saw Mary and Mark
saw them there would be analyses corresponding
to the sentence itself, as well as John saw Mary, John saw Mark, John saw them, Mary saw them,
Mary and Mark saw them, etc.
By setting the t h r e s h o l d to 1, only those par- t i a l analyses t h a t have no u n a c c o u n t e d for ter- minals w i t h i n their spans can succeed. Hence,
Mark saw them would receive a valid analysis, as
would Mary and Mark saw them, p r o v i d e d t h a t the g r a m m a r contains a rule for conjoined NPs;
John saw them, on the other hand, would not. As
this e x a m p l e illustrates, a p a r t i a l analysis of this k i n d m a y not in fact correspond to a true sub- parse of the i n p u t (since Mary and Mark was not a conjoined s u b j e c t in the original). Some care m u s t therefore be t a k e n in i n t e r p r e t i n g results.
A n u m b e r of b u i l t - i n p r e d i c a t e s are p r o v i d e d which allow the user to c o n s t r a i n the b e h a v i o u r of the p a r s e r in various ways, based on the notions of coverage, s p a n a n d threshold:
lhip _ p h r a s o (+C, + S )
Succeeds if the i n p u t S can be p a r s e d as an instance of c a t e g o r y C.
lhip_ cv_phrase (+C, +S)
As for lhip_phrase/2, except t h a t all of the i n p u t m u s t be covered.
lhip_phras e (+C, +S, -B, -E, -Coy)
As for lhip_phrase/2, except that B binds to
the b e g i n n i n g of the island described by this a p p l i c a t i o n of C, E binds to the p o s i t i o n imme- d i a t e l y following the end, a n d Coy binds to the ntunber of ternfinals covered.
lhip_mc_phrasos (+C, +S, -Coy, -Ps )
The m a x i m a l coverage of $ by C is Cov. Ps is the set of parses of S by C with coverage Coy. l h i p _ r a i n m a x _ p h r as e s (+C, +S, - C o y , - P s )
As for l h ± p _ m c _ p h r a s e s ] 4 , except t h a t Ps is a d d i t i o n a l l y the set of parses with the least span.
lhip seq_phrase(+C,+S,-Seq)
Succeeds if Soq is a sequence of one or more parses of S by C such t h a t they are non- overlapping and each consumes i n p u t t h a t pre- cedes t h a t consumed by the next.
l h i p maxT_phras os (+C, +S, -MaxT)
MaxT is the set of parses of S by C t h a t have the tfighest threshold value. On b a c k t r a c k i n g it r e t u r n s the set with the next highest threshold value.
In a d d i t i o n , other predicates can be used to search the chart for constituents t h a t have been identified b u t have not been a t t a c h e d to the parse tree. These include:
lhip_success
Lists successful rules, indicating island position a n d coverage.
lhip_ms_success
As for lhip_success, but lists ouly the most specific successful rules (i.e. those which have themselves succeeded but whose results have not been used elsewhere).
lhip_ms_success (N)
As for lhip_ms_succoss, but lists only suc- cessful instances of rule N.
Even if a sentence receives no complete analysis, it is likely to contain some parsalfle substrings; re- sults from these are recorded t o g e t h e r with their p o s i t i o n w i t h i n the input. By using these predi- cates, p a r t i M but possibly useful i n f o r m a t i o n can be e x t r a c t e d from a sentence despite a global fail- ure to parse it (see section 4).
3 U s i n g L H I P
As previously m e n t i o n e d , L H I P facilitates a cyc- lic a p p r o a c h to g r a m m a r development. Suppose one is w r i t i n g an English g r a m m a r for the M a p Task Corpus, a n d t h a t the following is the first a t t e m p t at a rule for n o u n phrases (with appro- p r i a t e rules for d e t e r m i n e r s a n d nouns):
up(N,
D, A) # 0.5 ,-~--~>
determiner(D), * n o , , n ( N )W h i l e tiffs rule will a d e q u a t e l y anMyse simple NPs such as your map, or a missionary camp, on a N P such as the bottom right-hand corner it will give analyses for the bottom, the right-hand a n d
the corner. Worse still, in a long sentence it will
join d e t e r m i n e r s f r o m the s t a r t of the sentence to nouns t h a t occur in the l a t t e r hMf of the sen- tence. The n u m b e r o f superfluous anMyses can be reduced b y i m p o s i n g a local t h r e s h o h l level, of say 0.5. By looking at the various analyses of sentences in the corpus, one can see t h a t this rule gives the skeleton for n o u n phrases, b u t from the fraction of coverage of these parses one c,'m also see t h a t it leaves out an i m p o r t m l t feature, adjec- fives, which are o p t i o n a l l y found in noun phrases.
np(N, D, A)
#0.5 ~,-~>
determiner(D),
(? adjectives(A) ?),
* noun(N).
W i t h rids rule, one can now h a n d l e such phrases as the left-hand bottom corner, and a ba-
nana tree. Suppose further t h a t this rule is now
a p p l i e d to tile corpus, a n d then the rule is ap- plied a g a i n b u t w i t h a local t h r e s h o l d level of 1. By looking at items p a r s e d in the first case b u t not in the second, one can identify features of nolm phrases found in tlle corpus t h a t are not covered b y the current rules. Tiffs might include, for instance, phrases of the form a slightly dip-
ping line. One can t h e n go h a r k to the g r a m m a r
azld see t h a t the noun p h r a s e rule needs to bc changed to account for c e r t a i n types of modifier including adjectives a n d a d v e r b i a l modifiers.
It is Mso possible to set loom thresholds dy- namically, by m a k i n g use of the '{ prolog code }' facility:
np(N, D, A) # T ~,,~>
determiner(D),
(? adjectives(A) ?),
* . o ~ . ( N ) .
{
set_dynamic_threshold(A,T)
}.In this way, the strictness of a rule m a y be var- ied according to i n f o r m a t i o n originating either w i t h i n the p a r t i c u l a r r u n - t i m e invocation of the rule, or elsewhere in the current parse. For exam- ple, it would be possible, by providing a suitable definition for set_dynamic_threshold/2, to set T to 0.5 when m o r e titan one o p t i o n a l adjective has been found, and 0.9 otherwise.
Once a given rule or set of rules is stabl% and tile w r i t e r is satisfied w i t h the performtmce of t h a t p a r t of the g r a m m a r , a local threshold value of 1 m a y bc assigned so t h a t superfluous parses will not interfere with work elsewhere.
The use of the chart to store known results and failures allows the user to develop h y b r i d parsing techniques, r a t h e r t h a n relying on the default depth-first t o p - d o w n s t r a t e g y given by analysing with respect to the t o p - m o s t category. For instance, it is possible to anMyse the input in 'layers' of linguistic categories, p e r h a p s s t a r t - ing b y analysing noun-phrases, then prepositions, verbs, relative clauses, clauses, conjuncts, and fi- nally complete sentences. Such a s t r a t e g y allows the user to p e r f o r m processing of results between these layers, w:hich can be useful in t r y i n g to find the ' b e s t ' analyses first.
4 P a r t i a l r e s u l t s
The discussion of built-ln predicates m e n t i o n e d facilities for recovering p a r t i a l parses. Here we ilhlstrate this process, and indicate what further use might be m a d e of tile i n f o r m a t i o n titus ob- tained.
In the following example, tile chart is inspected to reveal w h a t constituents have been built dur- ing a t~iled parse of the t r u n c a t e d sentence Have you the tree by the brook t h a t . . . :
> lhip_phrase(~(S),
[have,you,the,tree,by,the,brook,that]). no
> lhip success.
( - 1 ) [ 4 - - 5 ) / l " ' > ©tree (-1) [ 2 - - 3 ) /1 " ' > ©you (4) [2--8) 14 - - >
np(nppp(you,
pp(by,np(the,brook,B)))) (4) [ 3 - - 8 ) / 5 " ' >
n p ( n p p p ( n p ( t h e , t r e e , C ) ,
p p ( b y , n p ( t h e , b r o o k , D ) ) ) ) (5) [ 3 - - 8 ) /2 " ' > r i p ( r i p ( t h e , b r o o k , A ) ) (5) [ 6 - - 8 ) /2 " ' > n p ( n p ( t h e , b r o o k , G ) ) (5) [3--5) /2 " ' > n p ( n p ( t h e , t r e e , E ) ) (7) [4--5) /I "'> noun(tree)
(8) [7--8) /I "'> noun(brook) (9) [ 2 - - 3 ) l 1 " ' > np(you) (10) [5--8) /3 - - >
pp(pp(by,np(the,brook,F)))
(11) [ 3 - - 4 ) / l "'> d e t ( t h e ) (11) [6--7) /1 "'> d e t ( t h e ) yes
E a c h rule is listed with its identifier ( ' - 1 ' for lex- ical rules), the island wtfich it has a n a l y s e d with beginning a n d ending positions, its coverage, and the r e p r e s e n t a t i o n t h a t was c o n s t r u c t e d for it. F r o m this o u t p u t it can be seen t h a t the gram- m a r manages reasonably well with n o u n phrases, b u t is unable to deM w i t h questions (the i n i t i a l a u x i l i a r y have remains u n a t t a c h e d ) .
Users will often be m o r e i n t e r e s t e d in the successful a p p l i c a t i o n of rules which represent m a x i m a l constituents. These are displayed by l h i p _ m s _ suc cos s:
> l h i p _ m s _ s u c c e s s . ( - 1 ) [ 1 - - 2 ) /1 " ' > ©have ( - 1 ) [ 8 - - 9 ) /1 " ' > ©that (4) [2--8) /4 - - >
np(nppp(you,
pp(by,np(the,brook,J)))) (4) [ 3 - - 8 ) /5 "'>
n p ( n p p p ( n p ( t h e , t r e e , I { ) ,
p p ( b y , n p ( t h e , b r o o k , I ) ) ) ) (5) [ 3 - - 8 ) /2 "-> n p ( n p ( t h e , b r o o k , K ) )
yes
Here, two u n a t t a c h e d lexical items have been identified, together with two instances of rule 4, which combines a NP with a p o s t m o d i f y i n g PP. The first of these has a n a l y s e d the island you the tree by the brook, ignoring the tree, while the sec- ond has a n a l y s e d the tree by the brook, consum- ing all terminals. There is a second analysis for the tree by the bTvok, due to rule 5, which has been o b t a i n e d by ignoring the sequence tree by the. F r o m this information, a user might wistt to
r a n k the three results according to their respec- tive span:coverage r a t i o s , p r o b a b l y preferring the second.
5 D i s c u s s i o n
The a b i l i t y to deal with large amomlts of possi- bly ill-formed text is one of the p r i n c i p a l objec- tives of current NLP research. Recent proposals include the use of p r o b a b i l i s t i c m e t h o d s (see e.g. Briseoe and Carroll, 1993) and large r o b u s t deter- ministic systems like Hindle's F i d d i t c h (Hindle, 1989). 4 Experience so far suggests t h a t systems like L I I I P m a y in the right circumstances provide an a l t e r n a t i v e to these approaches. It combines the advantages of P r o l o g - i n t e r p r e t e d D C G s (ease of modification, p a r s e r o u t p u t suitable for direct use by other p r o g r a m s , etc.) with the a b i l i t y to relax tile adjacency constraints of t h a t form&llsm in a flexible and d y n a m i c manner.
LIHP is based on the a s s u m p t i o n t h a t p a r t i a l results can be useful (often much more useful t h a n no result at all), and t h a t an a p p r o x i m a - tion to complete coverage is more useful when it comes with indications of how a p p r o x i m a t e it is. This l a t t e r point is especially i m p o r t a n t in cases where a g r a m m a r must be usable to some degree at a relatively early stage in its development, as is, for example, the case with the development of a g r a m m a r for the Map Task Corpus. In the near future, we expect to a p p l y LHIP to a different problem, t h a t of defining a restricted language for specialized parsing.
The rationale for the distinction between sanc- tioned and u n s a n c t i o n e d non-coverage of i n p u t is twofold. F i r s t , the qgnore' facility p e r m i t s dif- ferent categories of unidentified input to be dis- tinguished. For example, it may be interesting to separate m a t e r i a l which occurs at the s t a r t of the input from t h a t appearing elsewhere. Ig- nore rules have a similar flmctionality to t h a t of n o r m a l rules. In p a r t i c u l a r , they can have ar- guments, and m a y therefore be used to assign a structure to unidentified input so t h a t it m a y be flagged as such within an overall parse. Sec- ondly, by setting a threshold value of 1, L t I I P can be m a d e to p e r f o r m llke a standaxdly i n t e r p r e t e d Prolog D C G , t h o u g h somewhat more efficiently
due to the use of the chart. ~
A number of possible extensions to the sys- tem can be envisaged. Whereas at present each rule is compiled individually, it would be prefer- able to enhance preprocessing in order to com- pute certain kinds of global information from the grammar. One improvement would be to deter- mine possible linking of 'root-to-head' sequences of rules, and index these to terminal items for use as an oracle during anMysis. A second would be to identify those items whose early analysis would most strongly reduce the search space for sub- sequent processing and sc,'m the input to begin parsing at those points rather titan proceeding strictly front left to right. This further suggests the possibility of a parallel approach to parsing. We expect that these measltres would increase the efficiency of LHIP.
Currently, also, results are returned in an order determined by the order of rules in the grammar. It would be preferable to arrange matters in a more cooperative fashion so that the best (those with the highest coverage to span ratio) are dis- played first. Support for bidirectional parsing (see Satta and Stock, to appear) is another candi- date for inclusion in a later version. These appear to be longer-term research goals, however. 6
A c k n o w l e d g m e n t s : The authors would like to thank Louis des Tombe and Dominique Estival for comments on earlier versions of this paper.
R e f e r e n c e s
Anderson, A.tI., M. Bader, E.G. Bard, E. Boyle, G. Doherty, S. Garrod, S. Isard, J. Kowtko, J. McAllister, J. Miller, C. Sotillo, It. Thompson and It. Weinert (1991) "The IIC]LC Map Task Corpus",
Language and Speech
34(4), 351-366. Briscoe, T. and J. Carroll (1993) "Generalized Probabilistie LR Parsing of Natural Language (Corpora) with Unification-Based Grammars"Computational Linguistics
19(1), 25-59.Hindle, D. (1989) "Acquiring Disambiguation Rules from Text".
Proceedings of the 27th An-
nual Meeting of the Association for Computa-
tional Linguistics,
118-125.Sin large grammars there is a significant time gain. The chart's main advantage, however, is in identify- ing unattached constituents and allowing a 'layered' approach to analysis of input.
6Source code for the LHIP system has been made publicly available. For information, contact the authors.
Kay, M. (1989) "Head-Driven Parsing",
Proceed-
ings of the Workshop on Parsing Technologies,
52-62.
Pereira, F.C.N. and S.M. Shieber (1987)
Prolog
and Natural Language Analysis,
CSLI LectureNotes No. 10, Stanford University.
Satta, G. and O. Stock (to appear) "Bidirec- tional Context-Free Grammar Parsing for Nat- ural Language Processing",
Artificial Intelli-
gence.
van Noord, G. (1991) "Head Corner Parsing for Discontinuous Constituency",
