=Paper=
{{Paper
|id=Vol-1170/CLEF2004wn-QACLEF-OsenovaEt2004
|storemode=property
|title=Bulgarian-English Question Answering: Adaptation of Language Resources
|pdfUrl=https://ceur-ws.org/Vol-1170/CLEF2004wn-QACLEF-OsenovaEt2004.pdf
|volume=Vol-1170
|dblpUrl=https://dblp.org/rec/conf/clef/OsenovaSSTK04
}}
==Bulgarian-English Question Answering: Adaptation of Language Resources==
https://ceur-ws.org/Vol-1170/CLEF2004wn-QACLEF-OsenovaEt2004.pdf
Bulgarian-English Question Answering: Adaptation
of Language Resources
Petya Osenova∗ Alexander Simov† Kiril Simov‡ Hristo Tanev§
Milen Kouylekov¶
Abstract
This paper describes the Bulgarian part of a Bulgarian–English question answer-
ing system. The Bulgarian modules are implemented as a question analysis procedure
within a Bulgarian question answering system — BulQA. The paper presents the avail-
able language resources and corresponding technology which is used for the analysis
of the questions in Bulgarian and their translation into English format necessary for
answer extraction. As implementation platform CLaRK System is used.
1 Introduction
This paper describes the first steps in the development of a question answering system for
Bulgarian — BulQA. The system is planned to have three main modules: Question analysis
module, Interface module, Answer extraction module. The Question analysis module deals
with the syntactic and semantic interpretation of the question. The result of this module is
independent on task and domain representation of the syntactic and semantic information in
the question. The Interface module maps the interpretation received from the first module
to the input necessary for the third module. The Answer extraction module is responsible for
the actual finding of the answer in the corresponding corpus. This architecture allows reusing
some of these modules in other tasks such as - Bulgarian as source language in a multilingual
question answering or Bulgarian sa target language. In general, only the Interface module has
to be re-implemented in order to tune the connection between Bulgarian modules and the
modules for the other languages.
Here we describe the current question analysis module and the Interface module in a
Bulgarian to English question answering system. In this system the Answer searching module
is based on the Diogene system implemented at the ITC-Irst, Trento, Italy.
The structure of the paper is as follows: first we describe the language resources and tools
developed within the BulTreeBank Project, then in section 3 we discuss their adaptation
for the analysis of Bulgarian questions, section 4 describes briefly the DIOGENE system for
document retrieval and answer extraction, in section 5 we discuss the interface between the
system BulQA for the analysis of Bulgarian questions and DIOGENE System, section 6 gives
a general overview of the CLaRK System in which the modules of BulQA are implemented,
the last section reports on the results of the question answering track and concludes the paper.
2 The BulTreeBank Language Resources and Tools
In this section we describe the available language resources and tools which we have adapted in
order to implement the Question analysis module for Bulgarian. Generally, a language technol-
ogy is supposed to include the following modules: tokenization and named entities recognition,
∗ Laboratory for Linguistic Modelling, Bulgarian Academy of Sciences, Bulgaria, petya@bultreebank.org
† Laboratory for Linguistic Modelling, Bulgarian Academy of Sciences, Bulgaria, alex@bultreebank.org
‡ Laboratory for Linguistic Modelling, Bulgarian Academy of Sciences, Bulgaria, kivs@bultreebank.org
§ TC-irst, Trento, Italy, tanev@itc.it
¶ TC-irst, Trento, Italy, kouylekov@itc.it
morphological analyzer and disambiguator, syntactic and semantic analyzer. Most of them
we have already implemented during the creation of a syntactic treebank for Bulgarian —
[Simov, Popova and Osenova 2002].
2.1 BulTreeBank Language Technology
Here we list the tools that we had at our disposal before we started to implement our system:
Tokenization
There is a hierarchy of tokenizers within the CLaRK system, which tokenizes the texts in
an appropriate way. Additionally, one can decide what the category of the token is and to
assign it.
The Morphosyntactic analyzer
It assigns all possible analyses to the word tokens. The lexicon is too large to be loaded
as one grammar in CLaRK and this is why we have divided it into several grammars which
are applied in a group. The separation of the lexicon is on the basis of the frequencies of the
word forms within the corpus. In this way the application has been speeded up. As it was
mentioned above, together with the morphosyntactic analyzer we use the gazetteers. They
are also implemented within the CLaRK system. In the places where competing analyses arise
between a common word and a name or an abbreviation, we try to use the token classification
strategy and the prompts of the context.
MorphoSyntactic Disambiguation
We have already implemented a preliminary version of a rule-based morpho-syntactic
disambiguator, encoded as a set of constraints within the CLaRK system. This rule-based
disambiguator exploits context information like agreement between an adjective and a noun
in a noun phrase, specific positions like a noun after a preposition, but it also deals with some
fixed phrases. The disambiguator does not try to solve unsure cases, but leaves them for fur-
ther processing. Its coverage is about 80 %. For automatic disambiguation we have developed
a neural-network-based disambiguator (see [Simov and Osenova, 2001]). It achieves accuracy
of 95.25% for part-of-speech and 93.17% for complete morpho-syntactic disambiguation.
Partial Grammars
We have constructed grammars for:
1. Sentence splitting. At the moment it is fully automated and reliable only for the basic
and clear cases. For solving complex and ambiguous cases this grammar is combined
with supporting modules for abbreviation detection.
2. Named-entity recognition. Identifying numerical expressions, names, abbreviations,
special symbols (see [Ivanova and Dojkoff 2002], [Osenova and Kolkovska 2002]). They
are designed to work in cooperation with the morphosyntactic analyzer. If necessary,
the grammars can overwrite the analysis of the morphosyntactic analyzer.
3. Chunking. Two basic modules have been developed: an NP chunker ([Osenova 2002],
[Osenova and Kolkovska 2002]) and a VP chunker [Slavcheva 2002]. Generally speaking,
the chunking process conforms to the following requirements: it deals with non-recursive
constituents; relies on a clear-indicator strategy; delays the attachment decisions; ignores
the semantic information; aims at accuracy, not coverage. Additionally, there are chunk
grammars for APs, AdvPs, PPs and some non-problematic clauses.
2.2 BulTreeBank Language Data
From the language resources the most important for the task are the lexicons:
The Morphological Dictionary
The dictionary is an electronic version of [Popov, Simov and Vidinska 1998] extended with
new words from the corpus. It covers the grammatical information of about 100 000 lexemes
(1 600 000 word forms) and serves as a basis for the morphological analyzer.
The Gazetteers
Two basic lists with items, missing in the morphological dictionary, have been compiled
with respect to their frequency:
1. Gazetteers of names. These consist of 15 000 items and include Bulgarian as well as
foreign person names, international and national locations, organizations. The most
frequent names are additionally classified according to three criteria: (1) grammatical
(gender and number); (2) semantic - with respect to an extended SIMPLE core ontology
(names for different types of locations, organizations, artifacts, persons’ social roles etc.)
and (3) ontological - some person names were connected with specific individuals in the
world and thus some encyclopedic information was provided in addition to the semantic
classification. All this information was ready to be used for practical applications like
Information Extraction or Retrieval, Data Mining, Question Answering etc. Special
attention is paid to the names of mountains and artifacts (books, films, broadcasts),
because their internal agreement does not always coincide with the external one, which
is needed for the sentence analysis.
2. Gazetteers of the most frequent abbreviations. They consist of 1500 acronyms and
graphical abbreviations. The acronyms’ extensions were mapped against the names
(mostly organizations) and therefore, assigned the same semantic and grammatical label.
In cases of idiosyncratic grammatical behaviour, the relevant patterns have been added
as well.
3. Gazetteers of the most frequent introductory expressions and parentheticals. This is
considered to be a step towards a basic list of collocations. They were classified according
to their morphological type or behavior: verbal, adverbial, linking (for conjunctions),
nominal (vocatives), idiomatic etc. We use them as an extended supplementary lexicon
during the phase of the syntactic annotation.
The Valence Dictionary
It consists of 1000 most frequent verbs and their valence frames and it is based on a
paper dictionary (see [Balabanova and Ivanova, 2002]). Each frame defines the number and
the kind of the arguments and imposes morphosyntactic and semantic restrictions over them.
The semantic restrictions over the arguments are extracted and matched against the SIMPLE
core ontology. The frames of the most frequent verbs are compared to the corpus data
and repaired if necessary (new frames are added, some of the existing frames are deleted or
fine-grained). We envisage to enlarge the coverage of the dictionary with the help of some
derivational means, such as the verb prefixes.
The Semantic Dictionary
Semantic information plays a crucial role in the process of named entity recogniion. Thus,
in order to support the selectional restrictions imposed by the valence dictionary and to facil-
itate its usage, we decided to compile a semantic dictionary along the guidelines of SIMPLE
project. It is worth mentioning that we follow an extended variant of the SIMPLE core
ontology. At the moment we are classifying the most frequent nouns with respect to the
ontological hierarchy without specifying the synonymic relations between them. Up to now
we have classified about 3 000 nouns. Recall that the named entities also have been classified
with respect to the same ontology.
3 Adaptation to Question Answering Task
Although the above listed language processing tools were extensively tested during the com-
pilation of our treebank, they needed some additional tuning to the task of question analysis.
The main difference is that most of them were implemented in such a way that in unsure cases
the ambiguity remained unresolved or the analysis was not produced. This tools’ application
was required in case when an annotator had to inspect the result of the processing.
With respect to the Question Answering task some ambiguities were resolved in the fol-
lowing way: (1) in ambiguities between 2 and 3 person or 1 and 3 person, always the 3
person was selected; (2) in ambiguities between present and past verb tense, the past tense
was selected, etc. The first ambiguity was resolved because the questions given in CLEF are
never in 1 or 2 person. Resolving between the different tenses in the question with respect to
validation of the found answers is not currently supported by the Answer extraction module.
Some other ambiguities we resolved on the frequency basis only — for each ambiguity class
the most frequent option was selected.
The major addition with respect to the available tools was the construction of a lemmatizer
for Bulgarian. We defined the lemma to be functionally determined by the wordform and its
morphosyntactic characteristics. The cases of ambiguous lemmas are not resolved and all
possible lemmas are assigned to the corresponding wordform. Lemma is used later to access
the semantic information from the semantic dictionary and the English equivalents in the
Bulgarian–English dictionary.
Here is an example of the analysis of the question “През коя година Томас Ман получи
Нобелова награда?” (in English: Which year did Thomas Mann receive the Nobel Prize?):
През
коя
година
Томас
Ман
получи
Нобелова
награда
?
Here each common word is annotated within the following XML element wordform, where the value of attribute ana is the correct morpho-
syntactic tag for the wordform in the given context of its usage. The value of the attribute bf is
a list of the lemmas assigned to the wordform. Names are annotated within the following XML
element Name, where the value of the attribute
ana is the same as above. The value of the attribute sort determines whether this is a name
of a person, a location, an organization or some other entity.
The next level of analysis is the result of the chunk grammars. In the example there
are three NPA elements (NPA stands for a noun phrase of head-adjunct type) and one PP
element. Also, one of the noun phrases is annotated as a name with a sort attribute with
value: NE-Pers.
The result of this analysis has to be translated into the format which the answer extraction
module is used as input.
4 DIOGENE System in Brief
In this section we briefly describe DIOGENE System which was used for document retrieval
and answer extraction. DIOGENE [Negri et al 2002] relies on the knowledge in multilingual
ontology MultiWordNet [Pianta et al 2002], manually created rules for named entity recogni-
tion and question type identification, a set of handcrafted answer extraction templates and
statistical information collected from the Web and off-line multilingual corpora.
In cross-language mode DIOGENE works as follows:
1. The question is processed and all the possible translations of the keywords from the
source language in English are found (for the Bulgarian-English task this is performed
in the BulQA system).
2. Finding correct combination of translations: We chose this combination of keyword
translations (k1 , k2 ,. . .,kn ) which has the highest frequency of co-occurrence in an En-
glish corpus (we used AQUAINT and TIPSTER collections). The main assumption is:
the more often a keyword translation combination appears with the translations close
to each other (in one and the same paragraph), the more plausible this combination is.
3. From keywords and their synonyms DIOGENE forms a Boolean query which is passed
to Managing Gigabytes (MG) search engine. Some keywords can be deleted from the
query if it generates no hit or just a few hits. In this way several feedback loops can be
performed. The output of this processing stage is a list of paragraphs where question
keywords and their synonyms appear together.
4. Named entity recognition and answer extraction templates are applied to extract can-
didate answers. In the cross-language mode DIOGENE applies answer extraction tem-
plates just for the definition questions. Candidate answers of the factoid questions are
captured using named entity recognition and proximity to the question keywords.
5. Finally, the candidate-answers of the factoid questions are evaluated using Web based
answer validation technique described in [Magnini et al 2002]. On the other hand, we
evaluate the answers of the definition questions using different syntactic and semantic
clues, among them is the presence of hyponym of the “person” or “organization” concept,
presence of definite lexical templates, etc.
The format which was necessary to be supplied to DIOGENE System was as follows:
• Head of the question. The head of each question depends on the interrogative word in
the question and helps to determine the kind of the answer. Some examples of question
heads are: what, who, what-who etc.
• Type of the question. It determines the semantic category of the possible answers.
• Head word of the question. It is the word in the question which provides the type
of the question. It can be a non-functional word or the interrogative word.
• Sense of the head word. This is the sense derived from WordNet for the head word.
If such a sense cannot be determined, then the value is NIL.
• Part of speech of the head word. This is the POS tag of the head word with respect
to the Pentreebank tagset.
• Position of the head word. A digit which determines where in the question the head
word is.
• List of key words. A list of the non-functional words in the question. Each keyword
is also annotated with its part of speech.
5 Interface module
Here we describe the implemented interface module which translates the result of the ques-
tion analysis module into the template necessary for DIOGENE System, which extracts the
answers of the questions. The process includes the following steps:
• Determining the head of the question.
The determination of the question head is done by searching for the chunk which contains
the interrogative pronoun. There are cases in which the question is expressed with the
help of imperative forms of verbs: назовете (to name), кажете (to point out; to say),
избройте (to list; to enumerate). After the chunk had been selected we classify the
interrogative pronoun within a hierarchy of question’s heads. In this hierarchy some
other elements of the chunks — mainly prepositions — play an important role as well.
• Determining the head word of the question and its semantic type.
The chunk determined in the previous step also is used for determining the head word
of the question. There are five cases. First, the chunk is an NP chunk in which the
interrogative pronoun is a modifier. In this case the head noun is the head word of the
question. In the second case the chunk is a PP chunk in which there is an NP chunk
similar to the NP chunk from the previous case. Thus, again the head noun is a head
word for the question. Third, the interrogative pronoun is a subject of a copula verb. In
this case the head word of the question is the head noun of the complement NP chunk
of the copula. Here the important moment is the distinction between the cases when
the interrogative pronoun is a subject of the copula and when it is its complement. The
rule covers only the subject case. The fourth case is similar, but it covers the questions
whit imperative verbs. Then again the head of the question is the head noun of the
complement NP chunk. The last case covers all the remaining questions. Then the head
word of the question is the interrogative pronoun itself.
The semantic type of the head word is determined by the annotation of the words with
semantic classes from the semantic dictionary. When there are more than one semantic
classes we add all of them. The type of the interrogative pronoun is used later for
disambiguation. If no semantic class is available in the dictionary, then the class ‘other’
is assigned.
• Determining the type of the question.
The type of the question is determined straightforwardly by the semantic type of the
head word.
• Determining the keywords of the question and their part of speech.
The keywords are determined by the non-functional words in the question. Their part
of speech is determined by a mapping from the Bulgarian tagset into the tagset used
by DIOGENE System. Sometimes it is possible to construct multi-token keywords like
names (Thomas Mann), terms or collocations (Nobel prize). To some extend this is done
after the translation into English.
• Translation of the question head word and the keywords into English.
We have two Bulgarian–English dictionaries: one for the common vocabulary and one
for the names. The dictionary of names contains the transliterations of most frequent
names that we found in Bulgarian corpus and in the English corpus. This dictionary is
necessary because a vast amount of foreign names do not follow the same transliteration
principles for Bulgarian. For instance, Washington as a name of the president George
Washington, the state Washington and the capital of the USA is written as Вашингтон
(Vashington), which follows the literal traditional transliteration, i.e. letter by letter.
However, in all other cases this name is written as Уошингтън, which follows the new
principles of transliteration, i.e. closer to the original pronunciation of the word. For
the names which are not in the dictionary we apply the transliteration for Bulgarian
into Latin as it is defined by the Bulgarian Post Services. Note that the last solution
is far form perfect and it has to be improved afterwards. The main problem is that
this transliteration does not take into account the sound representation of the names
in the original language. For instance, the name Thomas (Томас in Bulgarian) will be
transliterated as Tomas without ‘h’. For that reson, this problem will require much
more work in future. Some names of famous people and places are kept as one whole
expression in the dictionary. For example, ‘Thomas Mann’ is a multi-token name in the
dictionary. This helps us during the translation phase, because of the following: if we
take the two names separately, we can receive, wrongly, also Thomas Man as a potential
translation, where ‘Man’ is transliterated with one ‘n’. For the words which have more
than one translation we give all possibilities.
Another very useful resource is the collocation dictionary for English. For example, the
chunk Нобелова награда (Nobel prize)is a collocation in English, but we also translate
it into Bulgarian as ‘Nobel award’. If we have a collocation dictionary we could use it
in order to recognize such multi-token expressions. In future work we also will try to
use Internet to judge between the different possibilities. For the above examples, ‘Nobel
prize’ is much more frequent than ‘Nobel award’.
Here we give the result of the analysis for the above question:
През
коя
година
Томас
Ман
получи
Нобелова
награда
?
Here the new element is QHead which determines the chunk head of the question. It has
two attributes: qhead which has as a value the question head — what in the example;
and qtype which has as a value the type of the question — time here. Some of the words
received additional attributes: sort for the semantic class of the word, and eng for the
possible translations into English.
• Filling the template.
This step means the conversion of the information that has already beed explicated into
the form necessary for the DIOGENE System. Here also we try to produce multi-token
keywords. In the example above, such a keyword is Thomas Mann — a name that we
had in the dictionary.
All the steps during the analysis of the questions and their transformation into the DIO-
GENE format are implemented in the CLaRK system, which is shortly described in the next
section.
6 CLaRK System
In this section we describe the basic technologies of the CLaRK System1 ([Simov et. al. 2001]).
CLaRK is an XML-based software system for corpora development. It incorporates several
technologies: XML technology; Unicode; Regular Grammars; and Constraints over XML Doc-
uments.
XML Technology
The XML technology is at the heart of the CLaRK System. It is implemented as a
set of utilities for data structuring, manipulation and management. We have chosen the
XML technology because of its popularity, its ease of understanding and its already wide
use in description of linguistic information. In addition to the XML language [XML 2000]
processor itself, we have implemented an XPath language [XPath 1999] engine for navigation
in documents and an XSLT engine [XSLT 1999] for transformation of XML documents. We
started with basic facilities for creation, editing, storing and querying XML documents and
developed further this inventory towards a powerful system for processing not only single
XML documents but an integrated set of documents and constraints over them. The main
goal of this development is to allow the user to add the desirable semantics to the XML
documents. The XPath language is used extensively to direct the processing of the document
pointing where to apply a certain tool. It is also used to check whether some conditions are
present in a set of documents.
Tokenization
The CLaRK System supports a user-defined hierarchy of tokenizers. At the very basic
level the user can define a tokenizer in terms of a set of token types. In this basic tokenizer
each token type is defined by a set of UNICODE symbols. Above this basic level tokenizers the
user can define other tokenizers for which the token types are defined as regular expressions
over the tokens of some other tokenizer, the so called parent tokenizer. For each tokenizer an
alphabetical order over the token types is defined. This order is used for operations like the
comparison between two tokens, sorting and similar.
Regular Grammars
The regular grammars in CLaRK System [Simov, Kouylekov and Simov 2002] work over
token and element values generated from the content of an XML document and they incor-
porate their results back in the document as XML mark-up. The tokens are determined by
1 For the latest version of the system see http://www.bultreebank.org/clark/index.html.
the corresponding tokenizer. The element values are defined with the help of XPath expres-
sions, which determine the important information for each element. In the grammars, the
token and element values are described by token and element descriptions. These descriptions
could contain wildcard symbols and variables. The variables are shared among the token de-
scriptions within a regular expression and can be used for the treatment of phenomena like
agreement. The grammars are applied in cascaded manner. The evaluation of the regular ex-
pressions, which define the rules, can be guided by the user. We allow the following strategies
for evaluation: ‘longest match’, ‘shortest match’ and several backtracking strategies.
Constraints over XML Documents
The constraints that we have implemented in the CLaRK System are generally based on
the XPath language (see [Simov, Simov and Kouylekov 2003]). We use XPath expressions
to determine some data within one or several XML documents and thus we evaluate some
predicates over the data. Generally, there are two modes of using a constraint. In the first
mode the constraint is used for validity check, similar to the validity check, which is based
on a DTD or an XML schema. In the second mode, the constraint is used to support the
change of the document to satisfy the constraint. The constraints in the CLaRK System are
defined in the following way: (Selector, Condition, Event, Action), where the selector
defines to which node(s) in the document the constraint is applicable; the condition defines
the state of the document when the constraint is applied. The condition is stated as an XPath
expression, which is evaluated with respect to each node, selected by the selector. If the result
from the evaluation is improved, then the constraint is applied; the event defines when this
constraint is checked for application. Such events can be: selection of a menu item, pressing
of key shortcut, an editing command; the action defines the way of the actual constraint
application.
Cascaded Processing
The central idea behind the CLaRK System is that every XML document can be seen as a
“blackboard” on which different tools write some information, reorder it or delete it. The user
can arrange the applications of the different tools to achieve the required processing. This
possibility is called cascaded processing. For more on application construction abilities of
CLaRK System see [Simov, Simov and Osenova 2004].
7 Results and outlook
Here we report on the result from the Bulgarian–English QA track. From all 200 questions
the correct answers were extracted for 26 questions. 168 answers were wrong, 5 inexact and
1 unsupported. The distribution of the correct answers among the question categories is as
follows: 5 definition questions: 2 for organizations and 3 for persons; 21 factoid questions:
5 for locations, 2 for manner, 1 for measure, 2 for objects, 1 for organizations, 2 for other
categories, 4 for persons, and 4 for time. The main problem that caused the wrong answer
extraction was the degree of the ambiguity in the translation from Bulgarian to English.
Interestingly, the percentage of the ambiguities for nouns has bigger impact on the results
that the ambiguity of verbs. Another problem is that our semantic dictionary does not have a
mapping to the English WordNet synsets which is a crucial information for DIOGENE System
for answer extraction.
Our plans for future work are in two directions. First, we plan to implement a complete
question answering system for Bulgarian. With respect to the Bulgarian–English task we
envisage: to extend the dictionaries, to map our semantic dictionary (at least the top part)
to the WordNet synsets and to implement an efficient translation disambiguation module.
References
[Balabanova and Ivanova, 2002] Elisaveta Balabanova and Krassimira Ivanova. 2002. Creat-
ing a machine-readable version of Bulgarian valence dictionary: (A case study of CLaRK
system application). In: Proc. of The First Workshop on Treebanks and Linguistic Theories.
Sozopol, Bulgaria.
[Ivanova and Dojkoff 2002] Krassimira Ivanova and Dimitar Doikoff. 2002. Cascaded Regular
Grammars and Constraints over Morphologically Annotated Data for Ambiguity Resolution.
In: Proc. of The Workshop on Treebanks and Linguistic Theories. Sozopol, Bulgaria.
[Magnini et al 2002] Magnini, B., Negri, M., Prevete, R., Tanev, H. 2002. Comparing Statis-
tical and Content-Based Techniques for Answer Validation on the Web. Proceedings of the
VIII Convegno AI*IA, Siena, Italy.
[Negri et al 2002] Negri M., Tanev H., and Magnini B. 2003. Bridging Languages for Question
Answering: DIOGENE at CLEF-2003. Proceedings of CLEF-2003, Trondheim, Norway.
[Osenova 2002] Petya Osenova. 2002. Bulgarian Nominal Chunks and Mapping Strategies for
Deeper Syntactic Analyses. In: Proc. of The Workshop on Treebanks and Linguistic Theo-
ries. Sozopol, Bulgaria.
[Osenova and Kolkovska 2002] Petya Osenova and Sia Kolkovska. 2002. Combining the
named-entity recognition task and NP chunking strategy for robust pre-processing. In: Proc.
of The Workshop on Treebanks and Linguistic Theories. Sozopol, Bulgaria.
[Osenova and Simov 2002] Petya Osenova and Kiril Simov. 2002. Learning a token classifica-
tion from a large corpus. (A case study in abbreviations). In: Proc. of the ESSLLI Workshop
on Machine Learning Approaches in Computational Linguistics, Trento, Italy.
[Pianta et al 2002] Pianta, E., Bentivogli, L., Girardi, C. 2002. MULTIWORDNET: Develop-
ing an Aligned Multilingual Database. Proceedings of the 1st International Global WordNet
Conference, Mysore, India.
[Popov, Simov and Vidinska 1998] Dimitar Popov, Kiril Simov and Svetlomira Vidinska.
1998. A Dictionary of Writing, Pronunciation and Punctuation of Bulgarian Language.
Atlantis LK, Sofia, Bulgaria.
[Simov et. al. 2001] Kiril Simov, Zdravko Peev, Milen Kouylekov, Alexander Simov, Marin
Dimitrov, Atanas Kiryakov. 2001. CLaRK - an XML-based System for Corpora Develop-
ment. In: Proc. of the Corpus Linguistics 2001 Conference. pp 558–560.
[Simov and Osenova, 2001] Kiril Simov and Petya Osenova. 2001. A Hybrid System for Mor-
phoSyntactic Disambiguation in Bulgarian. In: Proc. of the RANLP 2001, Tzigov chark,
Bulgaria.
[Simov, Popova and Osenova 2002] Kiril Simov, Gergana Popova, Petya Osenova. 2002.
HPSG-based syntactic treebank of Bulgarian (BulTreeBank). In: “A Rainbow of Corpora:
Corpus Linguistics and the Languages of the World”, edited by Andrew Wilson, Paul
Rayson, and Tony McEnery; Lincom-Europa, Munich, pp. 135-142.
[Simov, Kouylekov and Simov 2002] Kiril Simov, Milen Kouylekov, Alexander Simov. 2002.
Cascaded Regular Grammars over XML Documents. In: Proc. of the 2nd Workshop on
NLP and XML (NLPXML-2002), Taipei, Taiwan.
[Simov, Simov and Kouylekov 2003] Kiril Simov, Alexander Simov, Milen Kouylekov. 2003.
Constraints for Corpora Development and Validation. In: Proc. of the Corpus Linguistics
2003 Conference, pages: 698–705.
[Simov, Simov and Osenova 2004] Kiril Simov, Alexander Simov, Petya Osenova. 2004. An
XML Architecture for Shallow and Deep Processing. In: Proc. of the ESSLLI 2004 Workshop
on Combining Shallow and Deep Processing for NLP. Nancy, France. pages: 51–60.
[Slavcheva 2002] Milena Slavcheva. 2002. Segmentation Layers in the Group of the Predicate:
a Case Study of Bulgarian within the BulTreeBank Framework. In: Proc. of The Workshop
on Treebanks and Linguistic Theories. Sozopol, Bulgaria.
[XML 2000] XML. 2000. Extensible Markup Language (XML) 1.0 (Second Edition). W3C
Recommendation. http://www.w3.org/TR/REC-xml
[XPath 1999] XPath. 1999. XML Path Language (XPath) version 1.0. W3C Recommenda-
tion. http://www.w3.org/TR/xpath
[XSLT 1999] XSLT. 1999. XSL Transformations (XSLT). version 1.0. W3C Recommenda-
tion. http://www.w3.org/TR/xslt