Tagging accurately- Don't guess if you know 
Pasi Tapanainen 
Rank Xerox Research Centre 
Grenoble Laboratory 
6, chemin de Maupertuis 
38240 Meylan, France 
Pasi.Tapanainen@xerox.fr 
Atro Voutilainen 
Research Unit for Computational Linguistics 
University of Helsinki 
P.O. Box 4 
00014 University of Helsinki, Finland 
Atro.Voutilainen@helsinki.fi 
Abstract 
We discuss combining knowledge-based (or 
rule-based) and statistical part-of-speech 
taggers. We use two mature taggers, 
ENGCG and Xerox Tagger, to indepen- 
dently tag the same text and combine the 
results to produce a fully disambiguated 
text. In a 27000 word test sample taken 
from a previously unseen corpus we achieve 
98.5 % accuracy. This paper presents the 
data in detail. We describe the problems 
we encountered in the course of combining 
the two taggers and discuss the problem of 
evaluating taggers. 
1 Introduction 
This paper combines knowledge-based and statisti- 
cal methods for part-of-speech disambiguation, tak- 
ing advantage of the best features of both ap- 
proaches. The resulting output is fully and accu- 
rately disambiguated. 
We demonstrate a system that accurately resolves 
most part-of-speech ambiguities by means of syntac- 
tic rules and employs a stochastic tagger to elimi- 
nate the remaining ambiguity. The overall results 
are clearly superior to the reported results for state- 
of-the-art stochastic systems. 
The input to our part-of-speech disambiguator 
consists of lexically analysed sentences. Many words 
have more than one analysis. The task of the dis- 
ambiguator is to select the contextually appropriate 
alternative by discarding the improper ones. 
Some of the inappropriate alternatives can be dis- 
carded reliably by linguistic rules. For example, we 
can safely exclude a finite-verb reading if the previ- 
ous word is an unambiguous determiner. The appli- 
cation of such rules does not always result in a fully 
disambiguated output (e.g. adjective-noun ambigui- 
ties may be left pending) but the amount of ambigu- 
ity is reduced with next to no errors. Using a large 
collection of linguistic rules, a lot of ambiguity can 
be resolved, though some cases remain unresolved. 
47 
The rule system may also exploit the fact that cer- 
tain linguistically possible configurations have such 
a low frequency in certain types of text that they 
can be ignored. A rule that assumes that a prepo- 
sition is followed by a noun phrase may be a useful 
heuristic rule in a practical system, considering that 
dangling prepositions occur relatively infrequently. 
Such heuristic rules can be applied to resolve some of 
the ambiguities that survive the more reliable gram- 
mar rules. 
A stochastic disambiguator selects the most likely 
tag for a word by consulting the neighbouring tags 
or words, typically in a two or three word window. 
Because of the limited size of the window, the choices 
made by a stochastic disambiguator are often quite 
naive from the linguistic point of view. For instance, 
the correct resolution of a preposition vs. subordi- 
nating conjunction ambiguity in a small window is 
often impossible because both morphological cate- 
gories can have identical local contexts (for instance, 
both can be followed by a noun phrase). Some of the 
errors made by a stochastic system can be avoided 
in a knowledge-based system because the rules can 
refer to words and tags in the scope of the entire 
sentence. 
We use both types of disambiguators. The 
knowledge-based disambiguator does not resolve all 
ambiguities but the choices it makes are nearly al- 
ways correct. The statistical disambiguator resolves 
all ambiguities but its decisions are not very reliable. 
We combine these two disambiguators; here this 
means that the text is analysed with both systems. 
Whenever there is a conflict between the systems, 
we trust the analysis proposed by the knowledge- 
based system. Whenever the knowledge-based sys- 
tem leaves an ambiguity unresolved, we select that 
alternative which is closest to the selection made by 
the statistical system. 
The two systems we use are ENGCG (Karlsson 
et al., 1994) and the Xerox Tagger (Cutting et al., 
1992). We discuss problems caused by the fact that 
these taggers use different tag sets, and present the 
results obtained by applying the combined taggers 
to a previously unseen sample of text. 
2 The taggers in outline 
2.1 English Constraint Grammar Parser 
The English Constraint Grammar Parser, ENGCG 
(Voutilainen et al., 1992; Karlsson el al., 1994), is 
based on Constraint Grammar, a parsing framework 
proposed by Fred Karlsson (1990). It was developed 
1989-1993 at the Research Unit for Computational 
Linguistics, University of Helsinki, by Atro Voutilai- 
nen, Juha Heikkil~i and Arto Anttila; later on, Ti- 
mo J£rvinen has extended the syntactic description, 
and Pasi Tapanainen has made a new fast imple- 
mentation of the CG parsing program. ENGCG is 
primarily designed for the analysis of standard writ- 
ten English of the British and American varieties. 
In the development and testing of the system, over 
100 million words of running text have been used. 
The ENGTWOL lexicon is based on the two-level 
model (Koskenniemi, 1983). The lexicon contains 
over 80,000 lexical entries, each of which represents 
all inflected and central derived forms of the lexemes. 
The lexicon also employs a collection of tags for part 
of speech, inflection, derivation and even syntactic 
category (e.g. verb classification). 
Usually less than 5 % of all word-form tokens in 
running text are not recognised by the morphological 
analyser. Therefore the system employs a rule-based 
heuristic module that provides all unknown words 
with one or more readings. About 99.5 % of words 
not recognised by the ENGTWOL analyser itself get 
a correct analysis from the heuristic module. The 
module contains a list of prefixes and suffixes, and 
possible analyses for matching words. For instance, 
words beginning with un... and ending in ...al are 
marked as adjectives. 
The grammar for morphological disambiguation 
(Voutilainen, 1994) is based on 23 linguistic gen- 
eralisations about the form and function of essen- 
tially syntactic constructions, e.g. the form of the 
noun phrase, prepositional phrase, and finite verb 
chain. These generalisations are expressed as 1,100 
highly reliable 'grammar-based' and some 200 less 
reliable add-on 'heuristic' constraints, usually in a 
partial and negative fashion. Using the 1,100 best 
constraints results in a somewhat ambiguous out- 
put. Usually there are about 1.04-1.07 morpholog- 
ical analyses per word. Usually at least 997 words 
out of every thousand retain the contextually appro- 
priate morphological reading, i.e. the recall usually 
is at least 99.7 %. If the heuristic constraints are 
also used, the ambiguity rate falls to 1.02-1.04 read- 
ings per word, with an overall recall of about 99.5 %. 
This accuracy compares very favourably with results 
reported in (de Marcken, 1990; Weisehedel et al., 
1993; Kempe, 1994) - for instance, to reach the recall 
of 99.3 %, the system by (Weischedel et al., 1993) 
has to leave as many as three readings per word in 
its output. 
2.2 Xerox Tagger 
The Xerox Tagger 1, XT, (Cutting et al., 1992) is 
a statistical tagger made by Doug Cutting, Julian 
Kupiec, Jan Pedersen and Penelope Sibun in Xerox 
PARC. It was trained on the untagged Brown Cor- 
pus (Francis and Kubera, 1982). 
The lexicon is a word-list of 50,000 words with al- 
ternative tags. Unknown words are analysed accord- 
ing to their suffixes. The lexicon and suffix tables are 
implemented as tries. For instance, for the word live 
there are the following alternative analyses: JJ (ad- 
jective) and VB (uninflected verb). Unknown words 
not recognised by suffix tables get all tags from a 
specific set (called open-class). 
The tagger itself is based on the Hidden Markov 
Model (Baum, 1972) and word equivalence classes 
(Kupiec, 1989). Although the tagger is trained with 
the untagged Brown corpus, there are several ways 
to 'force' it to learn. 
• The symbol biases represent a kind of lexical 
probabilities for given word equivalence classes. 
• The transition biases can be used for saying that 
it is likely or unlikely that a tag is followed by 
some specific tag. The biases serve as default 
values for the Hidden Markov Model before the 
training. 
• Some rare readings may be removed from the 
lexicon to prevent the tagger from selecting 
them. 
• There are some training parameters, like the 
number of iterations (how many times the same 
block of text is used in training) and the size of 
the block of the text used for training. 
• The choice of the training corpus affects the re- 
sult. 
The tagger is reported (Cutting el al., 1992) to 
have a better than 96 % accuracy in the analysis of 
parts of the Brown Corpus. The accuracy is similar 
to other probabilistic taggers. 
3 Grammatical representations of 
the taggers 
A major difference between a knowledge-based and 
a probabilistic tagger is that the knowledge-based 
tagger needs as much information as possible while 
the probabilistic tagger requires some compact set 
of tags that does not make too many distinctions 
between similar words. The difference can be seen 
by comparing the Brown Corpus tag set (used by 
XT) with the ENGCG tag set. 
The ENGTWOL morphological analyser employs 
139 tags. Each word usually receives several tags 
(see Figure 1). There are also 'auxiliary' tags for 
derivational and syntactic information that do not 
1 We use version 1. 
48 
ENGCG 
has V PRES SG3 VEIN 
have V PRES -SG3 VEIN 
V INF 
V IMP VFIN 
V SUBJUNCTIVE VFIN 
was V PAST SG1,3 VEIN 
do V PRES -SG3 VEIN 
V INF 
V IMP VEIN 
V SUBJUNCTIVE VEIN 
done PCP2 
cook 
cool 
V PRES -SG3 VEIN 
V INF 
V IMP VEIN 
V SUBJUNCTIVE VEIN 
N NOM SG 
V PRES -SG3 VFIN 
V INF 
V IMP VFIN 
V SUBJUNCTIVE VEIN 
A ABS 
cooled PCP2 
V PAST VEIN 
cooling PCP1 
JXT\[ 
hvz 
hv 
bedz 
do 
vbn 
vb 
nn 
vb 
JJ 
nn 
rb 
vbn 
vbd 
vbg 
nn 
Figure 1: Some morphological ambiguities for verbs. 
increase morphological ambiguity but serve as addi- 
tional information for rules. If these auxiliary tags 
are ignored, the morphological analyser produces 
about 180 different tag combinations. 
The XT lexicon contains 94 tags for words; 15 of 
them are assigned unambiguously to only one word. 
There are 32 verb tags: 8 tags for have, 13 for be, 6 
for do and 5 tags for other verbs. ENGCG does not 
make a distinction in the tagset between words have, 
be, do and the other verbs. To see the difference with 
ENGCG, see Figure 1. 
The ENGCG description differs from the Brown 
Corpus tag set in the following respects. ENGCG is 
more distinctive in that a part of speech distinction 
is spelled out (see Figure 2) in the description of 
• determiner-pronoun homographs, 
• preposition-conjunction homographs, 
• determiner-adverb-pronoun homographs, and 
• uninflected verb forms (see Figure 1), which are 
represented as ambiguous due to the subjunc- 
tive, imperative, infinitive and present tense 
readings. 
On the other hand, ENGCG does not spell out part- 
of-speech ambiguity in the description of 
• -ing and nonfinite -ed forms, 
~ Two most probable ENGCG tags (%) 
cs cs (70 %) 
PREP (28 %) 
DT DET DEM SG (48 %) 
PRON DEM SG (27 %) 
DTI DET SG/PL (68 %) 
PRON SG/PL (28 %) 
IN PI~EP (99 %) 
ADV (0.5 %) 
JJ A ABS (93 %) 
N NOMSG (3 %) 
NN N NOM SG (88 %) 
N NOM SG/PL (7 %) 
NP N NOM SG (80 %) 
NNOMPL(7%) 
VB VINF (84 %) 
V PRES -SG3 VEIN (12 %) 
* NEG-PART (100 %) 
Figure 2: Some mappings from the Brown Corpus 
to the ENGCG tagset. 
• noun-adjective homographs when the core 
meanings of the adjective and noun readings are 
similar, 
• ambiguities due to proper nouns, common 
nouns and abbreviations. 
4 Combining the taggers 
In our approach we apply ENGCG and XT indepen- 
dently. Combining the taggers means aligning the 
outputs of the taggers and transforming the result 
of one tagger to that of the other. 
Aligning the output is straightforward: we only 
need to match the word forms in the output of the 
taggers. Some minor problems occur when tokeni- 
sation is done differently. For instance, XT handles 
words like aren't as a single token, when ENGCG 
divides it to two tokens, are and not. Also ENGCG 
recognises some multiple word phrases like in spite 
of as one token, while XT handles it as three tokens. 
We do not need to map both Brown tags to 
ENGCG and vice versa. It is enough to transform 
ENGCG tags to Brown tags and select the tag that 
XT has produced, or transform the tag of XT into 
ENGCG tags. We do the latter because the ENGCG 
tags contain more information. This is likely to be 
desirable in the design of potential applications. 
There are a couple of problems in mapping: 
• Difference in distinctiveness. Sometimes ENG- 
TWOL makes a distinction not made by the 
Brown tagset; sometimes the Brown tagset 
makes a distinction not made by ENGTWOL 
(see Figure 2). 
• Sometimes tags are used in a different way. A 
49 
case in point is the word as. In a sample of 76 
instances of as from the tagged Brown corpus, 
73 are analysed as CS; two as QL and one as 
IN, while in the ENGCG description the same 
instances of as were analysed 15 times as CS, 
four times as ADV, and 57 times as PREP. 
In ENGCG, the tag CS represents subordinat- 
ing conjunctions. In the following sentences 
the correct analysis for word as in ENGCG is 
PREP, not CS, which the Brown corpus sug- 
gests. 
The city purchasing department, the 
jury said, is lacking in experienced 
clerical personnel as(CS) a result of 
city personnel policies. -- The pe- 
tition listed the mayor's occupation 
as(CS) attorney and his age as(CS) 71. 
It listed his wife's age as(CS) 74 and 
place of birth as(CS) Opelika, Ala. 
The sentences are the three first sentences 
where word as appears in Brown corpus. In the 
Brown Corpus as appears over 7000 times and it 
is the fourteenth most common word. Because 
XT is trained according to the Brown Corpus, 
this is likely to cause problems. 
XT is applied independently to the text, and the 
tagger's prediction is consulted in the analysis of 
those words where ENGCG is unable to make a 
unique prediction. The system selects the ENGCG 
morphological reading that most closely corresponds 
to the tag proposed by XT. 
The mapping scheme is the following. For each 
Brown Corpus tag, there is a decision list for possible 
ENGCG tags, the most probable one first. We have 
computed the decision list from the part of Brown 
Corpus that is also manually tagged according to the 
ENGCG grammatical representation. The mapping 
can be used in two different ways. 
• Careful mode: An ambiguous reading in the 
output of ENGCG may be removed only when 
it is not in the decision list. In practise this 
leaves quite much ambiguity. 
• Unambiguous mode: Select the reading in the 
output of ENGCG that comes first in the deci- 
sion list 2. 
5 Performance test 
5.1 Test data 
The system was tested against 26,711 words of 
newspaper text from The Wall Street Journal, The 
Economist and Today, all taken from the 200-million 
word Bank of English corpus by the COBUILD team 
at the University of Birmingham, England (see also 
(J/irvinen, 1994)). None of these texts have been 
2In some cases a word may still remain ambiguous. 
used in the development of the system or the de- 
scription, i.e. no training effects are to be expected. 
5.2 Creation of benchmark corpus 
Before the test, a benchmark version of the test 
corpus was created. The texts were first analysed 
using the preprocessor, the morphological analyser, 
and the module for morphological heuristics. This 
ambiguous data was then manually disambiguated 
by judges, each having a thorough understanding of 
the ENGCG grammatical representation. The cor- 
pus was independently disambiguated by two judges. 
In the instructions to the experts, special empha- 
sis was given to the quality of the work (there was 
no time pressure). The two disambiguated versions 
of the corpus were compared using the Unix sdiff 
program. At this stage, slightly above 99 % of all 
analyses agreed. The differences were jointly exam- 
ined by the judges to see whether they were caused 
by inattention or by a genuine difference of opinion 
that could not be resolved by consulting the docu- 
mentation that outlines the principles adopted for 
this grammatical representation (for the most part 
documented in (Karlsson et al., 1994)). It turned 
out that almost all of these differences were due to 
inattention. Only in the analysis of a few words it 
was agreed that a multiple choice was appropriate 
because of different meaning-level interpretations of 
the utterance (these were actually headings where 
some of the grammatical information was omitted). 
Overall, these results agree with our previous expe- 
riences (Karlsson et al., 1994): if the analysis is done 
by experts in the adopted grammatical representa- 
tion, with emphasis on the quality of the work, a 
consensus of virtually 100 % is possible, at least at 
the level of morphological analysis (for a less opti- 
mistic view, see (Church, 1992)). 
5.3 Morphological analysis 
The preprocessed text was submitted to the ENG- 
TWOL morphological analyser, which assigns to 
25,831 words of the total 26,711 (96.7 %) at least 
one morphological analysis. The remaining 880 
word-form tokens were analysed with the rule-based 
heuristic module. After the combined effect of these 
modules, there were 47,269 morphological analyses, 
i.e. 1.77 morphological analyses for each word on 
an average. At this stage, 23 words missed a con- 
textually appropriate analysis, i.e. the error rate of 
the system after morphological analysis was about 
0.1%. 
5.4 Morphological disambiguation 
The morphologically analysed text was submitted 
to five disambiguators (see Figure 3). The first one, 
D1, is the grammar-based ENGCG disambiguator. 
In the next step (D2) we have used also heuristic 
ENGCG constraints. The probabilistic information 
50 
is used in D3, where the ambiguities of D2 are re- 
solved by XT. We also tested the usefulness of the 
heuristic component of ENGCG by omitting it in 
D4. The last test, D5, is XT alone, i.e. only proba- 
bilistic techniques are used here for resolving ENG- 
TWOL ambiguities. 
The ENGCG disambiguator performed somewhat 
less well than usually. With heuristic constraints, 
the error rate was as high as 0.63 %, with 1.04 mor- 
phological readings per word on an average. How- 
ever, most (57 %) of the total errors were made af- 
ter ENGCG analysis (i.e. in the analysis of no more 
than 3.6 % of all words). In a way, this is not very 
surprising because ENGCG is supposed to tackle all 
the 'easy' cases and leave the structurally hardest 
cases pending. But it is quite revealing that as much 
as three fourths of the probabilistic tagger's errors 
occur in the analysis of the structurally 'easy' cases; 
obviously, many of the probabilistic system's deci- 
sions are structurally somewhat naive. Overall, the 
hybrid (D3#) reached an accuracy of about 98.5 % - 
significantly better than the 95-97 % accuracy which 
state-of-the-art probabilistic taggers reach alone. 
The hybrid D3~ is like hybrid D3~, but we have 
used careful mapping. There some problematic 
ambiguity (see Figure 2) is left pending. For in- 
stance, ambiguities between preposition and infini- 
tive marker (word to), or between subordinator and 
preposition (word as), are resolved as far as ENGCG 
disambiguates them, the prediction of XT is not con- 
sulted. Also, when XT proposes tags like JJ (adjec- 
tive), AP (post-determiner) or VB (verb base-form) 
very little further disambiguation is done. This hy- 
brid does not contain any mapping errors, and on 
the other hand, not all the XT errors either. 
The test without the heuristic component of 
ENGCG (D4) suggests that ambiguity should be re- 
solved as far as possible with rules. An open ques- 
tion is, how far we can go using only linguistic infor- 
mation (e.g. by writing more heuristic constraints to 
be applied after the more reliable ones, in this way 
avoiding many linguistically naive errors). 
The last test gives further evidence for the use- 
fulness of a carefully designed linguistic rule compo- 
nent. Without such a rule component, the decrease 
in accuracy is quite dramatic although a part of the 
errors come from the mapping between tag sets 3. 
6 Conclusion 
In this paper we have demonstrated how knowledge- 
based and statistical techniques can be combined to 
improve the accuracy of a part of speech tagger. Our 
system reaches a better than 98 % accuracy using a 
relatively fine-grained grammatical representation. 
Some concluding remarks are in order. 
3Even without the mapping errors, the reported 4 % 
error rate of XT is considerably higher than that of our 
hybrid. 
• Using linguistic information before a statistical 
module provides a better result than using a 
statistical module alone. 
• ENGCG leaves some 'hard' ambiguities unre- 
solved (about 3-7 % of all words). This amount 
is characteristic of the ENGCG rule-formMism, 
tagset and disambiguation grammar. It does 
not necessarily hold for other knowledge-based 
systems. 
• Only about 20-25 % of errors made by the sta- 
tistical component occur in the analysis of these 
'hard' ambiguities. That means, 75-80 % of the 
errors made by the statistical tagger were re- 
solved correctly using linguistic rules. 
• Certain kinds of ambiguity left pending by 
ENGCG, e.g. CS vs. PREP, are resolved rather 
unreliably by XT. 
• The overall result is better than other state-of- 
the-art part-of-speech disambiguators. In our 
27000 word test sample from previously unseen 
corpus, 98.5 % of words received a correct anal- 
ysis. In other words, the error rate is reduced 
at least by half. 
Although the result is better than provided by 
any other tagger that produces fully disambiguated 
output, we believe that the result could still be im- 
proved. Some possibilities: 
• We could use partly disambiguated text 
(e.g. the output of parsers D1, D2 or D3~) 
and disambiguate the result using a knowledge- 
based syntactic parser (see experiments in (Vou- 
tilainen and Tapanainen, 1993)). 
• We could leave the text partly disambiguated, 
and use a syntactic parser that uses both lin- 
guistic knowledge and corpus-based heuristics 
(see (Tapanainen and J//rvinen, 1994)). 
• Some ambiguities are very difficult to resolve in 
a small window that statistical taggers currently 
use (e.g. CS vs. PREP ambiguity when a noun 
phrase follows). A better way to resolve them 
would probably be to write (heuristic) rules. 
• We could train the statistical tagger on the out- 
put of a knowledge-based tagger. That is prob- 
lematic because generally statistical methods 
seem to require some compact set of tags, while 
a knowledge-based system needs more informa- 
tive tags. The tag set of a knowledge-based sys- 
tem should be reduced down to some subset. 
That might prevent some mapping errors but 
there is no quarantee that the statistical tagger 
would work any better. 
• We could try the components in a different or- 
der: using statistics before heuristical knowl- 
edge etc. However, currently the heuristic com- 
ponent makes less errors than the statistical 
tagger. 
51 
DO (Morphological analysis) 
D1 (DO + ENGCG) 
D2 (D1 + ENGCG heuristics) 
D3~ (D2 + XT + C-mapping) 
D3Z (D2 + XT + mapping) 
D4 (D1 + XT + mapping) 
D5 (DO + XT + mapping) 
\[ Amb. words 
37.6 % 
6.4 % 
3.6 % 
2.2 % 
0.0 % 
0.0 % 
0.7 % 
Figure 3: Performance of the 
Readings Readings / 
47269 1.77 
28815 1.08 
27681 1.04 
27358 1.02 
26744 1.00 
26794 1.00 
26977 1.01 
word\]Errors \[Error rate (%) 
23 0.09 % 
94 0.35 % 
169 0.63 % 
220 0.82 % 
391 1.46 % 
6.38 % 
taggers on a 26,711-word corpus. 
• We could use a better statistical tagger. But 
the accuracy of XT is almost the same as the 
accuracy of any other statistical tagger. What 
is more, the accuracy of the purely statistical 
taggers has not been greatly increased since the 
first of its kind, CLAWS1, (Marshall, 1983) was 
published over ten years ago. 
We believe that the best way to boost the accu- 
racy of a tagger is to employ even more linguistic 
knowledge. The knowledge should, in addition, con- 
tain more syntactic information so that we could re- 
fer to real (syntactic) objects of the language, not 
just a sequence of words or parts of speech. Statisti- 
cal information should be used only when one does 
not know how to resolve the remaing ambiguity, and 
there is a definite need to get fully unambiguous out- 
put. 
7 Acknowledgements 
We would like to thank Timo J£rvinen, Lauri Kart- 
tunen, Jussi Piitulainen and anonymous referees for 
useful comments on earlier versions of this paper. 

References 

L. E. Baum. 1972. An inequality and associated 
maximization technique in statistical estimation 
for probabilistic functions of a Markov process. 
Inequatics, 3:1-8, 1972. 

Kenneth W. Church. 1992. Current Practice in Part 
of Speech Tagging and Suggestions for the Fu- 
ture. In Simmons (ed.), Sbornik praci: In Honor 
of Henry Ku~era. Michigan Slavic Studies. 

Doug Cutting, Julian Kupiec, Jan Pedersen and 
Penelope Sibun. 1992. A Practical Part-of-Speech 
Tagger. In Proceedings of ANLP 92. 

W. N. Francis and F. KuSera. 1982. Frequency Anal- 
ysis of English Usage. Houghton Mifflin. 
Timo Jarvinen. 1994. Annotating 200 million 
words: the Bank of English project. In proceedings of COLING-94, Vol. 1,565-568. Kyoto. 

Fred Karlsson. 1990. Constraint Grammar as a 
Framework for Parsing Running Text. In Proceedings of COLING-90. Helsinki. Vol. 3, 168-173. 
Fred Karlsson, Atro Voutilainen, Juha Heikkile and 
Arto Anttila (eds.). 1994. Constraint Grammar: 
a Language-Independent System for Parsing Un- 
restricted Text. Berlin: Mouton de Gruyter. 

Andre Kempe. 1994. A Probabilistic Tagger and 
an Analysis of Tagging Errors. Research Report, 
Institut fiir Maschinelle Sprachverarbeitung, Uni- 
versit~t Stuttgart. 

Kimmo Koskenniemi. 1983. Two-level Morphology. 
A General Computational Model for Word-form 
Production and Generation. Publication No. 11, 
Department of General Linguistics, University of 
Helsinki. 

Julian M. Kupiec. 1989. Probabilistic models of 
short and long distance word dependencies in running text. In Proceedings of the 1989 DARPA 
Speech and Natural Language Workshop pp. 290-295. Philadelphia. Morgan Kaufman. 

Carl de Marcken. 1990. Parsing the LOB Corpus. 
In Proceedings of the 28th Annual Meeting of the 
ACL. 243-251. 

Ian Marshall. 1983. Choice of grammatical word- 
class without global syntactic analysis: tagging 
words in the LOB Corpus. Computers in the Hu- 
manities 17. 139-150. 

Pasi Tapanainen and Timo Jarvinen. 1994. Syntactic analysis of natural language using linguistic rules and corpus-based patterns. In proceedings of 
COLING-94, Vol. 1,629-634. Kyoto. 

Atro Voutilainen. 1994. Morphological disambiguation. In Karlsson et al.. 

Atro Voutilainen, Juha Heikkil£ and Arto Anttila. 
1992. Constraint Grammar of English. A Per- 
formance- Oriented Introduction. Publication No. 
21, Department of General Linguistics, University 
of Helsinki. 

Atro Voutilainen and Pasi Tapanainen. 1993. Ambiguity resolution in a reductionistic parser. Proceedings of EACL'93. Utrecht, Holland. 394-403. 

Ralph Weischedel, Marie Meteer, Richard Schwartz, 
Lance Ramshaw and Jeff Palmuzzi. 1993. Coping with ambiguity and unknown words through 
probabilistic models. Computational Linguistics, 
Vol. 19, Number 2. 
