Hypothesis Selection in Grammar Acquisition 
Masaki KIYONO * and Jun'ichi TSUJII 
Centre for Compul, ational Linguistics 
University of Manchester Institute of Science and Tcchnoh)gy 
PO \]3ox 88, Manchester M60 1QI) 
United Kingdom 
kiyonoOccl, umist, ac. uk, tsuj i iOccl, umist, ac. uk 
Abstract 
This pat)er t)resents some techniques for selecting 
linguistically adequate hypol;heses of new gram- 
mat.foal knowledge to be used as resources of 
gralmnatical knowledge acquisition. In our frame- 
work of liuguistic knowledge acquisition, a rule- 
based hypothesis generator is inw)ked in case of 
parsing failures and all the possible hypotheses of 
Llew graliLlnar rules or lexical entries are gener- 
ated ffOln partial parsing results. Although each 
hypothesis could recover the (lef>cl, s of the exist- 
ing grammar, the greater part of hypotheses are 
linguistically unnatural. The techniques we pro- 
pose here prevent such unnatural hypotheses Dont 
being gene,'ated without discarding plausible ones 
and make the following corpus-based acquisition 
l)rocess In(ire ellieien(, and more reliable. 
1 Introduction 
Reusability of existing linguistic knowledge is the most 
import,~mt requirement for the rapid development of 
pra.ctical nal, ural \]augtlage l)rocessing systems. In or- 
der to realize, automatic customizatiou of existing lin- 
guistic knowledge to each applicat;ion domain, we pro- 
posed a new approach of linguistic knowledge acquisi- 
tion, which is a combination of symbolic and statistical 
approaches \[Kiyono and Tsujii, 1993\]. 
The fi:amework of our al)proach is shown in Figure 1. 
'1'111". acquisilion flow starl;s with executing the l)arse of 
each sentence in a corpus. If parsing lhiled, |,he 'tiy- 
poi;hesis Generator' produces the hyl)otheses of addi- 
tional gramnu~tical knowh;dge, each of which could re- 
cover t;lle incompleteness of the existing grammar Af- 
ter iterating t;his hypothesis generation process for all 
the senten('es in the corpus, the hypotheses are passed 
to the statistical analysis procc.ss and finally plausible 
hypotheses are chosen as new knowh'.dge by observing 
statistical properties of tile hypotheses. 
Unlike robusl; parsing \[Mellish, 1989; Goeser, 1992; 
l)ouglas and I)ale, 1 !)92\] or nou-statisl.ical alll)roach for 
grallunar a(:lluisil;ioll , our al/proach does Ilol; require 
a mechanism to detect tile cause of the parsing fail-- 
ure in the sentencial analysis phase and therefore the 
'Ilypothesis (~eneral;or' may output ;111 I,he possible hy- 
potheses, l\[owever, the greater part 1)t' hypotheses gen- 
erated by a simple deductive mechanism are unnatural 
revisions of the e.xisting grammar. For example, even ~ 
rule which derives a tot) node category ,9 direcl,ly from 
the input string of words might be hypol,hesize(I. 
*a.lso a staff member ol7 Mattsus|dt~t Electric lndustri~d 
Co.,Ltd., Shina.gawa, Tokyo, .IAPAN. 
Q__._ C°rlnts ) 
i 
17 k 
Ilypothesis \[ 
(lener;ttor 
C nyl,oth~s,: ,~ i~)-- 
..... i 
~ (~ra111111\[tr) 
St a.tistiea.I 
Analyser 
-<\[Iypothesis I)B) 
INdc-Bascd Processing Corpus-Betsed lSocessiu 
Figure 1: Framework of Graulmar Acquisition 
Linguistically unnatural hypqtheses have harnfful et: 
fects on lille lbllowing corpus-based process, not only 
making the process inefficient but, also int, erl'ering wil, ll 
statistical dnl, a as noise. In this paper, some techniques 
to remove such inadequate hytml, heses are proposed 
and the results of exl)eriments which show the efl'ec- 
l.iwme.ss of the proposed techniques are also discussed. 
2 Grammar Hypothesizing 
2.1 Grammar Formalism 
The grammar lbrlnalisnl we use is a conventional 
uniIication-based grammar. F, ach gramLnar rule is 
written in the form of a combination of a conl, ext-free 
rule and feature unilication functions. This R)rmalism 
is not specitic to any linguistic ~;heory, but we inl.ro- 
duced a number of concepts widely accepted in lil,- 
guisl, ic theories, such as grammatical flmctions, sub- categorization |'ralLies, 
aim X-bar theory. 
'FILe parsing system we introdttced 1,11 allply our 
grammar lbrlnalism is a sysl,enl called SAX \[Mat- 
smnoto, 1986\]. SAX uses the concepts of act, iw~' and 
inactive edges of Chart P~lrsing and analyses an in- 
put sentence with a bottom-up and parallel algoril;hm. 
As the grammar hyl)othesiziLLg algorithm is supposed 
l:o refer partial parsiug results of unsuccessflflly parsed 
sentences, we slightly modified SAX so that it ou/.lmts 
inactive edges as partial parsing results. 
837 
2.2 Basic Algorithm 
When SAX fails to parse a sentence, no inactive edge 
of category 5' sl/anning the whole sentencc exists in the 
parsing result. Grammar tlypothesizing is a process to 
introduce this inactive edge by augnlentmg the current 
gramnmr. The basic part of the hypothesis generation 
algorithm is written as fbllows: 
\[Algorithm\] An inactive edge lie(A) : a:o, '%1 Call be 
introduced from x0 to x,~, with label A, by each of 
the \]lypotheses generated by the following two steps. 
\[Step 1\] For each sequence of inactive edges, 
lie(B1): :r0,Xl\],..., \[ie(Bn) : .Vn_l,a:n\], spanning 
from x0 to x~, generates a yew rule. 
A :=> /31,-..,B n 
\[Step 2\] For each existing rule A -~ AI,...,An, 
find an ineonlplete sequence of inactive edges, 
lie(A1) : x0,xl\], ..., \[ie(Ai_l) : xi-~,Xi_l\], 
\[ie(Ai+l ): :ci, .r:i+l\] .... , \[ie(A,) : :,:,,_~, x,~\], and 
call this algorithm ibr \[ie(Ai) : xi-1, .vii. 
Feature Structures: A rule generated in \[Step 1\] 
could be a lexieal entry when this top-down algorithm 
reaches the bottom. As we adopted a unification- 
based grmnmar fbrmalism, we extended the algorithm 
so thai, it can hyt)othesize a feature structure of a lex- 
ical entry by observing surrounding successful cate- 
gories. As the algorithm works even ibr a eOml)lex 
feature like a subcategorization frame, it can be used 
to acquire a subeategorization dictionary. While some 
previous works on subcategorization fi'alne acquisition 
assumed very little prior knowledge concerning the 
classification of subcategorization frames \[Brent, 1991; 
Manning, 1993\], our apllroach assumes the existence 
of grammar rules Sllecifying subcategorization fi'ame 
assignment, which enables more accurate learning of 
subcategorization frames. 
Multiple Defects: In \[Step 2\] of the algorithm, it 
is SUl)t)osed that each unsuccessfully parsed sentence 
has exactly one cause of failure but a sentmtee in actual 
texts often contains two or more causes of failure (for 
example, two unknown words). To solve this problem, 
we extended the algorithm so that it searches for a 
multiple hypothesis which is a set of rewriting rules 
and lexlcal entries. 
3 Hypothesis Selection 
3.1 Basic Grammatical Constraints 
From a linguistic point of view, hypotheses generated 
by the algorithnl given above might contain nlany un- 
natural hypotheses because the algorithm itself does 
not have any linguistic knowledge to judge the appro- 
priateness of hypotheses. To remove unnatural hy- 
potheses, we have introduced the following criteria 
\[Kiyono and Tsujii, 1993\]. 
• The maximum number of adjacent unsuccessful cat- 
egories is set to 2 in order not to decrease the efIi- 
ciency of the algorithm. 
• The lnaxilnuln nulnber of daughter nodes is set to 3. 
• Supl)osing that the existing grammar contains all 
the category conversion rules, a mmry rule which 
has only one daughter node is not generate.d. 
• Using generalizations embodied ill the existing 
grammar, a hypothesis contaiuing a sequence of 
subnodes which are collected into a larger category 
by existing grammar rules is not generated. 
• Distinguishing non-lexieal categories from lexica.l 
categories, a hypothesis whose mother category is 
a lexical category is not generated. 
• Assuming that tile existing gramznar has a complete 
set of fllnctional words, a lexical hypothe~sis is re- 
stricted to the open lexical categories, such as noun, 
verb, adjective, and adverb 
3.2 Constraint based on /meal Boundaries 
A new constraint on the violation of the boundary 
condition given to phrases was introduced to avoid 
any collection of adja/:ent successfifl categories in rule 
hypothesizing. The bomnlary condition is given by 
putting parentheses at both ends of a phrase, such 
as a noun phrase, a verb l)hrase, and a prelmsitional 
phrase. This constraint tilters out a hylmthesis which 
crosses either end, not both ends, of a phrase. For 
example, when parentheses are put like "\[Tile default 
blocking factor\] is \[20 blocks\]", a hyl)othesis 'VP ='e 
VP, NP, VE'R.BBIs"' covering "blocking factor is" is 
discarded because of the violation of the boumlary con 
dition of a noun phrase "The defimlt blocking factor" 
This constraint requires the hunlan task of putting 
parentheses before the hypothesis generator is invoked. 
hi comparison with writing a constituent structure of 
tile whole sentence, this work is much easier because 
we have only to give parentheses to delinite phrases. 
Moreover, instead of giving parentheses by hand, we 
can even obtain various tagged corpora. 
As this constraint is also atlplieable to other con- 
stituents of the input sentence, it might improve the 
etliciency of the top-down hypothesizing algorith m. 
3.3 Constraint based on X-bar Theory 
Most of the criteria in 3.1 are based on linguis- 
tic category classification but none of them com- 
mits itself to dealing with the rcla.tionship among the 
nlother node and the daughter nodes. For example, 
supposing the existing granmmr does not contain a 
rule for participial adjuncts ill t10UU phrases, the hy- 
pothesizing program generates a new rewriting rule 
'NP ~. VP, NP' t¥om the phrase "blocking tactor" 
in the sentence "The default blocking factor is 20 
blocks". Ilowever, tile program also general,es other 
alternative hyt)otheses from the same phrase, such as 
'PP ~ VP, NP', 'INFINIg'IVI'~ ~ VP, NP', and 
'THAT_CLAUS'I,/~ VP, NP', each of which derives 
a. 1lost-positional adjunct for "default" by believing 
"default" is a head noun of the noun phrase. Liuguisli- 
cally, such combinations of nlother nodes and daughter 
nodes are not allowed. 
As a general l)rinciple for explaining phrase struc- 
tures, X-bar lheoryis widely accepted. According to 
X-bar theory, a grannnar rule is (or can be converted 
to) either of the following forlns, where each prime(') 
expresses the l)rojection level of' a head X. The l)rojec - 
tion level increases as gramnlar rules are applied and 
X" is called a maa:imal projection of that category. U 
and W are adjuncts of X' and should I)e maximal pro 
jeetions of some ca.tegories. 
838 
X" ~ YX'Z 
X' -> HXW 
ItLhe cxisi.ing grammar is wriLl,en in X-bar L\]mory, 
l,his const\]'aint is drastically etl'(,cliv(~ in reduciug Li.~ 
mmd)er of hyi)()Lhes(~s. 
3.4 Plausibility of Hypotheses 
Among the hypo|,h(~scs which passed throug, l~ all th(- 
C()llSi;lll,i\[\[l;S, ea, c}l oI\[c \[\[;bq ll~ (till'er(ml, plausibility as 
gramnl;~l;icM knowl(~dg('.. Assumin I, l;\[laI, the cxisl;iug 
grammm' is rtmsonahly COmln'(~hensive , lcxical or id- 
iosyncratic km~wledg(~ should be lllOl:(! i)lausil)le than 
gen<:ral r(~wril,iug rules, lu oMcr 1;o eulphasizc this ten- 
(hmcy, each hyl)othcsis is given the rolhm, ing i)lausihil. 
ity value. 
W ( I1 !q)oi) x II ( I1 ypo,:) :'(/::vj,o~) 
: l- Iv(,v) x n(,~') 
This wdue is relat(~d Io tshc I)rOl)orl, ion of tlw si/e, or 
tim 1)roducl, of the width and l,he hcighl,, of the m\[l)l,r(,c. 
(:omt)osed by l,he hylmthesis in the whole sLrll(:l;llrO, o\[' 
the senl;(mcc. The wduc ral~ges I'rom 0 l,o 1 and gel,s 
bigger i\[' the hypol, hesis cowers a smaller pro'l, o\[" l,\[l(~ 
seHl,ew::e, The widLh or t,h¢~ hyl)ol, hcsis , 14:(llypoi), is 
delined as l;he woM cot\[l|t ,Jl" t,hc sul)lrce aml lhe h(~ight 
II(llgpoi) is as t, he shorl,est, path I'rom h!xical uodes I,o 
the Lo 1) node ol: I, hc sHt)l, re(!. 
4 Experiments 
4.:1 Corlms 
In order to check I, he eIl'ecl;s of the hyp(Mwsis sclecl,ion 
tcchuiqucs, wc carried out some eXlw.rim(ml;s wil,h I, hc 
UNIX I.)ll lille \[11111\[11111. 100 ,q(!ItLCIICOH W(!I'O chos(211 ;~ls HII 
ext)('~rimt!nLal set from the ll\[alttl;t\[. 'l'h(~ ch;u'acl;(~risl.ics 
ol7 this Col'pUS are as fl)llows. 
• Number o\[' s(mLen(:(~s: 10ft 
• L(;ngl.h of senl.cnces: 9.08 w,)rds (av(!rage) 
• Number of dilrcr(mt words: 381 
• I,'xaull)lcs: 
'l'}ler(~ is \]no (:Seal)C: scquc~ll(:c l, hal, prints a 
(Iot\[l)\[e-.(lUOl;(~= 
IJsc Lhc llCXI. ;ll'gUlllt~lt{, ~IS l.he hlocldng fact;or 
\['or t;al)C r(~cor(Is. 
The' d('faull, I)lo(:killg factor is 20 I)lo(:ks. 
4.2 Given (~rammaticai Knowledge 
Two set;s (71' grammar rules were prel)ar(~d ti)t' l.h(~ cx 
perimenl,s, (:ramma'v A and (/ru'ntntar II. (ll?;tllll\[lar A 
contaius 118 r(~writing rules LhaL ('(~vcr basic exprcs 
sions o1' I,;I@ish. (fl'amma.r l~l is a subset of Grammm" A 
and (:on(.ains only 25 rewriting rules. The conl,eul,s o1' 
(~ra\]\[llll,q.r A all(l (\]l';\[llllllar \[{ ;\[I'C ,~howll ill 'rablc 1. 
The dictionary we use is th(' I','\])1~ f';~\[glish l)iclioo 
nary containing 200,000 enl,rics, 'l'hc eifl, rics of Lhis dic- 
Lionary ar(, not wril.L(m in t.lm I;:)rm o1" a feal.ur(~ sLruc.. 
1,\[11'(: t)llt hav0 l,h(~ cllcodcd ild'ormation ()f Lhe SyllL~ctic 
pM;terns, which wc interpr(% its a f(~at, m'e sl, ru(:Lnre. As 
Lhe 1';1)1{ I)icl, ionary was d(;vclol)(>d as a masl,cr dicLio- 
nary for various al)plications , il; took ht 1;}1(2 il\[forilm- 
tioH concerning all 1,he apl)(mran('es of ca(:h word wit\]> 
out scre(!niug 1)y h'(!quen(:ies. Th is characl;crisl,ic of t,lm 
S(!\[\[ I,(;ll(:(? 
Verh Phras(~ 
Noun l'hras,~ 
I'rel)osit, ionM I'hrase 
A(Ijectivc 1)tu'as( ' 
Adverbia.l \]'hrase 
Inlinitive Claus(! 
'l'\]t;d; ( Jhulse 
Rchttive Clause 
Sul)ordinate ( ',l;ttls(, 
2:1 1 
4O 12 
27 Y 
2 1 
9 1 
5 I 
4 l 
1 l 
(i 0 
1 0 
118 25 'l'(fl, al 
'l'abh'. \[: I{.ule (;ounLs of Two (~l:H\[llll\[~ir SC\[,S 
EI)IL I)icl, ionary i\[lcreascs I;h(~ aml)iguity of parsing. \[u 
fact, each word wil, hin the Saml)h', s(mt('~uces I'r()m IJw 
I;NIX mamIM has 1.49 l)arts of speecl, in l;he I:,l)l/ 
I)icl, iona, ry while l;hc same vahlc is IAI according I,o 
the COL/)INS COIl UII, I) I)iclio~lar:ll. 
4.3 Generated Ilypotheses 
(4choral ()utcoll\[m: 'Fh<: CXl)r.rimcnts ()\[ gencraLiug 
hyl)ot, hen(~s were carried ()ttl, with (;rammar A umh~r 
thrt,.e (li\[li~rent ctmdMons, (a) using tim basic grammai.- 
ical (:onsl, rMnLs only, (h) adding Lhe conslraiul with 
local I)hras;d boundaries giv(m as l)areld;h(:ses, and (c) 
adding l;tte consLraint wit,h X-I)ar theory, To carry 
out exp(,rimcnts (h) aud (c), within 1,he targt,t sen- 
tl!ll(t(!S, i)arellLh(!s(~s weft! given t() llOllll l)llr;is(~s, ill\[illi-. 
Live chmses, that-cla, uses, aml rod)ordinate clauses A 
lmrt of I.lw resltll, of (~Xl)erimcnt, (a) is shown in Tahle 2, 
e;.:h c()hmm of which disl)lays l, he numb('r of hyl)othc- 
sos g(!nm'atcd '\['lm cohmms 'SiHgle' and 'MultilflC ~ 
,~d.)w I,h(~ mmthcrs of single and multilflC hypol, h(:s('.s 
resl)eCl,ivcdy. 
The r(~sull, s of the t;hre(~ CXl)erimeul,s art. summarized 
in Table 3. The i)arser failed Lo amdyse 61 out of 
100 senL(m(:es and I,he grammar hyl)olJ,esizing program 
was iuvoked lot 1,hose scnl,tm(;(:.~;. While no hypol, heses 
were g(mera, Lt~d \[rom 2(1 or 30% of unsu(:ct!ssfully parsed 
s(mt.ca~ces I)ccause tim current hypothesizing algonil, hm 
does m)t allow verth:M duplical, ion of incOnll)l(%cncss 
and also because I,he l)armH(d,crs of I, hc basic gra, mnmt- 
ical (:ons|,raints do 11ol, allow th('. (':xisl,Cll(:(: ()\['lllor(; l, ll;lll 
l.wo adjaceul; inc.omplet,~ nodes, Lh('. results on 1,he nttm. 
l)crs ot acLual hypotheses made show t\]ml, lhc ,q|,r()llger 
l;hc coHsl.raint, we pose, Lhc l}'~wer hYl)ol.heses are gell(W 
al,(!(I. The average hypol, lieses per S(;ll\[,(!ll(;(!, cah:ulatcd 
1)y d ivi(ling l,\]le total hyl)ol,hc,~;is (:omfl; of 1,30l in (a), 
708 in (b), and 23l ill (c), I)y the number or actual 
scntt'.nccs From which llyl)ol, hcses wer(~ g(m(:\]'al,cd, 50 
in (a), .~4 in (I,), and 41 i,, ((:), ,w,~ ,.~,h,c(~d r,.(,,,, ~.0 
t;() 5.6. 
Ill ,~ollie cas(:s, a\]\[ 1,11(~ hyl)oLh(,scs ;IFC l'(~lllOV(!d I)y 
ncwly introduced COllSl, r;tilli,s, 6 Sclll,(!lic(!s by Lhe local 
boundary consl,l:ailfl, and 3 more ,q(~llliellC(!S I)y I,he COil- 
sl;raint o1" X-I)ar I,hcory. hlvestigation of the iuiLial set. 
of hyl)oLh(~scs .g(:uer;ttct\] t'rOllI Sll(:ll SOH{,OIlC(~S r(wc;41(?d 
thaL no plausible hypot,hesis was included iu it. '\['h(~rc 
%re, Lhcse seni;ences are liOt; criLical to 1,he hyl)oLlmsis 
seh:cl;ion nmt, hod we inl,roduce(I. 
In tim linal sel; o/' hypol.heses, 3{) plausible hyl)(ll.h(! 
839 
Sentence II Single ~_ Multiple II T°t'all 
Lex Rule Lex Mixed Rule 
Tie efault oc ing factor is 20 ~ "~1800021 
The output device in use is not capable of backspacing. 
l{emove initial definitions for all predefined symbols. 
The escaped NEWLINE is not included in the macro value. 
Components of an expression are separated by white space. 
The name of this directory is listed in tile fohler wu'iable. 
The name of the editor is listed in tile EDITOR variable. 
Table 2: }'art of tile Result of Experiment (a) 
No. of Unsuccessfully l'arsed Seutences 
No. of Sentences which generated No llypothesis 
No. of Sentences which generated Single Ilypotheses 
No. of Sentences whieh generated Multiple IIypotheses 
No. of Sentences which generated Plausible Ilypotheses 
Rank of Plausible llyl)otheses (Average) 
No. of llypotheses (Total) 
No. of ttypotheses (Average) 
Experiment (a) EXllerin,eut (b) l':xperiment (e) 
61 61 6I 
11 
43 
7 
' 33 
7.4 
17 
39 
5 
32 
2.8 
708 
16.1 
1301 
26.0 
20 
37 
4 
30 
1.6 
231 
5.6 
Table 3: tlypotheses Generated fi'om Different Conditions 
ses, 7 new rewriting rules and 23 new or modified lex- 
teal entries, remained without being filtered out by 
newly introduced constraints Some of the plausible 
hypotheses are listed below. 
New Rule: np => np,adjp. 
New Rule: nil =¢. np,np~ 
New Rule: np =¢, vp,np. (from 3 sentences) 
New Rule: up => vppsv,np. 
New Rule: vp :e~ vp,p. 
New Lexical Entry: n => \['DF, LI'~TE'\]. 
New Lexical Fmtry: n => \[patlummes\]. 
Modified Lexical Entry: v => \[default\]. 
Modified Lexical Fmtry: adj => \[invisible\]. 
Modified Lexical Entry: adj => \[callable \]. 
New Lexical Entry: adv => \[recursively\]. 
The weighting function explained ill 3.4 was not 
used for selecting hypotheses but tim validity of it was 
proved by counting the order of each plausible hypoth- 
esis in the set of generated hypotheses. The row of 
'Rank of Plausible IIypotheses' in Table 3 indicates 
that plausible hypotheses stand much higher than the 
middle of the order. 
ExamI)les: IIereafter, in order to show how hy- 
potheses were selected by each constraint, we explain 
tile results for some typical examples. 
Ex.1) "The default blocking factor is 20 blocks." 
As Grammar A does not contain a rule for participial 
adjuncts, the parser fails to analyse the noun phrase 
"tile det~ult blocking fact, of' and the grammar hy- 
pothesizing program is iuvoked. While this program 
generates 21 hypotlmses in experiment (a), it filters 
out the following 12 hypotheses in experiment (b). 
While checking local boundary violation, the pro- 
gram removes those grammatically unnatural com- 
binations of categories, though it does not use any 
linguistic knowledge. 
New Rule: advp => np,vp. 
New Rule: infinitive => np,Vpo 
New Rule: infinitive => vp,np,vp. 
New Rule; np => s,vp,np. 
New Rule: np=> vp,np,vpo 
New Rule: pp => np,vp. 
New Rule: pp => vp,np,vp. 
New Rule: that_clause => vp,np,vpo 
New Rule: vp => vp,np,auxbe. 
New Rule: vp --> vp,np,s. 
New Rule: vppsv => vp,np,auxbe. 
New Rule: vppsv => vp,np,vp. 
Moreover, the l)rogram filters otlt (;lie following 4 
llypothese with tile constraint of X-bar theory. 
New Rule: infinitive => vp,npo 
New Rule: pp => vp,np. 
New Rule: that clause => vp,np~ 
New Rule: vppsv => vp,np. 
Finally, the fol\]owing 5 l~ypotheses, alnong which the 
expected hypothesis 'NI'=> VP, NI" still remains 
are generate.d 
Modified Lexical Entry: n => \[factor\]. 
New Lexical Entry: adv => \[factor\] 
New Lexical Entry: n => \[blocking\]. 
New Rule; np => vp,np. 
New Rule: np => s,vp,np. 
Ex.2) "Tile output dcviee ill use is not capal)h~ of 
backspacing." 
This sentence is also parsed unsuccessfully beeause 
the current version of the EI)R 1)ictionary does not 
have information that "capable" subcategorizes a 
prepositional phrase. Among the initial set of 30 
hyl)otheses , the following 8 llypotheses pass through 
840 
No. of Unsuccessfully Parsed Sentences 
No. of Senl,ences which generated No Ilypothesis 
No. of Sentences which generated Single llypotheses 
No. of Sentences which generated Multiple llyl)otheses 
No. (if Sentences which generated Plausible Hyl)otheses 
No. of Ilylmtheses (Total) 
No. of llypotheses (Aw'.rage) 
(;raimnar A Grammar B 
97 61 
11 
43 
7 
31 
1301 
26.0 
45 
41 
11 
16 
550 
10.6 
Table d: Ilylmtheses Generated from Two Grammar Sets 
the constraints o1' local bomMaries and X-bar the- 
ory. The lirst hypothesis iu the list is the plausi- 
ble hypothesis obtained in search of the real cause 
of the feature disagreement between "capable" and 
"of backsl~aeiug". This lexical hypothesis for "capa 
hie" contains a modified version of its subcategoriza- 
lion frame so that it subcategorizes @prepositional 
phrase. 
NodJ_*ied Lexical Entry: adj => \[capable\]. 
New Lexical Entry: n => \[capable\]. 
New Lexical Entry: v => \[capable\]. 
New Lexical Entry: v => \[not\]. 
New Rule: adjp => neg,adjp. 
New Rule: adjp => neg,adjp. 
New Rule: s => s,adjp. 
Re~r Rule: vp :> vp,p. 
4.4 tIypotheses from Smaller Knowledge 
Another experiment was pertbrmed with (;ramlmtr B 
under the basic grammatical eoustraints in order to 
compare 1,he effects of the maturity of existing gram- 
ma.tieal knowledge. The nmnl)ers el' hypotheses gener- 
ated from two grammar sets are shown iu Tal)le 4. 
The eoverage of (lranunar II is so limited that 97 ()tit 
of 100 sentences were parsed unsuccessfully and passed 
to the Ilypol, hesis Generator \[lowew!r, as the imma- 
turity of Grammar B also al\[ects the number of gen- 
erated hypotheses, the mmlber of plausible hypotheses 
among the 550 hyl)otheses (10.6 hyl)otheses per sen- 
tence) generated l'rom 97 sentellces was only I(L This 
result claims that cyelie acquisition of grammatical 
knowledge is wdid. I",w'.n the sentences frolu which no 
hylmt,heses are gellerated with a small grammar would 
be taken into consideration ill a. later acquisition cycle 
with a larger grammar. 
5 Conclusion 
This paper l)rOl)osed l,eehuiques for selecting appro- 
priate hyl)¢)these.s in (.he rule-hased l)roeessing stage 
Of grallllllar aeqllisitioll. The experinlents tt) (;X~Llnille 
the etl'eets o\[' these techniques indicate that (,hey have 
several advant;~ges 
• The newly introduced COllstraiuls reduce the nmn- 
bet of hypotheses per sentence, from 26.0 1;o only 
5.6, small enough to be treal;ed in a corpus-based 
processing enviromnent. This hyl~othesis selection 
is done without discarding plausilAe hypotheses. Al- 
though, all the initial hypotheses may be, in certain 
cases~ removed by the lle.W constraints, this hapl)ens 
only if no plansilAe hyl)othesis is inch,tied iu the ini 
tM set. 
• Even if no hylmthesis is generated from all unsuc- 
cessfully l)arsed sentence (20 out of 61 sentences 
ill experiment (el) or no plausible hypothesis is ill- 
eluded ill I, he initial hypothesis set (11 out of 41 
sentences in experinmnt (el), a plausible hypothesis 
will be generated in the later acquisition cycle al: 
ter adding grammatical knowledge vital lot the sen- 
tenee~ 
• Among the generated hypotheses, lexical hypotheses 
arc more Idausible than rule hypotheses (2a out of 30 
plausible hypothese.s were lexical in expermmnt (el). 
This fact amahs l;hat the. grammar used R)r the ex- 
perilnents has an almost sullieient set of rewriting 
rules and that, after t;he grammar reaches such a 
mature situat;ion during the acquisition cycle, only 
lexieal or idiosyncratic knowledge has to be added. 
As our method has a facility to hypothesize a lexical 
entry with it, s l~;aI, ure structure including a subcate- 
gorization frame, we can set the target of acquisition 
only to lexieal knowledge for a large dictionary. 
,, '\['he loeaI boundary const;raint was introduced for 
automatie hypothesis selection, but it might also be 
used in an interactive debugging tool lbr grammar 
lllailltellanee. 
References 
\[lh'ent, 1991\] Michael R. Brent. Automatic Acqui- 
sition of Subeategorization Frames from Untagged 
Text. In Prec. of th.e ~29st ACL meeting, pl).209 
2t4, 1991. 
\[l)ouglas and I)ale, 1!)92\] Shona l)ouglas and l{obert 
l)ale. Towards Robust PNI'IL In Prec. of COLING- 
02, pp.468 474, 1992. 
\[Goeser, 1992\] Sel)astian Goeser. Chart Parsing of II.o- 
I)ust (~h:amnmrs. hi Prec. of COLING-92, pp.120 
126, 1992~ 
\[Kiyono and 'l)sujii, 1993\] Masaki l(iyono and aun'ichi 
Tsujii. Linguistic knowledge acquisition from pars- 
ing taihu'es. In Prec. oJ£'AUL-93, pp.222-231, 1993. 
\[Maturing, 1993\] (Jhristopher I). Manning. Automatic 
Acquisition of a I,arge Subcategorizatiou l)ictionary 
from Corpora. in PTvc. oJ' lhe 31st ACL mccling, 
1)p.235 242, 1993. 
\[Matsumoto, 1986\] Yuqi Matsumoto. A I'm:aiM I'ars- 
ing Sysl;em for Natural Language Analysis. \]n 
Lechtrc Noles in Uompuler Science 225, Springer- 
Verlag, pp.39d 409, 1986. 
\[Mellish, \[989\] Chris S. Meltish. Some Chart-hased 
Techniques for l:'axsing I\]l-\[brmed Input. In Prec. 
oJ lhc 27th ACL meeting, pp.102- 109, 1989. 
841 
