i 
Combining Hand-crafted Rules and Unsupervised Learning 
in 
Constraint-based Morphological Disambiguation 
Kemal Oflazer and GSkhan Tfir 
Department of Computer Engineering and Information Science 
Bilkent University, Bilkent, Ankara, TR-06533, TURKEY 
{ko, tur}~cs, bilkent, edu. tr 
Abstract 1 Introduction 
This paper presents a constraint-based 
morphological disambiguation approach 
that is applicable languages with complex 
morphology-specifically agglutinative lan- 
guages with productive inflectional and 
derivational morphological phenomena. In 
certain respects, our approach has been 
motivated by Brill's recent work (Brill, 
1995b), but with the observation that his 
transformational approach is not directly 
applicable to languages like Turkish. Our 
system combines corpus independent hand- 
crafted constraint rules, constraint rules 
that are learned via unsupervised learn- 
ing from a training corpus, and additional 
statistical information from the corpus to 
be morphologically disambiguated. The 
hand-crafted rules are linguistically moti- 
vated and tuned to improve precision with- 
out sacrificing recall. The unsupervised 
learning process produces two sets of rules: 
(i) choose rules which choose morpholog- 
ical parses of a lexical item satisfying con- 
straint effectively discarding other parses, 
and (ii) delete rules, which delete parses 
satisfying a constraint. Our approach also 
uses a novel approach to unknown word 
processing by employing a secondary mor- 
phological processor which recovers any rel- 
evant inflectional and derivational informa- 
tion from a lexieal item whose root is un- 
known. With this approach, well below 
1% of the tokens remains as unknown in 
the texts we have experimented with. Our 
results indicate that by combining these 
hand-crafted, statistical and learned infor- 
mation sources, we can attain a recall of 96 
to 97% with a corresponding precision of 
93 to 94%, and ambiguity of 1.02 to 1.03 
parses per token. 
Automatic morphological disambiguation is a very 
crucial component in higher level analysis of natural 
language text corpora. Morphological disambigua- 
tion facilitates parsing, essentially by performing a 
certain amount of ambiguity resolution using rela- 
tively cheaper methods (e.g., Gfing6rdii and Oflazer 
(1995)). There has been a large number of studies 
in tagging and morphological disambiguation using 
various techniques. Part-of-speech tagging systems 
have used either a statistical approach where a large 
corpora has been used to train a probabilistic model 
which then has been used to tag new text, assign- 
ing the most likely tag for a given word in a given 
context (e.g., Church (1988), Cutting et al. (1992), 
DeRose (1988)). Another approach is the rule-based 
or constraint-based approach, recently most promi- 
nently exemplified by the Constraint Grammar work 
(Karlsson et al., 1995; Voutilainen, 1995b; Vouti- 
lainen et al., 1992; Voutilainen and Tapanainen, 
1993), where a large number of hand-crafted linguis- 
tic constraints are used to eliminate impossible tags 
or morphological parses for a given word in a given 
context. Brill (1992; 1994; 1995a) has presented a 
transformation-based learning approach, which in- 
duces rules from tagged corpora. Recently he has 
extended this work so that learning can proceed 
in an unsupervised manner using an untagged cor- 
pus (Brill, 1995b). Levinger et al. (1995) have re- 
cently reported on an approach that learns morpho- 
lexical probabilities from untagged corpus and have 
the used the resulting information in morphological 
disambiguation in Hebrew. 
In contrast to languages like English, for which 
there is a very small number of possible word forms 
with a given root word, and a small number of tags 
associated with a given lexical form, languages like 
Turkish or Finnish with very productive agglutina- 
tive morphology where it is possible to produce thou- 
sands of forms (or even millions (Hankamer, 1989)) 
for a given root word, pose a challenging problem 
for morphological disambiguation. In English, for 
example, a word such as make or set can be verb 
69 
or a noun. In Turkish, even though there are ambi- 
guities of such sort, the agglutinative nature of the 
language usually helps resolution of such ambiguities 
due to restrictions on lnorphotactics. On the other 
hand, this very nature introduces another kind of 
ambiguity, where a lexical form can be morpholog- 
ically interpreted in many ways, some with totally 
unrelated roots and morphological features, as will 
be exemplified in the next section. 
Our previous approach to tagging and morpho- 
logical disambiguation for Turkish text had em- 
ployed a constraint-based approach (Oflazer and 
Kuru6z, 1994) along the general lines of similar pre- 
vious work for English (Karlsson et al., 1995; Vouti- 
lainen et al., 1992; Voutilainen and Tapanainen, 
1993). Although the results obtained there were rea- 
sonable, the fact that all constraint rules were hand 
crafted, posed a rather serious impediment to the 
generality and improvement of the system. 
In this paper we present a constraint-based mor- 
phological disambiguation approach that uses unsu- 
pervised learning component to discover some of the 
constraints it uses in conjunction with hand-crafted 
rules. It is specifically applicable to languages with 
productive inflectional and derivational morpholog- 
ical processes, such as Turkish, where morpholog- 
ical ambiguity has a rather different nature than 
that found in languages like English. Our approach 
starts with a set of corpus-independent hand-crafted 
rules that reduce morphological ambiguity (hence 
improve precision) without sacrificing recall. It then 
uses an untagged training corpus in which all lexical 
items have been annotated with all possible morpho- 
logical analyses, incrementally proposing and eval- 
uating additional (possibly corpus dependent) con- 
straints for disambiguation of morphological parses 
using the constraints imposed by unambiguous con- 
texts. These rules choose or delete parses with spec- 
ified features. In certain respects, our approach has 
been motivated by Brill's recent work (Brill, 1995b), 
but with the observation that his transformational 
approach is not directly applicable to languages like 
Turkish, where tags associated with forms are not 
predictable in advance. 
In the following sections, we present an overview 
of the morphological disambiguation problem, high- 
lighted with examples from Turkish. We then 
present the details of our approach and results. We 
finally conclude after a discussion and evaluation of 
our results. 
2 Tagging and Morphological 
Disambiguation 
In almost all languages, words are usually ambigu- 
ous in their parts-of-speech or other lexical features, 
and may represent lexical items of different syntac- 
tic categories, or morphological structures depend- 
ing on the syntactic and semantic context. Part-of- 
speech (POS) tagging involves assigning every word 
its proper part-of-speech based upon the context the 
word appears in. In English, for example a word 
such as set can be a verb in certain contexts (e.g., 
He set the table for dinner) and a noun in some oth- 
ers (e.g., We are now facing a whole set of problems). 
In Turkish, there are ambiguities of the sort 
above. However, the agglutinative nature of the 
language usually helps resolution of such ambigui- 
ties due to the restrictions on morphotactics. On 
the other hand, this very nature introduces another 
kind of ambiguity, where a whole lexical form can be 
morphologically interpreted in many ways not pre- 
dictable in advance. For instance, our full-scale mor- 
phological analyzer for Turkish returns the following 
set of parses for the word oysa: 1,2 
1. \[\[CAT CONN\] \[ROOT oysa\]\] 
(on the other hand) 
2. \[\[CAT NOUN\] \[ROOT oy\] \[AGR 3SG\] 
\[POSS NONE\] \[CASE NOM\] 
\[c0Nv VERB NONE\] 
\[TAM1COND\] \[AGR 3SG\]\] 
(if it is a vote) 
3. \[\[CAT PRONOUN\] \[ROOT o\] \[TYPE DEMONS\] 
\[AGR 3SG\] \[POSS NONE\] 
\[CASE NOM\] \[CONV VERB NONE\] 
\[TAM1COND\]\[AGR 3SG\]\] 
(if it is) 
4. \[\[CAT PRONOUN\] \[ROOT o\] \[TYPE PERSONAL\] 
\[AGR 3SG\] \[POSS NONE\] \[CASE NOM\] 
\[CONV VERB NONE\] \[TAM1COND\] 
\[AGR 3SG\]\] 
(if s/he is) 
s. \[\[CAT VERB\] \[ROOT oy\] \[SENSE POS\] 
\[TAM1 DES\] \[AGR 3SG\]\] 
(wish s/he would carve) 
On the other hand, the form oya gives rise to the 
following parses: 
1. \[\[CAT NOUN\] \[ROOT oya\] \[AGR 3SG\] 
\[POSS NONE\] \[CASE NOM\]\] (lace) 
2. \[\[CAT NOUN\] \[ROOT oy\] \[AGR 3SG\] 
\[POSS NONE\] \[CASE DAT\]\] (to the vote) 
3. \[\[CAT VERB\] \[ROOT oy\] \[SENSE POS\] 
\[TAM1 OPT\] \[AGR 3SG\]\] (let him carve) 
and the form oyun gives rise to the following parses: 
iOutput of the morphological anaJyzer is edited for 
clarity, and English glosses have been given. 
2Glosses are given as linear feature value sequences 
corresponding to the morphemes (which are not shown). 
The feature names are as follows: CAT-major category, 
TYPE-minor category, R00T-main root form, AGR -number 
and person agreement, POSS- possessive agreement, CASE 
surface case, CONV- conversion to the category follow- 
ing with a certain suffix indicated by the argument after 
that, TAMl-tense, aspect, mood marker 1, SENSE-verbal 
polarity, DES- desire mood, IMP-imperative mood, 0PT- 
optative mood, COND-Conditional 
70 
I 
I. \[\[CAT NOUN\] \[ROOT oyun\] \[AGR 3SG\] 
\[POSS NONE\] \[CASE NOM\]\] (game) 
2. \[\[CAT NOUN\] \[ROOT oy\] \[AGR 3SG\] 
\[POSS NONE\] \[CASE GEN\]\] (of the vote) 
3. \[\[CAT NOUN\] \[ROOT oy\] \[AGR 3SG\] 
\[POSS 2SG\] \[CASE NOM\]\] (your vote) 
4. \[\[CAT VERB\] \[ROOT oy\] \[SENSE POS\] 
\[TAM1 IMP\] \[AGR 2PL\]\] (carve it!) 
On the other hand, the local syntactic context 
may help reduce some of the ambiguity above, as 
in: 3 
sen-in oy-un .. 
PRON(you)+GEN NOUN(vote)+POSS-2SG 
your vote 
oy-un reng-i .. 
NOUN(vote)+GEN NOUN(color)+POSS-3SG 
(NOUN-GEN NOUN-POSS form) 
color of the vote 
oyun reng-i .. 
NOUN(game) NOUN(color)+POSS-3SG 
game color 
(NOUN NOUN-POSS form) 
using some very basic noun phrase agreement con- 
straints in Turkish. Obviously in other similar cases, 
it may be possible to resolve the ambiguity com- 
pletely. 
There are also numerous other examples of word 
forms where productive derivational processes come 
into play: 4 
geldiGimdeki (at the time I came) 
\[\[CAT VERB\] \[ROOT gel\] \[SENSE POS\] 
(basic form) 
\[CONV NOUN DIE\] \[AGR 3SG\] 
\[POSS iSG\] \[CASE LOC\] 
(participle form) 
\[CONV ADJ REL\]\] 
(final adjectivalization by the 
relative (ki) suffix) 
Here, the original root is verbal but the final part- 
of-speech is adjectival. In general, the ambiguities of 
the forms that come before such a form in text can be 
resolved with respect to its original (or intermediate) 
parts-of-speech (and inflectional features), while the 
ambiguities of the forms that follow can be resolved 
based on its final part-of-speech. 
The main intent of our system is to achieve a mor- 
phological ambiguity reduction in the text by choos- 
ing for a given ambiguous token, a subset of its 
ZWith a slightly different but nevertheless common 
glossing convention. 
4Upper cases in morphological output indicates one of 
the non-ASCII special Turkish characters: e.g., G denotes 
~, U denotes /i, etc. 
parses which are not disallowed by the syntactic con- 
text it appears in. It is certainly possible that a given 
token may have multiple correct parses, usually with 
the same inflectional features or with inflectional fea- 
tures not ruled out by the syntactic context. These 
can only be disambiguated usually on semantic or 
discourse constraint grounds. 5 
We consider a token fully disambiguated if it has 
only one morphological parse remaining after auto- 
matic disambiguation. We consider as token as cor- 
rectly disambiguated, if one of the parses remain- 
ing for that token is the correct intended parse. 6 
We evaluate the resulting disambiguated text by a 
number of metrics defined as follows (Voutilainen, 
1995a): 
#Parses Ambiguity - 
#Tokens 
Recall = #Tokens Correctly Disambiguatcd 
#Tokens 
~Tokcns Correctly Disambiguated Precision = 
~Parses 
In the ideal case where each token is uniquely and 
correctly disambiguated with the correct parse, both 
recall and precision will be 1.0. On the other hand, a 
text where each token is annotated with all possible 
parses, 7 the recall will be 1.0 but the precision will 
be low. The goal is to have both recall and precision 
as high as possible. 
3 Constraint-based Morphological 
Disambiguation 
This section outlines our approach to constraint- 
based morphological disambiguation incorporating 
unsupervised learning component. Our system with 
the structure presented in Figure 1 has three main 
components: 
1. the preprocessor, 
2. the learning module, and 
3. the morphological disambiguation module. 
Preprocessing is common to both the learning and 
the morphological disambiguation modules. The 
module takes as input to the system raw Turkish 
text and preprocesses it in a manner to be described 
shortly. 
If the text is to be used for training, the learning 
module then 
1. applies an initial set of linguistically motivated 
hand-crafted constraint rules to choose and/or 
delete certain parses, and 
5For instance the third and fourth parses for oysa 
above. 
6It is certainly possible that, a parse that is deleted 
may also be a valid parse in that context. 
rAssuming no unknown words. 
71 
2. uses an unsupervised learning procedure to in- 
duce some additional (an possibly corpus de- 
pendent) rules to choose and delete some parses. 
Morphological disambiguation of previously un- 
seen text proceeds as follows: 
1. The hand-crafted rules are applied first. 
2. Certain parses are deleted using context statis- 
tics on the corpus to be tagged. 
3. Rules learned to choose and delete parses are 
then applied. 
3.1 The Preprocessor 
The preprocessing module takes as input a Turk- 
ish text, segments it into sentences using various 
heuristics about punctuation, tokenizes and runs it 
through a wide-coverage high-performance morpho- 
logical analyzer developed using two-level morphol- 
ogy tools by Xerox (Karttunen, 1993). This module 
also performs a number of additional functions: 
• it groups lexicalized collocations such as id- 
iomatic forms, semantically coalesced forms 
such as proper noun groups, certain numeric 
forms, etc. 
• it groups any compound verb formations which 
are formed by a lexically adjacent, direct or 
oblique object, and a verb, which for the pur- 
poses of syntactic analysis, may be considered 
as single lexical item: e.g., saygz durmak (to pay 
respect), kafay~ yemek (literally t0 eat the head 
- to get mentally deranged), etc. 
• it groups non-lexicalized collocations: Turkish 
abounds with various non-lexicalized colloca- 
tions where the sentential role of the colloca- 
tion has (almost) nothing to do with the parts- 
of-speech of the individual forms involved. Al- 
most all of these collocations involve duplica- 
tions, and have forms like w + x w + y where 
w is the duplicated string comprising the root 
and certain sequence of suffixes and x and y are 
possibly different (or empty) sequences of other 
suffixes. 
The following is a list of multi-word constructs 
for Turkish that we handle in our preproces- 
sor. This list is not meant to be comprehensive, 
and new construct specifications can easily be 
added. It is conceivable that such a function- 
ality can be used in almost any language. (See 
Oflazer and Kuru6z (1994) and KuruSz (1994) 
for details of all other forms for Turkish.) 
1. duplicated optative and 3SG verbal forms 
functioning as manner adverb. An example 
is ko~a ko~a, where each lexical item has the 
morphological parse 
\[\[CAT VERB\] \[ROOT koS\] \[SENSE POS\] 
\[TAM1 0PT\] \[AGR3SG\]\] 
The preprocessor recognizes this and gen- 
erates the feature sequence: 
2. 
\[\[CAT VERB\] \[ROOT koS\] \[SENSE POS\] 
\[TAM1 OPT\] \[AGR 3SG\] 
\[CONV ADVERB DUPi\] \[TYPE MANNER\]\] 
aorist verbal forms with root duplications 
and sense negation, functioning as tem- 
poral adverbs. For instance for the non- 
lexicalized collocation yapar yapmaz, where 
items have the parses 
\[\[CAT VERB\] \[ROOT yap\] \[SENSE P0S\] 
\[TAM1 AORIST \] \[AGR 3SG\]\] 
\[\[CAT VERB\] \[ROOT yap\] \[SENSE BEG\] 
\[TAM1 AORIST \] \[AGR 3SG\]\] 
respectively, the preprocessor generates the 
feature sequence 
\[\[CAT VERB\] \[ROOT koS\] \[SENSE POS\] 
\[TAM1 AORIST\] \[AGR 3SG\] 
\[CONV ADVERB DUP-AOR\] \[TYPE TEMP\]\] 
3. duplicated verbal and derived adverbial 
forms with the same verbal root acting as 
temporal adverbs, e.g., gitti gideli, 
4. emphatic adjectival forms involving dupli- 
cation and question clitic, e.g., g71zel mi 
g~zel (beautiful question-clitic beautiful- 
very beautiful) 
5. adjective or noun duplications that act as 
manner adverbs, e.g., hzzh hzzh, evev, 
This module recognizes all such forms and coa- 
lesces them into new feature structures reflect- 
ing the final structure along with any inflec- 
tional information. 
• The preprocessor then converts each parse into 
a hierarchical feature structure so that the in- 
flectional feature of the form with the last cat- 
egory conversion (if any) are at the top level. 
Thus in the example above for geldi~imdeki, the 
following feature structure is generated: 
\[\[CAT VERB\] \[ROOT gel\] \[SENSE POS\] 
\[CONV NOUN DIK\] \[AGR 3SG\] 
\[P0SS ISG\] \[CASE LOC\] 
\[CONV ADJ REL\]\] 
"CAT ADJ 
"CAT 
AGR 
POSS 
CASE 
NOUN 
3SG 
iSG 
LOC 
STEM CAT VERB' 
STEM \[ROOT gel 
\[.SENSE POS 
SUFFIX DIK 
SUFFIX ~EL 
• Finally, each such feature structure is then pro- 
jected on a subset of its features. The features 
selected are 
- inflectional and certain derivational mark- 
ers, and stems for open class of words, 
72 
TOKEN IZATION MORPHOLOGY NON-LEXICAL UNKNOWN FORMAT 
COLLOCATION WORD CONVERSION 
RECOGNIZER PROCESSOR ( / PRO/ECTION ) 
PREPROCESSOR 
MORPHOLOGICAL 
DISAM BIGUATION 
MODULE 
LEARNING LEARNED RULES 
MOI)ULE 
Figure h The structure of the constraint-based morphological disambiguation system. 
-- roots and certain relevant features such as 
subcategorization requirements for closed 
classes of words such as connectives, post- 
positions, etc. 
The set of features selected for each part-of- 
speech category is determined by a template 
and hence is controllable, permitting experi- 
mentation with differing levels of information. 
The information selected for stems are deter- 
mined by the category of the stem itself recur- 
sively. 
Under certain circumstances where a token has 
two or more parses that agree in the selected 
features, those parses will be represented by 
a single projected parse, hence the number of 
parses in the (projected) training corpus may be 
smaller than the number of parses in the origi- 
nal corpus. For example, the feature structure 
above is projected into a feature structure such 
as: 
-CAT ADJ 
\[OAT NOUN \]\] 
|AGR 3SG 
|POSS 1SG 
STEM /CASE LOC 
/STEM \[CAT VERB 
\[SUFFIX DIK 
SUFFIX REL 
3.2 Unknown Words 
Although the coverage of our morphological analyzer 
for Turkish (Oflazer, 1993), with about 30,000 root 
words and about 35,000 proper names, is very sat- 
isfactory, it is inevitable that there will be forms 
in the corpora being processed that are not recog- 
nized by the morphological analyzer. These are al- 
most always foreign proper names, words adapted 
into the language and not in the lexicon, or very 
obscure technical words. These are nevertheless in- 
flected (using Turkish word formation paradigms) 
with inflectional features demanded by the syntactic 
context and sometimes even go through derivational 
processes. For improved disambiguation, one has to 
at least recover any morphological features even if 
the root word is unknown. To deal with this, we 
have made the assumption that all unknown words 
have nominal roots, and built a second morphologi- 
cal analyzer whose (nominal) root lexicon recognizes 
S + where S is the Turkish surface alphabet (in the 
two-level morphology sense), but then tries to in- 
terpret an arbitrary postfix of the unknown word 
as a sequence of Turkish suffixes subject to all mor- 
phographemic constraints. For instance when a form 
such as talkshowumun is entered, this second ana- 
lyzer hypothesizes the following analyses: 
I. \[\[CAT NOUN\] \[ROOT talkshowumun\] 
\[AGR 3SG\] \[POSS NONE\] \[CASE NOM\]\] 
2. \[\[CAT NOUN\] \[ROOT talkshowumu\] 
\[AGR 3SG\] \[POSS 2SG\] \[CASE NOM\]\] 
3. \[\[CAT NOUN\] \[ROOT talksho~um\] 
\[AGR 3SG\] \[POSS NONE\] \[CASE GEN\]\] 
4. \[\[CAT NOUN\] \[ROOT talkshowum\] 
\[AGR 3SG\] \[POSS 2SG\] \[CASE NOM\]\] 
5. \[\[CAT NOUN\] \[ROOT talksho~u\] 
\[AGR 3SG\] \[POSS 1SG\] \[CASE GENII 
6. \[\[CAT NOUN\] \[ROOT talkshow\] 
\[AGR 3SG\] \[POSS ISG\] \[CASE GEN\]\] 
which are then processed just like any other during 
disambiguation.S 
This however is not a sufficient solution for some 
very obscure situations where for the foreign word 
is written using its, say, English orthography, while 
suffixation goes on according to its English pronun- 
ciation, which may make some constraints like vowel 
8Incidentally, the correct analysis is the 6 th, meaning 
o.\[ my talk show. The 5 th one has the same morphological 
features except for the root. 
73 
harmony inapplicable on the graphemic representa- 
tion, though harmony is in effect in the pronuncia- 
tion. For instance one sees the form Carter'a where 
the last vowel in Carter is pronounced so that it 
harmonizes with a in Turkish, while the e in the 
surface form does not harmonize with a. We are 
nevertheless rather satisfied with our solution as in 
our experiments we have noted that well below 1% 
of the forms remain as unknown and these are usu- 
ally item markers in formatted or itemized lists, or 
obscure foreign acronyms. 
3.3 Constraint Rules 
The system uses rules of the sort 
if LC and RC then choose PARSE or 
if LC and RC then delete PARSE 
where LC and RC are feature constraints on unam- 
biguous left and right contexts of a given token, and 
PARSE is a feature constraint on the parse(s) that is 
(are) chosen (or deleted) in that context if they are 
subsumed by that constraint. Currently the left and 
right contexts can be at most 2 tokens, hence we 
look at a window of at most 5 tokens of which one 
is ambiguous. We refer to the unambiguous tokens 
in the context as llc (left-left context) lc (left con- 
text), rc (right context) and rrc (right-right con- 
text). Depending on the amount of unambiguous 
tokens in a context, our rules can have one of the 
following context structures, listed in order of de- 
creasing specificity: 
i. llc, Ic .... rc, rrc 
2. llc, ic .... 
rc, rrc 
3. ic rc 
4. lc 
rc 
To illustrate the flavor of our rules we can give 
the following examples. The first example chooses 
parses with case feature ablative, preceding an un- 
ambiguous postposition which subcategorizes for an 
ablative nominal form. 
\[llc: \[\] ,Ic: \[\] , 
choose : \[case : abl\] , 
rc: \[\[cat :postp,subcat :abl\]\] ,rrc: \[\]\] 
A second example rule is 
\[llc : \[ \[cat : adj , type : determiner\] \] , 
ic: \[\[cat :adj ,stem: \[cat :noun\]\]\] , 
choose: \[cat :adj\] , 
rc:\[\[cat:noun,poss:'NONE'\]\], rrc:\[\]\]. 
which selects and adjective parse following a deter- 
miner, adjective sequence, and before a noun with- 
out a possessive marker. 
Another sample rule is: 
\[llc: \[\] ,Ic: \[ \[agr: '2SG' ,case:gen\]\] , 
choose: \[cat :noun,poss: '2SG'\] , 
rc: \[3 ,rrc: \[\]3 
which chooses a nominal form with a possessive 
marker 2SG following a pronoun with 2SG agree- 
ment and genitive case, enforcing the simplest form 
of noun-noun form noun phrase constraints. 
Our system uses two hand-crafted sets of rules, 
in combination with the rules that are learned by 
unsupervised learning: 
1. We use an initial set of hand-crafted choose 
rules to speed-up the learning process by cre- 
ating disambiguated contexts over which statis- 
tics can be collected. These rules (examples 
of which are given above) are independent of 
the corpus that is to be tagged, and are lin- 
guistically motivated. They enforce some very 
common feature patterns especially where word 
order is rather strict as in NP's or PP's. 9 
The motivation behind these rules is that they 
should improve precision without sacrificing re- 
call. These are rules which impose very tight 
constraints so as not to make any recall errors. 
Our experience is that after processing with 
these rules, the recall is above 99% while pre- 
cision improves by about 20 percentage points. 
Another important feature of these rules is that 
they are applied even if the contexts are also 
ambiguous, as the constraints are tight. That 
is, if each token in a sequence of, say, three am- 
biguous tokens have a parse matching one of the 
context constraints (in the proper order), then 
all of them are simultaneously disambiguated. 
In hand crafting these rules, we have used our 
experience from an earlier tagger (Oflazer and 
Kuruhz, 1994). Currently we use 288 hand- 
crafted choose rules. 
2. We also use a set of hand-crafted heuristic delete 
rules to get rid of any very low probability 
parses. For instance, in Turkish, postpositions 
have rather strict contextual constraints and if 
there are tokens remaining with multiple parses 
one of which is a postposition reading, we delete 
that reading. Our experience is that these rules 
improve precision by about 10 to 12 additional 
percentage points with negligible impact on re- 
call. Currently we use 43 hand-crafted delete 
rules. 
3.4 Learning Choose Rules 
Given a training corpus, with tokens annotated with 
possible parses (projected over selected features), we 
first apply the hand-crafted rules. Learning then 
goes on as a number of iterations over the training 
corpus. We proceed with the following schema which 
is an adaptation of Brill's formulation (Brill, 1995b): 
9Turkish is a free constituent order language whose 
unmarked order is SOV. 
74 
1. We generate a table, called incontext, of all 
possible unambiguous contexts which contain a 
token with an unambiguous (projected) parse, 
along with a count of how many times this 
parse occurs unambiguously in exactly the same 
context in the corpus. We refer to an en- 
try in table with a context C and parse P as 
incontext(C, P). 
2. We also generate a table, called count, of all 
unambiguous parses in the corpus along with a 
count of how many times this parse occurs in 
the corpus. We refer to an entry in this table 
with a given parse P, as count(P). 
3. We then start going over the corpus token by 
token generating contexts as we go. 
4. For each unambiguous context encountered, 
C = (LC, RC) 1° around an ambiguous token w 
with parses P1,. • • Pk, and for each parse Pi, we 
generate a candidate rule of the sort 
if LC and RC then choose Pi 
5. Every such candidate rule is then scored in the 
following fashion: 
(a) We compute 
Pma~ = argmaxpj (j#i) count(Pj)" 
incontext( C, Pj ). 
(b) The score of the candidate rule is then com- 
puted as: 
Scorei = incontext(C, Pi) - count(P~) count(Pmax)" 
incontext( C, Pmaz) 
6. We order all candidate rules generated during 
one pass over the corpus, along two dimensions: 
(a) we group candidate rules by context speci- 
ficity (given by the order in Section 3.3), 
(b) in each group, we order rules by descending 
score. 
We maintain score thresholds associated with 
each context specificity group: the threshold of 
a less specific group being higher than that of 
a more specific group. We then choose the top 
scoring rule from any group whose score equals 
or exceeds the threshold associated with that 
group. The reasoning is that we prefer more 
specific and/or high scoring rules: high scor- 
ing rules are applicable, in general, in more 
places; while more specific rules have stricter 
constraints and more accurate morphological 
parse selections, We have noted that choosing 
the highest scoring rule at every step may some- 
times make premature commitments which can 
not be undone later. 
1°Either of LC or RC may be empty. 
7. The selected rules are then applied in the 
matching contexts and ambiguity in those con- 
texts is reduced. During this application the 
following are also performed: 
(a) if the application results in an unambigu- 
ous parse in the context of the applied rule, 
we increment the count associated with this 
parse in table count. We also update the 
incontext table for the same context, and 
other contexts which contains the disam- 
biguated parse. 
(b) we also generate any new unambiguous 
contexts that this newly disambiguated to- 
ken may give rise to, and add it to the 
incontext table along with count 1. 
Note that for efficiency reasons, rule candidates 
are not generated repeatedly during each pass 
over the corpus, but rather once at the begin- 
ning, and then when selected rules are applied 
to very specific portions of the corpus. 
8. If there are no rules in any group that exceed 
its threshold, group thresholds are reduced by 
multiplying by a damping constant d (0 < d < 
1) and iterations are continued. 
9. If the threshold for the most specific context 
falls below a given lower limit, the learning pro- 
cess is terminated. 
Some of the rules that have been generated by this 
learning process are given below: 
1. Disambiguate around a coordinating conjunc- 
tion: 
\[llc: \[\] ,ic: \[\] , 
choose : \[cat :noun,agr: 3SG ,case :nom\], 
rc : \[ \[cat :conn, root : re\] \] , 
rrc : \[ \[cat : noun, agr : 3SG, poss : NONE\] \] \] 
2. Choose participle form adjectival over a nomi- 
nal reading: 
\[llc: \[\] ,Ic: \[\], 
choose : \[cat : adj, suffix : yah\], 
rc : \[ \[cat : noun, agr : 3SG, poss : NONE\] \] , 
rrc: \[\[cat :noun,agr:3SG,poss: 3SG\]\]\] . 
3. Choose a nominal reading (over an adjectival) 
if a three token compound noun agreement can 
be established with the next two tokens: 
\[llc: \[\] ,lc: \[\] , 
choose : \[cat :noun,agr: 3SG ,case :nom\], 
rc : \[ \[cat :noun, agr : 3SG,poss : 3SG\] \], 
rrc : \[ \[cat : noun, agr : 3SG,poss : 3SG\] \] \] 
3.4.1 Contexts induced by morphological 
derivation 
The procedure outlined in the previous section has 
to be modified slightly in the case when the unam- 
biguous token in the rc position is a morphologi- 
cally derived form. For such cases one has to take 
into consideration additional pieces of information. 
75 
We will motivate this using a simple example from 
Turkish. Consider the example fragment: 
... bir masa+dlr. 
... a table+is 
... is a table 
where the first token has the morphological parses: 
I. \[\[CAT ADJ\] \[ROOT bir\] \[TYPE CARDINAL\]\] 
(one) 
2. \[\[CAT ADJ\] \[ROOT bir\] \[TYPE DETERMINER\]\] 
(a) 
3. \[\[CAT ADVERB\] \[ROOT bir\]\] 
(only/merely) 
and the second form has the unambiguous morpho- 
logical parse: 
1. \[\[CAT NOUN\] \[ROOT masa\] \[AGR 3SG\] \[POSS NONE\] 
\[CASE NOM\] \[CONV VERB NONE\] 
\[TAM1PRES\] \[AGR 3SG\]\] (is table) 
which in hierarchical formcorresponds to the ~ature 
VERB 
PRES 
3SG 
ROOT masa 
STEM 3SG 
I POSS NONE 
LCASE NOM 
SUFFIX NONE 
In the syntactic context this fragment is interpreted 
as 
structure: 
"C AT 
TAM1 
%GR 
VP 
NP +dlr 
DET NOUN 
I I bir masa 
where the determiner is attached to the noun and 
the whole phrase is then taken as a VP although the 
verbal marker is on the second lexical item. If, in 
this case, the token bit is considered to neighbor a 
token whose top level inflectional features indicate 
it is a verb, it is likely that bit will be chosen as 
an adverb as it precedes a verb, whereas the correct 
parse is the determiner reading. 
In such a case where the right context of an am- 
biguous token is a derived form, one has to con- 
sider as the right context, both the top level features 
of final form, and the stem from which it was de- 
rived. During the set-up of the incontext table, such 
a context is entered twice: once with the top level 
feature constraints of the immediate unambiguous 
right-context, and once with the feature constraints 
of the stem. The unambiguous token in the right 
context is also entered to the count table once with 
its top level feature structure and once with the fea- 
ture structure of the stem. 
When generating candidate choose or delete rules, 
for contexts where rc is a derived form and rrc is 
empty, we actually generate two candidates rules for 
each ambiguous token in that context: 
1. if llc, ic and rc then choose/delete Pi. 
2. if llc, Ic and stem(re) then choose/delete 
P~. 
These candidate rules are then evaluated as de- 
scribed above. In general all derivations in a lexical 
form have to be considered though we have noted 
that considering one level gives satisfactory results. 
3.4.2 Ignoring Features 
Some morphological features are only meaningful 
or relevant for disambiguation only when they ap- 
pear to the left or to the right of the token to be 
disambiguated. For instance, in the case of Turkish, 
the CASE feature of a nominal form is only useful in 
the immediate left context, while the POSS (the pos- 
sessive agreement marker) is useful only in the right 
context. If these features along with their possible 
values are included in context positions where they 
are not relevant, they "split" scores and hence cause 
the selection of some other irrelevant rule. Using the 
maxim that union gives strength, we create contexts 
so that features not relevant to a context position are 
not included, thereby treating context that differ in 
these features as same. 11 
3.5 Learning Delete Rules 
For choosing delete rules we have experimented with 
two approaches. One obvious approach is to use 
the formulation described above for learning choose 
rules, but instead of generating choose rules, pick 
the parses that score (significantly) worse than and 
generate delete rules for such parses. We have imple- 
mented this approach and found that it is not very 
desirable due to two reasons: 
1. it generates far too many delete rules, and 
2. it impacts recall seriously without a correspond- 
ing increase in precision. 
The second approach that we have used is consid- 
erably simpler. We first reprocess the training cor- 
pus but this time use a second set of projection tem- 
plates, and apply initial rules, learned choose rules 
and heuristic delete rules. Then for every unambigu- 
ous context C = (LC, RC), with either an immediate 
left, or an immediate right components or both (so 
n Obviously these features are specific to a language. 
76 
the contexts used here are the last 3 in Section 3.3), 
a score incontext( C, Pi ) 
count ( Pi ) 
for each parse Pi of the (still) ambiguous token, is 
computed. Then, delete rules of the sort 
if LC and RC then delete Pi 
are generated for all parses with a score below a cer- 
tain fraction (0.2 in our experiments) of the highest 
scoring parse. In this process, our main goal is to 
remove any seriously improbable parses which may 
somehow survive all the previous choose and delete 
constraints applied so far. Using a second set of tem- 
plates which are more specific than the templates 
used during the learning of the choose rules, we in- 
troduce features we were originally projected out. 
Our experience has been that less strict contexts 
(e.g., just alc or rc) generate very useful delete 
rules, which basically weed out what can (almost) 
never happen as it is certainly not very feasible to 
formulate hand-crafted rules that specify what se- 
quences of features are not possible. 
Some of the interesting delete rules learned here 
are: 
1. Delete the first of two consecutive verb parses: 
\[llc : \[\] , lc: \[\] , 
delete : \[cat : verb\] , 
re: \[\[cat :verb\]\],rrc: \[\]\] 
2. Delete accusative case marked noun parse be- 
fore a postposition that subcategorizes for a 
nominative noun: 
\[llc: \[\] ,Ic: \[\] , 
delete : \[cat :noun, agr : 3SG,poss : NONE, case : acc\] , 
rc : \[ \[cat : postp, subcat : nora\] \] ,rrc : \[\] \] . 
3. Delete the accusative case marked parse with- 
out any possessive marking, if the previous form 
has genitive case marking (signaling a genitive- 
possessive NP construction): 
\[llc: \[\] , 
lc : \[ \[cat : noun, agr : 3SG,poss : NONE, case : gen\] \], 
delete : \[cat :noun, agr: 3SG ,poss : NONE, case : ace\], 
re: \[\] ,rrc: \[\]\]. 
3.6 Using context statistics to delete parses 
After applying hand-crafted rules to a text to be dis- 
ambiguated we arrive at a state where ambiguity is 
about 1.10 to 1.15 parses per token (down from 1.70 
to 1.80 parses per token) without any serious loss 
on recall. This state allows statistics to be collected 
over unambiguous contexts. To remove additional 
parses which never appear in any unambiguous con- 
text we use the scoring described above for choosing 
delete rules, to discard parses on the current text 
based on context statistics} 2 We make three passes 
12Please note that delete rules learned may be applied 
to future texts to be disambiguated, while this step is 
over the current text, scoring parses in unambigu- 
ous contexts of the form used in generating delete 
rules, and discarding parses whose score is below a 
certain fraction of the maximum scoring parse, on 
the fly. The only difference with the scoring used for 
delete rules, is that the score of a parse Pi here is a 
weighted sum of the quantity 
incontext(C, Pi) 
count( Pi ) 
evaluated for three contexts in the case both the lc 
and rc are unambiguous/ 
3.7 Steps in Disamblguating a Text 
Given a new text annotated with all morphological 
parses (this time the parses are not projected), we 
proceed with the following steps for disambiguation: 
1. The initial hand-crafted choose rules are applied 
first. These rules always constrain top level in- 
flectional features, and hence, any stems fromn 
derivational processes are not considered unless 
explicitly indicated in the constraint itself. 
2. The hand-crafted delete clean-up rules are ap- 
plied. 
3. Context statistics described in the preceding 
section are used to discard further parses. 
4. The choose rules that have been learned ear- 
lier, are then repeatedly applied to unambigu- 
ous contexts, until no more ambiguity reduc- 
tion is possible. During the application of these 
rules, if the immediate right context of a token 
is a derived form, then the stem of the right 
context is also checked against the constraint 
imposed by the rule. So if the rule right context 
constraint subsumes the top level feature struc- 
ture or the stem feature structure, then the rule 
succeeds and is applied if all other constraints 
are also satisfied. 
5. Finally, the delete rules that have been learned 
are applied repeatedly to unambiguous con- 
texts, until no more ambiguity reduction is pos- 
sible. 
4 Experimental Results 
We have applied our learning system to two Turk- 
ish texts. Some statistics on these texts are given 
in Table 1. The first text labeled ARK is a short 
text on near eastern archaeology. The second text 
from which fragments whose labels start with C are 
derived, is a book on early 20 ~h history of Turkish 
Republic. 
In Table 1, the tokens considered are that are gen- 
erated after morphological analysis, unknown word 
processing and any lexical coalescing is done. The 
applied to the current text on which disambiguation is 
performed. 
77 
Text 
ARK 
C2400 
C270 
Sentences 
492 
2,407 
270 
Tokens 
0 1 
7,928 0.15% 49.34% 
39,800 0.03% 50.56% 
5212 0.02% 50.63% 
Distribution 
of 
Morphological Parses 
2 3 4 >4 
30.93% 9.19% 8.46% 1.93% 
28.66% 10.12% 8.16% 2.47% 
30.68% 8.62% 8.36& 1.69% 
Table 1: Statistics on Texts 
words that are unknown are those that could not 
even be processed by the unknown noun proces- 
sor. Whenever an unknown word had more than one 
parse it was counted under the appropriate group. 
We learned rules from ARK itself, and on the 
first 500, 1000, and 2000 sentence portions of C2400. 
C270 which was from the remaining 400 sentences of 
C2400 was set aside for testing. Gold standard dis- 
ambiguated versions for ARK, C270 were prepared 
manually to evaluate the automatically tagged ver- 
sions. 
Our results are summarized in the following set 
of tables. Tables 2 and 3 give the ambiguity, re- 
call and precision initially, after hand-crafted rules 
are applied, and after the contextual statistics are 
used to remove parses - all applications being cu- 
mulative. The rows labeled BASE give the initial 
state of the text to be tagged. The rows labeled 
INITIAL CHOOSE give the state after hand-crafted 
choose rules are applied, while the rows labeled INI- 
TIAL DELETE give the state after the hand-crafted 
choose and delete rules are applied. The rows la- 
beled CONTEXT STATISTICS give the state after 
the rules are applied and context statistics are used 
(as described earlier) to remove additional parses. 
Disambiguation Ambiguity Recall Pre. 
Stage (~) (~) 
BASE 1.828 100.OO 54.69 
INITIAL CHOOSE 1.339 99.28 74.13 
INITIAL DELETE 1.110 99.08 88.91 
CONTEXT STATISTICS 1.032 97.38 94.35 
Table 2: Average parses, recall and precision for text 
ARK 
Disambiguation 
Stage 
Ambiguity 
BASE 1.719 
INITIAL CHOOSE 1.353 
INITIAL DELETE 1.130 
CONTEXT STATISTICS 1.038 
Recall Pre. (%) (%) 
100.00 58.18 
99.16 73.27 
98.73 87.24 
96.70 93.15 
Table 3: Average parses, recall and precision for text 
C270 
Tables 5 and 6 present the results of further dis- 
ambiguation of ARK, and C270 using rules learned 
from training texts C500, C1000, C2000 and ARK. 
These rules are applied after the last stage in the ta- 
bles above. 13 The number of rules learned are given 
in Table 4.14 
Training Choose Delete 
Text Rules Rules 
ARK 23 89 
C500 ii 113 
C1000 29 195 
C2000 61 245 
Table 4: Number of choose and delete rules learned 
from training texts. 
Disambiguation Ambiguity Recall Pre. 
Stage (%) (°Z~) 
Training Set ARK 
I I LEARNED DELETE 1.027 97.20 94.63 
Training Set C5O0 
LEARNED DELETE 1.028 97.30 94.61 
Training Set C1000 
I I LEARNED DELETE 1.026 97.18 94.68 
Training Set C2000 
LEARNED CHOOSE 1.028 I 97.24 I 94.60 
LEARNED DELETE 1.025 97.1394.71 
Table 5: Average parses, recall and precision for text 
ARK after applying learned rules. 
Table 7 gives some additional statistical results at 
the sentence level, for each of the test texts. The 
columns labeled UA/C and A/C give the number 
and percentage of the sentences that are correctly 
disambiguated with one parse per token, and with 
more than one parse for at least one token, respec- 
tively. The columns labeled 1, 2, 3, and >3 denote 
the number and percentage of sentences that have 
1, 2, 3, and >3 tokens, with all remaining parses 
incorrect. It can be seen that well 60% of the sen- 
tences are correctly morphologically disambiguated 
with very small number of ambiguous parses remain- 
ing. 
13Please note for ARK, in the first two rows, the train- 
ing and the test texts are the same. 
nLearning iterations have been stopped when the 
maximum rule score fell below 7. 
78 
Text 
ARK 
C270 
Sentences 
Total UA/C \[ A/C C (UA/C+A/C) I 2 3 >3 
494 220 (44.53%) 97 (19.6,I%) 317 (64.17%) 133 (26.92%) 41 (8.30%) 3 (0.61%) 0 (0.00%) 
270 116 (42.96%) 50 (18.52%) 166 (61.48%) 55 (20.37%,) 27 (10.00%) 17 (6.30%) 5 (1.85%) 
Table 7: Disambiguation results at the sentence level using rules learned from C2000. 
Disambiguation Stage Ambiguity \[ Recall (%) Pre. \] (~) 
Training Set ARK 
LEARNED CHOOSE 1.035 
LEARNED DELETE 1.029 
~ainingSet 
LEARNED CHOOSE 
LEARNED DELETE 
96.64 93.36 
96.40 93.71 
C500 I 1.035 96.66 93.32 
1.029 96.40 93.66 
Training Set C1000 
LEARNED CHOOSE 1.035 
LEARNED DELETE 1.029 
Training Set C20O0 
LEARNED CHOOSE 1.034 
LEARNED DELETE 1.030 
96.66 93.34 
96.42 93.64 
96.64 \] 93.41 
96.52 93.70 
Table 6: Average parses, recall and precision for text 
270 after applying learned rules. 
4.1 Discussion of Results 
We can make a number of observations from our 
experience: Hand-crafted rules go a long way in im- 
proving precision substantially, but in a language 
like Turkish, one has to code rules that allow no, or 
only carefully controlled derivations, otherwise lots 
of things go massively wrong. Thus we have used 
very tight and conservative rules in hand-crafting. 
Although the additional impact of choose and rules 
that are induced by the unsupervised learning is not 
substantiM, this is to be expected as the stage at 
which they are used is when all the "easy" work 
has been done and the more notorious cases re- 
main. An important class of rules we explicitly have 
avoided hand crafting are rules for disambiguating 
around coordinating conjunctions. We have noted 
that while learning choose rules, the system zeroes 
in rather quickly on these contexts and comes up 
with rather successful rules for conjunctions. Simi- 
larly, the delete rules find some interesting situations 
which would be virtually impossible to enumerate. 
Although it is easy to formulate what things can go 
together in a context, it is rather impossible to for- 
mulate what things can not go together. 
We have also attempted to learn rules directly 
without applying any hand-crafted rules, but this 
has resulted in a failure with the learning process 
getting stuck fairly early. This is mainly due to the 
lack of sufficient unambiguous contexts to bootstrap 
the whole disambiguation process. 
From analysis of our results we have noted that 
trying to choose one correct parse for every token is 
rather ambitious (at least for Turkish). There are a 
number of reasons for this: 
There are genuine ambiguities. The word o is 
either a personal or a demonstrative pronoun 
(in addition to being a determiner). One simply 
can not choose among the first two using any 
amount of contextual information. 
A given word may be interpreted in more than 
one way but with the same inflectional features, 
or with features not inconsistent with the syn- 
tactic context. This usually happens when the 
root of one of the forms is a proper prefix of 
the root of the other one. One would need se- 
rious amounts of semantic, or statistical root 
word and word form preference information for 
resolving these. For instance, in 
koyun sfirfisfi 
koyun sfirfi+sfi 
sheep herd+POSS-3SG 
(sheep herd) 
koy+un siirfi+sfi 
bay+GEN herd+POSS-3SG 
(?? bay's herd) 
both noun phrases are syntactically possible, 
though the second one is obviously nonsense. 
It is not clear how one would disambiguate this 
using just contextual or syntactic information. 
Another similar example is: 
kurmaya yardlm etti 
kur+ma+ya yardlm et+ti 
construct+INF+DAT help make+PAST 
helped construct (something) 
kurmay+a yard~m et+ti 
milit ary-officer+DAT help make+PAST 
helped the military-officer 
where again with have a similar problem. It 
may be possible to resolve this one using sub- 
categorization constraints on the object of the 
verb kur assuming it is in the very near preced- 
ing context, but this may be very unlikely as 
Turkish allows arbitrary adjuncts between the 
object and the verb. 
Turkish allows sentences to consist of a number 
of sentences separated by commas. Hence locat- 
ing a verb in the middle of a sentence is rather 
difficult, as certain verbal forms also have an 
adjectival reading, and punctuation is not very 
helpful as commas have many other uses. 
The distance between two constituents (of, say, 
a noun phrase) that have to agree in vari- 
ous morphosyntactic features may be arbitrar- 
79 
ily long and this causes occasional mislnatches, 
especially if the right nominal constituent has 
a surface plural marker which causes a 4-way 
ambiguity, as in masalam. 
masalarI 
I. \[\[CAT NOUN\] \[ROOT masa\] \[AGR 3PL\] 
\[POSS NONE\] \[CASE ACC\]\] 
(tables accusative) 
2. \[\[CAT NOUN\] \[ROOT masa\] \[AGR 3PL\] 
\[POSS 3SG\] \[CASE NOM\]\] 
(his tables) 
3. \[\[CAT NOUN\] \[ROOT masa\] 
\[POSS 3PL\] \[CASE NON\]\] 
(their tables) 
\[AGR 3PL\] 
4. \[\[CAT NOUN\] \[ROOT masa\] \[AGR 3SG\] 
\[POSS 3PL\] \[CASE NOM\]\] 
(their table) 
Choosing among the last three is rather prob- 
lematic if the corresponding genitive form to 
force agreement with is outside the context. 
Among these problems, the most crucial is the 
second one which we believe can be solved to a great 
extent by using root word preference statistics and 
word form preference statistics. We are currently 
working on obtaining such statistics. 
5 Conclusions 
This paper has presented a rule-based morphologi- 
cal disambiguation approach which combines a set of 
hand-crafted constraint rules and learns additional 
rules to choose and delete parses, from untagged text 
in an unsupervised manner. We have extended the 
rule learning and application schemes so that the 
impact of various morphological phenomena and fea- 
tures are selectively taken into account. We have ap- 
plied our approach to the morphological disambigua- 
tion of Turkish, a free-constituent order language, 
with agglutinative morphology, exhibiting produc- 
tive inflectional and derivational processes. We have 
also incorporated a rather sophisticated unknown 
form processor which extracts any relevant inflec- 
tional or derivational markers even if the root word 
is unknown. 
Our results indicate that by combining these 
hand-crafted, statistical and learned information 
sources, we can attain a recall of 96 to 97% with 
a corresponding precision of 93 to 94% and ambigu- 
ity of 1.02 to 1.03 parses per token, on test texts, 
however the impact of the rules that are learned is 
not significant as hand-crafted rules do most of the 
easy work at the initial stages. 
6 Acknowledgments 
We would like to thank Xerox Advanced Document 
Systems, and Lauri Karttunen of Xerox Pare and 
of Rank Xerox Research Centre (Grenoble) for pro- 
viding us with the two-level transducer development 
software on which the morphological and unknown 
word recognizer were implemented. This research 
has been supported in part by a NATO Science for 
Stability Grant TU-LANGUAGE. 

References 
Eric Brill. 1992. A simple-rule based part-of-speech 
tagger. In Proceedings of the Third Conference 
on Applied Natural Language Processing, Trento, 
Italy. 
Eric Brill. 1994. Some advances in rule-based part of 
speech tagging. In Proceedings of the Twelfth Na- 
tional Conference on Articial Intelligence (AAAI- 
9~), Seattle, Washinton. 
Eric Brill. 1995a. Transformation-based error- 
driven learning and natural language processing: 
A case study in part-of-speech tagging. Computa- 
tional Linguistics, 21(4):543-566, December. 
Eric Brill. 1995b. Unsupervised learning of dis- 
ambiguation rules for part of speech tagging. In 
Proceedings of the Third Workshop on Very Large 
Corpora, Cambridge, MA, June. 
Kenneth W. Church. 1988. A stochastic parts pro- 
gram and a noun phrase parser for unrestricted 
text. In Proceedings of the Second Conference 
on Applied Natural Language Processing, Austin, 
Texas. 
Doug Cutting, Julian Kupiec, Jan Pedersen, and 
Penelope Sibun. 1992. A practical part-of-speech 
tagger. In Proceedings of the Third Conference 
on Applied Natural Language Processing, Trento, 
Italy. 
Steven J. DeRose. 1988. Grammatical category dis- 
ambiguation by statistical optimization. Compu- 
tational Linguistics, 14(1):31-39. 
Zelal GiingSrdii and Kemal Oflazer. 1995. Pars- 
ing Turkish using the Lexical-Functional Gram- 
mar formalism. Machine Translation, 11(4):293- 
319. 
Jorge Hankamer. 1989. Morphological parsing and 
the lexicon. In W. Marslen-Wilson, editor, Lexical 
Representation and Process. MIT Press. 
Fred Karlsson, Atro Voutilainen, Juha HeikkilS, 
and Arto Anttila. 1995. Constraint Grammar-A 
Language-Independent System for Parsing Unre- 
stricted Text. Mouton de Gruyter. 
Lauri Karttunen. 1993. Finite-state lexicon com- 
piler. XEROX, Palo Alto Research Center- Tech- 
nical Report, April. 
ilker Kuru6z. 1994. Tagging and morphological 
disambiguation of Turkish text. Master's thesis, 
Bilkent University, Department of Computer En- 
gineering and Information Science, July. 
Moshe Levinger, Uzzi Ornan, and Alon Itai. 1995. 
Learning morpho-lexical probabilities fi'om an 
untagged corpus with an application to He- 
brew. Computational Linguistics, 21(3):383-404, 
September. 
Kemal Oflazer and ilker KuruSz. 1994. Tagging and 
morphological disambiguation of Turkish text. In 
Proceedings of the 4 th Applied Natural Language 
Processing Conference, pages 144-149. ACL, Oc- 
tober. 
Kemal Oflazer. 1993. Two-level description of Turk- 
ish morphology. In Proceedings of the Sixth Con- 
ference of the European Chapter of the Associa- 
tion for Computational Linguistics, April. A full 
version appears in Literary and Linguistic Com- 
puting, Vol.9 No.2, 1994. 
Atro Voutilainen and Pasi Tapanainen. 1993. Am- 
biguity resolution in a reductionistic parser. In 
Proceedings of EACL'93, Utrecht, Holland. 
Atro Voutilainen, auha Heikkilg, and Arto Anttila. 
1992. Constraint Grammar of English. University 
of ttelsinki. 
Atro Voutilainen. 1995a. Morphological disam- 
biguation. In Fred Karlsson, Atro Voutilainen, 
Juha Heikkilg, and Arto Anttila, editors, Con- 
straint Grammar-A Language-Independent Sys- 
tem for Parsing Unrestricted Tea:t, chapter 5. 
Mouton de Gruyter. 
Atro Voutilainen. 1995b. A syntax-based part-of- 
speech analyzer, In Proceedings of the Seventh 
Conference of the European Chapter of the Asso- 
ciation of Computational Linguistics, Dublin, Ire- 
land. 
