Aspects of Pattern-matching in Data-Oriented Parsing 
Guy De Pauw 
CNTS 
University of Antwert) 
Abstract 
Data-Oriented Parsing (DOP) ranks mnong the best pars- 
ing schemes, pairing state-of-the art parsing accuracy to 
the psycholinguistic insight that larger clmnks of syn- 
tactic structures are relevant grammatical and proba- 
bilistic units. Parsing with the DOp-model~ however, 
seems to involve a lot of CPU cycles and a consider- 
able amomtt of double work, brought on by the concept 
of multiple derivations, which is necessary for probabilis- 
tic processing, lint which is not convincingly related to a 
proper linguistic backbone. It is however possible to re- 
interpret the poP-model as a pattern-matching model, 
which tries to maximize the size of the substructures 
that construct the parse, rather than the probability of 
the parse. By emphasizing this memory-based aspect of 
the DoP-model, it is possible to do away with multiple 
derivations, opening up possibilities for efiqcient Viterbi- 
style optimizations, while still retaining acceptable pars- 
ing accuracy through enhanced context-sensitivity. 
1 Introduction 
The machine learning paradigm of Memory- 
Based Learning, based on the assumption that 
new problems are solved by direct refbrence to 
stored experiences of previously solved prob- 
lems, has beest successfully applied to a number 
of linguistic phenomena, such as part-of-speech 
tagging, NP-clmnking and stress acquisition 
(consult Daelemans (1999) for an overview). 
To solve these particular problems, linguistic 
information needed to trigger the correct dis- 
ambiguation, is encoded in a linear feature 
value representation and presented to a mem- 
ory based learner, such as TiMBL (Daelemans 
et al., 1999). 
Yet, many of the intricacies of the domain of 
syntax do not translate well to a linear repre- 
sentation, so that established MBL-methods are 
necessarily limited to low-level syntactic analy- 
sis, like the atbrementioned NP-chunking task. 
Data Oriented Parsing (Bod, 1999), a state- 
of-the art natural  parsing system, 
translates very well to a Memory Based Learn- 
ing context. This paper describes a re- 
interpretation of the soP-model, in which the 
pattern-match, infl aspects of tim model are ex- 
ploited, so that parses are analyzed by trying 
to match a stew analysis to the largest possible 
substructures recorded in memory. 
A short introduction to Data Oriented Pars- 
ing will be presented in Section 2, followed by an 
explanation of the term pattern-matehin9 in the 
context of this paper. Section 4 describes the 
experimental setup and the corlms. The parsing 
phase that precedes the disambiguation phase 
will be outlined in Section 5 and a description 
of the 3 disambiguating models, POFG, PMPG 
and the combined system PCFG@PMPG (:an be 
found in Sections 6, 7 and 8. 
2 Data Oriented Parsing 
Data Oriented Parsing, originally conceived by 
Remko Scha (Scha, 1990), has been successfully 
applied to syntactic natural  parsing 
by ll,ens Bod (1995), (1999). The aim of Data 
Oriented Parsing (henceforth DOP) is to develop 
a per\[ormanee model of natural , that 
models  use rather than some type of 
competence. It adapts the psycholinguistic in- 
sight that  users analyze sentences us- 
ing previously registered constructions and that 
not only rewrite rules, but cornt)lete substruc- 
tures of any given depth cast be linguistically 
relevant milts tbr parsing. 
2.1 Arehiteeture 
The core of a DOP-system is its TREEBANK: an 
annotated corlms is used to induce all substruc- 
t, ures of arbitrary depth, together with their re- 
spective probabilities, which is a expressed by 
236 
S J~ 
NP \111 
I 
Peter 
VP 
killed NP 
i~ raccooll 
S NP \;P 
NP VP Peter killed NP 
NI' 
.~l raCCOOll 
Figure 1: Multiple l)eriw~tions 
its fl:equency in the TREEBANK relative to l;he 
numl)er of substructures with the Sanle root- 
node. 
Figure 1 shows the coral)|nation ol)eral;ion 
that is needed to tbrm the correct l)arse tree 
for the sentence Peter" killed a raccoon. Given a 
treet)ank of substructures, the systcln tries to 
match the leftmost open nod(; of a substruc- 
ture |;hat is consistent with the parse tree, with 
the top-node of another sul)structur(;, consistent 
with the parse tree. 
Usually, ditferent conlt)inations of sul)struc- 
tllrO.s are possible, as is i~l(ti(:ated in Figure 
1: in the examl)le at the left-hand side the 
tree-structure (:an t)e built l)y (:o11111ining all S- 
structure wil;h a st)coiffed NP a.lld a flllly spe(:i- 
fled vp-structure. The right example shows an- 
other possible Colnl)ination, where a parse tree is 
1)uilt t)y conll)ining the \]ninimal sut)s|;rltcl;ures. 
Nol;e that t\]\]cse are (:(msisl;(mt wit\]l ol'dinary 
rewrite-rules, such as s -+ NP VP. 
One t)artit:ul;~r 1)~trse tree may t;hus (:()\]lsist ()f 
several (lill.(u'(ml; deriva, tio'n.s..To lind l;hc 1)rot) - 
al)ility (If ;I, (terivation, we lnultit)ly tim t)rot)a- 
1)ilities of the substructures thai; were used to 
l.()rm the derivation. To lind the t)robal)ility of 
a parse, we must; in tlrilmit)le sum the t)rol)at)il- 
ities of all its deriw~tions. 
It is COlnl/utationally hardly tra(:tat)h; to COil- 
sider all deriw~tiolls t.()r each pars('. Since 
VITF, RBI ol)timization only su('(:ceds in finding 
the most 1)robal)h'~ (teriw~tion as opposed to the 
most 1)robal)le l)arse, the MONTE CARLO al- 
gorithm is introduced as a proper al)proxima- 
tion I;hat randomly generates a large nlmfl)er of 
deriw~tions. The most prol)al/le l)arse is (:onsi(t- 
ered to be the parse that is most often observed 
in this derivation forest. 
2.2 Experimental Results of HOP 
The basic 1)op-model, POP1, was testc,(t (111 
a manually edited version (if the ATIS-corlnlS 
(Marcus, Sant(lrini, and Marcinkiewicz, 199a). 
The syst;eln was trained on 603 Selltelmes (t)arl; - 
ofstmech tag sequelmes) and (;wfluated on a test 
set (if 75 SCld;ences. Parse accuracy was used as 
an evahlation metric, expressing t;11(; percentage 
of sentences in the test set for which the tlarse 
l)rOl)osed by the system is COlnpletely identi- 
cal to the one in l;lle original eort)us, l)ifl'er- 
eat exl)erilnents were conducted in which max|- 
11111111 sul)structure size was varied. With DoPl- 
lillfited to a sul)sl;ructure-size (If 1 (equiw~lenl; 
1;O a PCFG), t)arse accuracy is 47%. hi the (/p- 
l;ima\] D()l'-mo(lel, in whi(:h sut)stru(:ture-siz(; is 
1lot limited, a 1)arse accuracy of 85% is (ll)- 
tni\]lc(t. 
2.3 Short Assessment of DOP 
DOI'I in its ot)tinlal fornl achieves a very high 
parse accuarcy. The comt)utational costs of the 
syste111, however, are equally high. Bed (19951 
reported an average t/arse tilne of 3.5 hours 11(;1 . 
Sellte.n(:e. Even though (:urrent 1)arse tilne is 
rcl)ortc.d to l)e 11,or(; reasollal)le, tile oi)timal 
D()P algoril:lml in whi(:h n(/('onstr;dlts are made 
on tll('~ size (1t' sut)structures, nlay not yet 1)e 
tract;able for life-siz( ~. COl'l)()ra. 
In a context-free grammar framework (con- 
sistent with \])()P limited to a sutlstru(:tm:e-size 
(If 1), there is only (me way a t/arse tree can 
t)e t'ornmd (t'(/1: exalnl/le, the right hand side of 
Figure \]), nleaning that there is Olfly one del:iva- 
tioll for a given 1)arse tree. This allows efficient 
VITEll.BI style Ol)tillfization. 
To elmo(le (:ontext-sellsitivity in the systeln, 
DOP is tbr(:ed to introduce multiple deriw~tiolls, 
so that repeatedly the same l)arse tree needs to 
1)e g(;lmrated, l)rillging at/(/ut a lot of COlll\])llta,- 
tional overhead. 
Even though the use of larger syntactic coil- 
texts is highly relewmt fl'om a psycholinguisI,ic 
t)oint-ofview, there is 11o explicit l)reference l)e- 
ing lnade t'(/1' larger substructures in the DOP 
nlodel. While the MONTE CARLO optimizatiolx 
scheme nlaxinlizes the prot)ability of the (teriw> 
tions and seelns to 1)refer derivations nlade up 
of larger substructures, it; may 1)e ild;eresting to 
237 
Disambiguator Parse Accuracy (/562) 
PCFG 
PMPG 
I'CFG@PMP(I 
(a) Correct Analysis 
S 
NP-SBJ VP , 
prp 
vbp NP 
NP PP PP 
dt m~ in NP to NP 
I I 
ltIl I) \]HI t) 
% F \] Parse Accuracy oi* parsable sentences (/456) % 
373 66.4 83.0 I 373 327 58.2 75.1 327 
402 71.5 85.2 402 
Table 1: Exl)erimental Results 
(b) PCFG-Analysis 
S 
NP-SBJ VP 
I 
prp 
vbp NP PP PP 
81.8 
71.7 
88.2 
ell; mt in NP to NP 
I I 
nilp nn t) 
Figure 2: PCFG Error AnMysis 
see if we can make this assumption explicit. 
3 Pattern-matching 
When we look at natural  parsing fl:om 
a memory-based point of view, one might say 
that a sentence is analyzed by looking u t) the 
most similar structure for the different analy- 
ses of that sentence in meinory. The parsing 
system described in this paper tries to mimic 
this 1)ehavior by interpreting the pop-model as 
a memory-t)ased model, in which analyses are 
being matched with syntactic patterns recorded 
in memory. Similarity t)etween the proposed 
analysis and tile patterns in memory is com- 
Imted according to: 
• the number of patterns needed to construct 
a tree (to be minimized) 
• the size of the patterns that are used to 
construct a tree (to be maximized) 
Tile nearest neighbor tbr a given analysis can 
be defined as the derivation that shares the 
largest amount of common nodes. 
4 The experimental Setup 
10-tbld cross-validation was used to appropri- 
ately evaluate the algorithms, as tile dataset 
(see Section 4.1) is rather small. Like DoPl the 
system is trained and tested on part-of-speech 
tag sequences. In a first phase, a simple bottom- 
up chart parser, trained on the training parti- 
tions, was used to generate parse forests tbr the 
1)art-of speech tag sequences of the test parti- 
tion. Next, the parse tbrests were sent to the 3 
algorithms (hencetbrth the disambiguators) to 
order these parse forests, the first parse of the 
ordered parse forest being the one proposed by 
the disanfl)iguator. 
In this paper, 3 disambiguators are described: 
• PCFG: siml)le Prol)abilistic Context-Free 
(~ralnlnar 
• PMPG: the DOP approximation, Patten> 
Matching Probabilistic Grammar 
• PCFGq-PMPG: a combined system, inte- 
grating PCFG and PMPG 
The evaluation metric used is pars(; accuracy, 
but also tile typical parser evaluation metric F- 
measure (precision/recall) is given ms a means 
of reference to other systems. 
4.1 The Corpus 
The ext)eriments were conducted oil all edited 
version of tile ATIS-II-corpus (Marcus, San- 
torini, and Marcinkiewicz, 1993), which con- 
sists of 578 sentences. Quite a lot of errors and 
inconsistencies were found, but not corrected, 
since we want our (probabilistic) system to be 
238 
lille to deal with this kind of noise. Seman- 
ti(:ally oriented tlags like -TMP all(1 -Dill,, lllOSI; 
often used in conjmml;ion with l'p, have been 
renlove(t~ since l;here is no way of rel;rieving this 
kind of semanti(: intbrmation from t;11(; t)art;-o5 
sl)ee(:h tags of the ATIS-(:ortms. Synta(:ti(: flags 
like -sILL on the other hand, \]lave 1)een main- 
taine(t. Internal relations (denoted by llllllleric 
tlags) were removed and tbr 1)ractical reasons, 
scntenee-lellgth was limited 1;o 15 words max. 
The edited (:orl)us retained 562 sentences. 
5 Parsing 
As a first phase, a 1)ottom-ut) (:hart parser 
i)al"sed t;he test sol;. This t)roved to t)e quite 
l)rol)lemati(:, since overall, 1()6 out of 562 sen- 
ten(:es (190/(0) could not 1)e t)arsed, (111(', to the 
sl,arsencss of the gramnmr, meanil,g I;ha(; l;he 
at)l)ropriate rewrite rule needed to (:onstru('l; the 
(:orre(:t t)~lrse tree tbr a senten(:c, in the test set, 
wasn't featured in the, in(tu(:ed grammar. NP- 
annol;at;ion seem(~(t 1;o 1)(; t;lle lml, in (:aus(~ \]'or 11n- 
l)arsal)ility. An NP like restriction code AP/57 
is repres(ml;ed 1)y the, rewrite rule: 
NP -~ NN NN sym sym sym C\]) CD 
Highly st)ccitt(: and tint stru(:tur(;s like these 
are s(:ar(:e an(t are usually ll()t induced from the 
training set whell nee(h;d to parse the test set. 
On-going re, sear(:h tries 1;o iml)h;ln(ml; gl"am- 
mal;i(:a.1 SlnOothing ;ts :t soluti(m to |;his 1)rol)hml, 
but one might also (:onsid('a: genera.ling parse 
fol"eSi;S with an in(tep(mdent ~,;l"allllll;Ll', ilMu(:e(l 
fronl the entire (:orlms (training setq-t('~si;s(',l;) or 
a difl'erent corlms. 111 t)()th cases, however, we 
would need to apply 1)robal)ilisti(" smoothing to 
be al)le to assign t)rot)at)ilities to llllkllown s(;,l;llc- 
lures/rules. Neither grammatical, nor t)rot)a- 
bilistic smoothing was imt)lemented in the (;ell- 
text of the exl)eriments, (les(:ril)ed in this 1)at)er. 
The sl/ars(mess of the grammar 1)roves t;o l)e 
a serious 1)otl;hme(:k fi)r pars(', a(:(:ura(:y, limiting 
our (lisamlliguators t;o a maximuln tlarsc act:u- 
racy of 81%. 
6 PCFd-experiments 
a PCFG constru(:ts parse trees by using simple 
rewrite-rules. The prot)al)ility of ~ parse tree 
(;~7tll })e (:omlml;ed l)y mull;it)lying the t)robat)ili- 
ties (1t" the. rewrite-rules that w(~.re used to (:on- 
st;fuel; the t)ars(:. Note that a l'CFd is i(h;nti(:al 
tO DOP\] whell we limit I;he maximum sul)Stl'UC- 
tures size to \], only Mlowing deriwd;ions of the 
type found at the right-hand side of Figure 1. 
6.1 Experimental Results 
The first line of Tat)le I shows the, rc, sull;s for the 
l'CF(~-(',xl)eriments: 66.4% parse accuracy is an 
adequate result for this baseline model. We also 
look at l)arsc accuracy for parsable sentences 
(an estimal;e of the parse accuracy we 1night 
get if we had a more suited parse forest gener- 
ator) and w(; notice that we are able to a(:hieve 
a 81.8% parse ae(:ur~my. This is already quite 
high, trot on exmnining the parsed data, serious 
and fluManmntal limitations to the POPO-mo(lcl 
can be el)served 
6.2 Error Analysis 
Figm'c 2, disl)lays the mosl; common tyl)c of mis- 
take mad(; l)y 1)CFG~S. :\]'lit; (;orr0,cl; t)arse l;ree 
('ouht r(;i)res(mt an mlalysis for 1;11(; senten(:e: 
I ".;ant o, fli.qht from \]h'us.scl.s to 2bronto. 
This examt)le shows thai; ~t PCFG h~ls a I;(~,n- 
dency to prctbr tlatter strueture, s over emt)edde, d 
stru(:t;ures. This is a trivial effect of 1;11(; mathc- 
mat;it'll tbrmula used to conqml;e the t)rol)at)il - 
il;y of a I)arse-tr(;(;: emt/cdded structure require 
more r(;writ(' rules, adding more fat:tots to the 
multii)li(:ation , whi(:h will alm(/st ilw, vit~d)ly r(;- 
suit in :t lower l)rol)al)ilit;y. 
11; is all 1111J'()ri;llllal;e 1)r()I)(;rl;y of I'CFG~s t;hal; 
the mmfl)er of no(l(;s in the 1)atse tree is invers(~ly 
1)rot)ortiomd;e to il;s t)rol)al)ility. ()n(; might t)e 
inclin(xl to n(n'malizc a parse tree's pr()bat)ility 
relative t(/the mnnt)er of nodes in the tree, but a 
more linguistically solmd alternative is at hand: 
the enhancenmnt of context sensii;ivity through 
the use of larger synl;tt(:ti(: (:ont(;xt; within t)arse 
tre(:s (:;/,11 make our disaml)iguat;or lnore rolmst. 
7 pMpo-experiments 
The 1)att(;rn-Matching Prol)al)ilistie Gramnmr 
is a memory-based interpretation of a \])OI'- 
model, in which a s(mtence is analyzed t)y 
matching the largest, possible chunks of syn- 
t;acti(" strut:lure Oll the sentence. To COml)ile 
t/~rse trees into pat, terns, all substructm'es ill 
the l;raining set are eneo(ted 1)y assigning l;hem 
specific indexes, NP(o)345 e.g. denotil~g a fully 
specified NP-sl;ruel;urc. This apt)roa(:h was in- 
sl)ired 1)y Goodman (199(i), in which Goodman 
239 
unsuccessflflly uses a system of indexed parse 
trees to transform DOP into aSl equivalent PCFG. 
The system of indexing (which is detailed in De 
Pauw (2000)) used in tim experiments described 
in this paper, is however specifically geared to- 
wards encoding contextual intbnnation in parse 
trees. 
Gives, an indexed training set, indexes can 
then be matched on a test set parse tree in a 
bottom-up fashion. In the tbllowing example, 
boxed nodes indicate nodes that have been re- 
trieved from memory. 
S 
vp 
prp vbp \[~ 
dt nn i 
I 
mlp 
In this example we can see that an NP, con- 
sisting of a flflly specified embedded NP and 
l'P, has l)een completely retrieved from men> 
ory, meaning that the NP in its entirety can 
be observed in the training set. However, no 
vp was tbund that consists of a VBP and that 
particular NP. Disambiguating with PMPG coil- 
sequently involves pruning all nodes retrieved 
frolu illeillory: 
S 
NP-SBJ VP 
vbp NP 
Finally, the probability for this pruned parse 
tree is computed in a pCFO-type manner, not 
adding the retrieved nodes to the product: 
P(parse) = P(s --+ NP-SBJ VP) . P(vp --+ vb I) NP) 
7.1 Experimental Results 
The results tbr the PMPG-exI)erinmnts can be 
ibund on the second line of Table 1. On some 
partitions, PMPG pcrtbrmed insignificantly bet- 
ter than PCFG, but Table 1 shows that tile re- 
sults for the context sensitive scheme are much 
worse. 58.2% overall parse accuracy and 71.7% 
parse accuracy on parsable sentences indicates 
that PMPG is *sot a valid approximation of DOP'S 
context-sensitivity. 
7.2 Error Analysis 
The dramatic drop in parsing accuracy calls tbr 
an error analysis of the parsed data. Figure 3 
is a prototypical mistake PMPG has made. The 
correct analysis could represent a parse tree for 
a sentence like: 
What flights can I get firm Brussels to 2brvnto. 
The PMPG analysis would never have been 
considered a likely candidate by a common 
PCFG. This particular sentence in fact was ef 
tbrtlessly disambignated by the PCFG . Yet 
the fact that large chunks of tree-structure are 
retrieved Dora memory, make it the preferred 
parse for the PMPG. We notice tbr instance that 
a large part of the sentence can be matched 
on an SBAR structure, which has no relevance 
whatsoever. 
Clearly, PMPG overestimates substructure 
size as a feature for disambiguation. It's inter- 
esting however to see that it is a working imple- 
mentation of context sensitivity, eagerly match- 
ing patterns from memory. At the same time, it 
has lost track of common-sense PCFG tactics, it 
is in the combination of the two that one may 
find a decent disambiguator and accurate im- 
plementation of context-sensitivity. 
8 A Combined System (PMPG@PCFG) 
Table 1 showed that 81.8(/o of the time, a PCFG 
finds the correct parse (Ibr t)arsable sentences), 
meaning that the correct parse is at the first 
place in the ordered parse tbrest. 99% of the 
time, the correct parse can be tbund among the 
10 most probable parses in the ordered pars(; 
forest. This opens up a myriad of possibili- 
ties tbr optin, ization. One might for instance 
use a best-first strategy to generate only the 10 
best parses, significantly reducing parse and dis- 
ambiguation time. An optimized disanNiguator 
might theretbre include a preparatory phase in 
wtfich a common-sense PCFG retains the most 
probable parses, so that a nlore sophisticated 
tbllow-up scheme ,teed not bother with sense- 
less analyses. 
In our experiments, we combined the 
common-sense logic of a PCFG and used its 
output as the PMPG'8 input. This is a well- 
established technique usually refi~rred to as sys- 
tent combination (see van Halteren, Zavrel, and 
Daelemans (1998) for an application of this 
240 
technique to I)art-ofst)ee('h tagging): 
I 
I)CFG 
I I, ,nos,: 
In'obable parso, s \[ 
I I'MI'(; 
I 
\]most In'obablc Imrse \[ 
We art'. also presented with th(', possibility to 
assign a weight to each algorithm's decision. 
The probability of a parse can the })e described 
with the following formula: 
I~/,(rewrito,-rule)i 
i 
l)(\]m'""s'(O = (# non-inde, xed nodes),, 
The weight of ea(:h algorithm's (lc(:ision, as 
well as the mnnt)er of 1HOSt t)robM)h; parses that 
m:e extrat)olated for the 1)attern-m~tt:hing algo- 
rithnq are parameters to 1)e optimized. Futm:e 
work will include evaluation on a validation set 
to retrieve the ol)timal va, hles for these 1)aram- 
e, tcrs. 
8.1 Results 
The third line in Tattle 1 shows that the com- 
1)ined system 1)ert'orlns better them either one, 
wit;h a parse accuracy of 71.5% and close I;o 90% 
1)~trs(; at:curacy on t)arsal)l(~ scnt(m(:es, whi(:h w(', 
(-nn consider an at)l)roximat;ion of results rc- 
porteA for DOP1. Error annlysis shows that 
the combined system is ilMe, ed M)Ie to overt:ore(; 
difficulties of both Mgorithms. The examtflo, 
in Figure 2 as well as the, ex~mlple in Figure 
3 were disanllfiguated correctly using the com- 
bined syst(;m 
9 Future Research 
Even thoug\]l t\]le PMPG shows a lot of promise 
in its parse at:curacy, the following extensions 
ne, ed to be researched: 
Optimizing PMPG@PCFG for comtmta- 
tional etfieieney: the graph in Section 8 
shows a possible optimized parsing system, 
in which a pre-processing POF(I generates 
the n most likely candidates to 1)e extrap- 
olated tbr the actual disantbiguator. Full 
parse forests were generated for the exper- 
iments descrit)e,d in this paper, so that the 
efiiciency gain of such a system Calmot t)e 
prot)erly estimated. 
PMPG@PCFG as all approximation needs to 
be compm'ed to actual D()P~ by having DOP 
parse the data used in this experiment, and 
by having PMPG-I-I'CFG parse the data used 
in the exl)erilnents described in Bod (1999). 
The l)ottlelmck of the sparse grammar 
1)roblem prevents us from flflly exploiting 
the disambiguating power of the pattern- 
matching algorithln. The ORAEL-system 
(GRammar Adaptation, Evolution and 
Learning) that is currently being devel- 
olmd , tries to address the t)roblem of gram- 
matical spars(mess by using evolutionary 
te(:lmiques to g('ncrate,, Ol)l;imizo, and com- 
l)lemeld, g~rallllllars. 
10 Conclusions 
Even though l)()l'\] exhil)its outstanding pars- 
ing 1)eh~vior, the et|iciency of the model is 
rathe, r problematic. The introduction of mul- 
tit fie deriwd;ions causes a considerable amount 
of computational overhead. Neither is it clear 
how the concept of multiple deriwd;ions trans- 
lal;es to a t)sycholinguistic context: there is no 
proof thai; lmlguage users consider (titf'(;rcnt in- 
st~mtiations of th(; same parse, whmt deciding 
on the correct anMysis for a given sentence. 
A 1)M;tcrn-m~t:chil~g schcnm w~s 1)rcsenLcd 
that tried to dis~mfl)iguate parse forests by 
trying to maximize the size of the sul)strnc- 
tures that can 1)e retrie, ved from inoanory. 
This straightforward memory-based intert)rcta- 
tion yields sut)-standm'd parsing accuracy. But 
the (:oml)ination of common-sense l)robal)ili- 
ties nnd enhanced context-sensitivity provides 
a workM)le t)arse forest disambiguator, indicat- 
ing that  users might exert a COml)lex 
corot)libation of memory-based recollection tech- 
niques and stored statistical data to analyze ut- 
terances. 

241 
(a) Correct Analysis 
S 
WHNP SQ 
WHNP PP PP 
wdt mm xxx xxx / 
(b) PMPG Analysis 
vbp NP-SBJ VP 
' 
prp 
vb NP PP PP 
xxx in NP to NP 
I I nnp mtp 
wdt fills 
S 
NP-SBJ \[~\] 
/ ~ nnp mlp 
xxx I NP- 
, 
XXX V 
I 
prp 
Figure 3: PMPG Error Analysis 

References 

Bod, R. 1995. Enriching linguistics with stat:istics: Per- 
fornmnce models of natural . Dissertation, 
II,LC, Univcrsiteit wm Alnsterdanl. 

Bod, l{.ens. 1999. Be, pond Grammar An E:rpericncc- 
Based ~lTu:ory of Language. Cambridge, Fngland: 
Cambridge University Press. 

Daelcmans, W., J. Zavrcl, K. Van der Sloot, and 
A. Van den Bosch. 19!)9. TiMt3L: Tillmrg Memory 
Based Learner, version 2.0, reference manual. Tech- 
nical Report ILK-9901, ILK, Tilburg University. 

Daelemans, Walter. 1999. Memory-based  pro- 
ccssing. Journal for Ez'perimcntal and Theoretical Ar- 
tificial Intelligence, 11:3:287 467. 

De Pauw, Guy. 2000. Probabilistischc Parsers - Con- 
te~:tgcvocligheid cn Pattcrn-Matehin 9. Antwerpen, 
Belgium: Antwerp Papers in Linguistics. 

Goodman, Joshua. 1996. Efficient algorithms for parsing 
the dop model. In Proceedings of the Conference on 
Empirical Methods in Natural Language Processing. 
pages 143-152. 

Marcus, M., B. Santorini, and Mary A. Marcinkiewicz. 
1993. Building a large annotated corpus of english: The Penn Treebank. Computational Linguistics, 19(2):313-330. 

Scha, R. 1990. Taaltheorie en taaltectmologie: com- 
petence cn performance. In Q. A. M. dt Kort 
and G. L. J. Lcerdam, editors, Computcrtocpassin- 
.qcn in dc Nccrlandistick, LVVN-jaarboek. Landelijke 
Vereniging van Ncerlandici. 

van Halteren, H., J. Zavrel, and W. Daelemans. 1998. 
Improving data-driven wordclass tagging by system 
combination. In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics, 
Montreal, Quebec, Canada, pages 491-497, Montreal, 
Canada, August 10-14. 
