Skip to main content

Marko Tadić

University of Zagreb, Faculty of Humanities and Social Sciences [Filozofski fakultet], Department of Linguistics, Faculty Member

Followers

83

Following

44

Co-authors

35

Public Views

Marko TADIĆ, linguist, professor at the University of Zagreb, Faculty of Humanities and Social Sciences, Dept. of Linguistics and associated member of the Croatian Academy of Sciences and Arts. He published 80 papers and 5 books. He is the co-author of the Croatian frequency dictionary (1999) and a number of important Croatian language resources: Croatian National Corpus, Croatian Morphological Lexicon, Croatian Dependency Treebank, Croatian Wordnet. His interests are in corpus and computational linguistics, language technologies and research infrastructures in humanities and social sciences.

less

University of Zagreb

Uppsala University

University of East London

University of Leicester

Gwen Robbins Schug

University of North Carolina at Greensboro

Gabriel Gutierrez-Alonso

University of Salamanca

Macquarie University

Universidade Federal do Rio Grande do Sul

Swansea University

Jesper Hoffmeyer

University of Copenhagen

Uploads

Papers by Marko Tadić

CroSentiNews 2.0: A Sentence-Level News Sentiment Corpus

arXiv (Cornell University), May 14, 2023

This article presents a sentence-level sentiment dataset for the Croatian news domain. In additio... more This article presents a sentence-level sentiment dataset for the Croatian news domain. In addition to the 3K annotated texts already present, our dataset contains 14.5K annotated sentence occurrences that have been tagged with 5 classes. We provide baseline scores in addition to the annotation process and inter-annotator agreement.

Building the Croatian-English Parallel Corpus

Language Resources and Evaluation, May 1, 2000

Language Report Croatian

Cognitive technologies, 2023

This chapter presents a summary of the Language Report on Croatian (Tadić 2022) on general featur... more This chapter presents a summary of the Language Report on Croatian (Tadić 2022) on general features of the language and the level of technological support it receives since the previous report ). The chapter includes information about the typological and structural features of Croatian, its status and usage in the digital sphere and its support through Language Technologies. The Croatian language belongs to the West-South Slavic subgroup of the Balto-Slavic branch of the Indo-European linguistic family. It is the native language of over 5 million speakers. Croatian consists of the dialects and standard national language of the Croats, and is the official language of just under 4 million people in Croatia. Along with Bosnian and Serbian, it is one of the three official languages in Bosnia and Herzegovina, where it is spoken by about 400,000 people. Croatian is also spoken by national minorities in Croatia as well as by autochthonous Croatian minorities in Serbia, Montenegro, Slovenia, Hungary, Austria, Slovakia and Italy. Croatian is also used abroad. The largest Croatian diaspora is located in Germany, followed by the US, Canada and Australia. In 2013 Croatian became the 24th official EU language. According to the 2011 census, 90.42% of the country's inhabitants are ethnic Croats, with Croatian the native language for 95.6%. Croatian is the main language used and taught in schools. The literacy ratio in Croatia is 99.2%. Croatian was written with three scripts (Glagolitic, Cyrillic, Latin), and the Latin script became dominant in the 16th century. It was standardised after 1835, when the Croatian Latin alphabet settled on its modern-day form. The phoneme inventory of the Croatian standard language consists of 6 vowels and 25 consonants. Croatian differentiates ten parts of speech, five of which inflect (nouns, adjectives, numbers (partially), pronouns, verbs) and four do not (prepositions, conjunctions, particles, exclamations), while some adverbs inflect only in com-

RECSA: Resource for Evaluating Cross-lingual Semantic Annotation

by Daša Berović, Marko Tadić, and M. Srebačić

Hrvatske glagolske imenice i njihove istovrijednice u talijanskome jeziku – korpusna analiza

by Maslina Ljubičić and Marko Tadić

National Language Technology Platform (NLTP): The Final Stage

The final stage and the demo of the National Language Technology Platform (NLTP) developed within... more The final stage and the demo of the National Language Technology Platform (NLTP) developed within the CEF action of the same name is presented in this paper. The action aims at combining the most advanced language technology tools and solutions in a new state-of-the-art, artificial-intelligencedriven, web-based national platform for language technology oriented primarily towards users from public administrations of partner states. The Platform combines into a single framework the CAT tools, the TMs usage and management, the terminology management, several different MT engines and other language technology modules.

Croatian Film Review Dataset (Cro-FiReDa): A Sentiment Annotated Dataset of Film Reviews

Proceedings of the 9th Workshop on Slavic Natural Language Processing 2023 (SlavicNLP 2023)

This paper introduces Cro-FiReDa, a sentimentannotated dataset for Croatian in the domain of movi... more This paper introduces Cro-FiReDa, a sentimentannotated dataset for Croatian in the domain of movie reviews. The dataset, which contains over 10,000 sentences, has been annotated at the sentence level. In addition to presenting the overall annotation process, we also present benchmark results based on the transformerbased fine-tuning approach.

Corpus-based Syntactic Typological Methods for Dependency Parsing Improvement

This article presents a comparative analysis of four different syntactic typological approaches a... more This article presents a comparative analysis of four different syntactic typological approaches applied to 20 different languages to determine the most effective one to be used for the improvement of dependency parsing results via corpora combination. We evaluated these strategies by calculating the correlation between the language distances and the empirical LAS results obtained when languages were combined in pairs. From the results, it was possible to observe that the best method is based on the extraction of word order patterns which happen inside subtrees of the syntactic structure of the sentences.

Corpus-based Syntactic Typological Methods for Dependency Parsing Improvement

Zenodo (CERN European Organization for Nuclear Research), May 2, 2023

This article presents a comparative analysis of four different syntactic typological approaches a... more This article presents a comparative analysis of four different syntactic typological approaches applied to 20 different languages to determine the most effective one to be used for the improvement of dependency parsing results via corpora combination. We evaluated these strategies by calculating the correlation between the language distances and the empirical LAS results obtained when languages were combined in pairs. From the results, it was possible to observe that the best method is based on the extraction of word order patterns which happen inside subtrees of the syntactic structure of the sentences.

Analysis of Corpus-based Word-Order Typological Methods

Zenodo (CERN European Organization for Nuclear Research), Dec 1, 2022

This article presents a comparative analysis of four different syntactic typological approaches a... more This article presents a comparative analysis of four different syntactic typological approaches applied to 20 different languages. We compared three specific quantitative methods, using parallel CoNLL-U corpora, to the classification obtained via syntactic features provided by a typological database (lang2vec). First, we analyzed the Marsagram linear approach which consists of extracting the frequency word-order patterns regarding the position of components inside syntactic nodes. The second approach considers the relative position of heads and dependents, and the third is based simply on the relative position of verbs and objects. From the results, it was possible to observe that each method provides different language clusters which can be compared to the classic genealogical classification (the lang2vec and the head and dependent methods being the closest). As different word-order phenomena are considered in these specific typological strategies, each one provides a different angle of analysis to be applied according to the precise needs of the researchers.

Multilingual Comparative Analysis of Deep-Learning Dependency Parsing Results Using Parallel Corpora

Zenodo (CERN European Organization for Nuclear Research), Jun 1, 2022

This article presents a comparative analysis of dependency parsing results for a set of 16 langua... more This article presents a comparative analysis of dependency parsing results for a set of 16 languages, coming from a large variety of linguistic families and genera, whose parallel corpora were used to train a deep-learning tool. Results are analyzed in comparison to an innovative way of classifying languages concerning the head directionality parameter used to perform a quantitative syntactic typological classification of languages. It has been shown that, despite using parallel corpora, there is a large discrepancy in terms of LAS results. The obtained results show that this heterogeneity is mainly due to differences in the syntactic structure of the selected languages, where Indo-European ones, especially Romance languages, have the best scores. It has been observed that the differences in the size of the representation of each language in the language model used by the deep-learning tool also play a major role in the dependency parsing efficacy. Other factors, such as the number of dependency parsing labels may also have an influence on results with more complex labeling systems such as the Polish language.

Proceedings of the Seventh International Conference Formal Approaches to South Slavic and Balkan languages

The article examines the argument structures of a group of intransitive verbs of non-directed mov... more The article examines the argument structures of a group of intransitive verbs of non-directed movement and their formal description in the Bulgarian FrameNet (Koeva, 2008; Koeva & Dekova, 2008). The verbs to be discussed include the Bulgarian verbs блуждая 'roam', бродя 'rove', лутам се 'wander', скитам 'ramble,' and шляя се 'stroll/loaf about'. They all share a common meaning of non-directed movement and they also display corresponding argument structures. We argue that the proper formal description of the semantic features of a verb is crucial in predicting the verb's syntactic behavior. This is particularly evident when describing a group of semantically related verbs. 1 The financial support is granted under Contract No. BG051PO001-3.3.04/27 of 28 August 2009 within the Operation Support to the development of PhD students, post-doctoral students, postgraduate students and young scientists of the General Directorate Structural Funds and International Educational Programmes with the Ministry of Education, Youth and Science. 2 The authors would like to thank the anonymous reviewer of the paper, whose remarks helped us to improve this article. on both semantic and syntactic levels. Each argument structure consists of two arguments, that is, an external argument (the subject) and an internal argument (a PP with a locative function). 2.1 Syntactic-semantic specification of the external argument The external argument can be expressed by a noun (or an NP) which denotes an animated object, a person or an animal. An interesting finding is that the first argument does not have to be the conscious doer of the action represented by the examined verbs. This refers especially to arguments expressed by nouns denoting an animal, as animals are known to act not consciously but according to their instincts, which is also evident in the corpus data (see the examples in (1) below). (1) a) Според последното преброяване в "Мазалат" бродят 150 диви свине, 120 сърни, 100 елена и 30 мечки (BulNC); 'According to the last census 150 wild-boars, 120 doe, 100 deer, and 30 bears rove in "Mazalat." b) Тук може да се прекарат незабравими дни и през лятото, когато планината е по-достъпна и из нея човек може да скита със седмици (BulNC); 'Unforgettable days can be spent here, when the mountain is more accessible and one can ramble there for weeks. c) А между тях се лутат рой врани, орли, псета. (BulNC); 'A flight of crows, eagles, dogs wander among them. d) Разбира се, сега кучетата се шляят по целия остров…(BulNC). 'Of course, now the dogs loaf about on the entire island… It is important to mention that the corpus data provided evidence that human subjects may also be regarded as acting unconsciously, as seen in the example in (2) below. (2) Хората блуждаеха по улиците безмълвни и унесени като зомбита. (BulNC); 'The people roamed the streets silent and rapt as zombies.' The presence of the comparative phrase като зомбита ('as zombies') does prove the possible lack of conscious control over the action performed by the external argument of the sentence in (2), that is Хората ('The people'). On the syntactic level, the subject does not have to be explicitly mentioned. This is one of the fundamental characteristics of Bulgarian as a Balkan language and therefore we do not intend to discuss it any further in this paper. 2.2 Syntactic-semantic specification of the internal argument The actions denoted by the verbs блуждая 'roam', бродя 'rove', лутам се 'wander', скитам 'ramble,' and шляя се 'stroll/loaf about' imply also a location where the specified movement takes place. Hence, we can expect to have a syntactic component with a locative function expressed overtly whenever these verbs are used. This participant (the internal argument or the complement) is semantically related to a comparatively large area. That is, the actions denoted by the verbs at hand usually cannot be executed in a small region, as for example 2 m 2. Instead, they all entail a non-directed movement in larger space, 4 as illustrated by the phrases in italics in the examples in (1) and (2) above. Therefore, whenever present the internal argument inevitably denotes some kind of a space, a PP with a locative function on the syntactic level. The PP is introduced by one of the following spacial prepositions: из 'through', по 'on', в 'in', сред 'among', между 'between', край 'along', покрай 'along', около 'around', зад 'behind', пред 'in front of'. However, only the first three prepositions are discussed presently, being the most interesting ones on the syntax-semantic interface. used without the internal argument, while the verb скитам 'ramble' and most of all the verb бродя 'rove' allow for this syntactic drop out only in cases of iteration or habitual action 5. (3) a) Струваше си да блуждая две хиляди години, за да науча това (BulNC) 'It was worth roaming for two thousand years to learn this' b) … един слуга ми каза да не се шляя вечерта, … (BulNC) '… a servant told me not to stroll about in the evening, …' c) Лутал се безпризорен 13 дни. (BulNC) '(He) has wandered homeless for 13 days.' d) Хубаво беше само това, дето сега не скиташе много нощем. (BulNC) 'The only good thing was that now he didn't ramble much at nights.' e) Най-често бродел неуморно пешком, като се подпирал на жезъл … (BulNC) 'Most often (he) roved tirelessly on foot, resting against a scepter …' 5 This syntactic change triggered by the meaning of the verb whenever a semantic component of iteration or habitual action is present has received wide attention both in English and Bulgarian literature. Therefore, we do not feel it is necessary to discuss it in this paper. 6 Cf. Koeva & Dekova (2008) for a detailed explanation on the formal descriton used in the Bulgarian FramenNet. 7 However, по 'on' bears no association to any boundaries as the preposition из 'through' does. (GCBL, 1983:436) The locative argument of the verb пребродя 'traverse / go through' is often elucidated by the words всичко 'all' or цялото 'the whole', which essentially repeat the meaning introduced by the prefix пре-, as shown in the examples in (7). (7) a) Вече си мислех, че ще трябва да пребродя цялото имение, за да те открия (BulNC); 'I've already thought that I'd have to cross the estate all over to find you .' b) Бях готов да пребродя …, всички улици, … (BulNC); 'I was ready to travel across all streets.' c) Като преброди така цялата мера, той отново се върна по същия път … (BulNC); 'When (he) traveled like this across the whole common land, he returned back on the same road …' d) Дори да пребродите целия този обширен континент, … (BulNC); 'Even if you travel across this whole vast continent, ...' The corpus data show that the locative argument in the syntactic realization of the verb избродя 'travel all over' also includes the semantics of всичко 'all' or цялото 'the whole', as demonstrated in the examples in (8) below. (8) a) … и изброди цялата околия от край до край … (BulNC); '… and (he/she) crossed the whole region from one end to another …' b) Сега ми се налагаше да живея в тази империалистическа държава, да я избродя от край до край, … (BulNC); 'Now I wished that I lived in this imperialistic country, to cross it from one end to another …'.' c) Не се и съмнявам, …, че може да избродите всичките мотели в страната, … (BulNC); 'I don't even doubt it, …, that (you) can go through all the motels in the country, ….'

The Optimization of Portuguese Named-Entity Recognition and Classification by Combining Local Grammars and Conditional Random Fields Trained with a Parsed Corpus

Springer eBooks, 2021

This article presents the results of the study concerning namedentity recognition and classificat... more This article presents the results of the study concerning namedentity recognition and classification for Portuguese focusing on temporal expressions. We have used Conditional Random Fields (CRF) probabilistic method and features coming from automatically annotated parsed corpus and local grammars. We were able to notice that Part-of-Speech (PoS) tags are the most relevant information coming from a parsed corpus to be used as feature for this task. No positive synergy emerges from the association of these tags with other linguistic information from the parsed corpus. NooJ local grammar, created to recognize "Time" category entities (without detailing types and subtypes) provides information that surpasses PoS tags as a feature for CRF training in terms of precision and recall. The combination of PoS and NooJ annotations does not bring any advantage.

Why do we Need Croatian Wordnet

Racunalno pohranjeni rjecnici i semanticke baze podataka, odnosno semanticke mreže predstavljaju ... more Racunalno pohranjeni rjecnici i semanticke baze podataka, odnosno semanticke mreže predstavljaju temelj racunalne analize leksickoga znacenja. Takve mreže za hrvatski uopce jos nisu razvijane metodologijom uporabljenom kod vecine europskih jezika, a rijec je o WordNetu. Rad prikazuje preliminarne postupke pri izradbi hrvatskog Wordneta u Zavodu za lingvistiku Filozofskog fakulteta u Zagrebu na primjeru ontologije za uskodefinirano podrucje i obrade znacenjskih odnosa među njezinim elementima.

Arabic Treebank: from Phrase-Structure Trees to Dependency Trees

Language Resources and Evaluation, May 1, 2012

The aim here is to create a dependency treebank from a phrase-structure treebank for Arabic. Arab... more The aim here is to create a dependency treebank from a phrase-structure treebank for Arabic. Arabic has a number of characteristics,described below, which make it particularly challenging to any natural language processing (NLP) applications. We describe an encouraging semi-automatic technique for converting phrase-structure trees to dependency trees by using a head percolation table.One of the most signi?cant challenges here is the determination of the head of each subtree. We therefore examined different versionsof the head percolation table to ?nd the best priority list for each entry in the table. Given that there is no absolute measure of the?correctness? of a conversion of a phrase structure tree to dependency form, we tested the various transformations by seeing how well astate-of-the-art dependency parser learnt the generalisations that were embodied by the converted trees.

Interacting Croatian NERC System and Intex/NooJ Environment

In this contribution, we present design and implementation details of an early version of Croatia... more In this contribution, we present design and implementation details of an early version of Croatian finite state transducer engine called NercFst. The engine currently implements a small subset of Intex/NooJ finite state transducer functionality developed for the purpose of deriving a standalone module for named entity recognition and classification (NERC) system applicable to Croatian texts, previously created as a module in Intex. We also provide some general notes on the Intex module for Croatian NERC and notes on porting the module from Intex to NooJ. Current NercFst engine functionality overview is given in more detail along with some upcoming export features for NooJ which are currently under development with a purpose of supporting portability to various other open source finite state transducer libraries by exporting systems designed and implemented within Intex or NooJ linguistic development environment.

Dani e-infrastrukture Srce DEI 2018 : izazovi (ne)otvorenosti : knjiga sažetaka

Dani e-infrastrukture - Srce DEI 2018: Izazovi (ne)otvorenosti, 2018

Identification of translational equivalents in Croatian-English Parallel Corpus

U radu se ispituju mogucnosti identificiranja prijevodnih ekvivalenta u Hrvatsko-engleskom parale... more U radu se ispituju mogucnosti identificiranja prijevodnih ekvivalenta u Hrvatsko-engleskom paralelnom korpusu sravnjenom (aligned) na razini recenica koji je sastavljen u Zavodu za lingvistiku Filozofskog fakulteta u Zagrebu. Pozornost je u pocetku usmjerena na identifikaciju prijevodnih ekvivalenata od po jedne rijeci u svakom jeziku. To je postignuto generiranjem svih mogucih prijevodnih parova rijeci unutar svake recenice sravnjene u odnosu 1:1. Na tako generirane parove primijenjen je izracun statisticke mjere uzajamne obavijesnosti (mutual information) koja daje podatak od statisticki relevantnim supojavljivanjima rijeci u takvu paru. Parovi s visokom vrijednoscu uzajamne obavijesnosti predstavljaju dobre kandidate za prijevodne ekvivalente. U nastavku se rada provodi identifikacija jedinica sastavljenih od vise rijeci (multi-word units) u izvornim (hrvatskim) tekstovima i to primjenom iste statisticke mjere. Identican se postupak rabi za ustanovljavanje relevantnog supojavljiv...

The European Language Technology Landscape in 2020: Language-Centric and Human-Centric AI for Cross-Cultural Communication in Multilingual Europe

Multilingualism is a cultural cornerstone of Europe and firmly anchored in the European treaties ... more Multilingualism is a cultural cornerstone of Europe and firmly anchored in the European treaties including full language equality. However, language barriers impacting business, cross-lingual and cross-cultural communication are still omnipresent. Language Technologies (LTs) are a powerful means to break down these barriers. While the last decade has seen various initiatives that created a multitude of approaches and technologies tailored to Europe's specific needs, there is still an immense level of fragmentation. At the same time, AI has become an increasingly important concept in the European Information and Communication Technology area. For a few years now, AI – including many opportunities, synergies but also misconceptions – has been overshadowing every other topic. We present an overview of the European LT landscape, describing funding programmes, activities, actions and challenges in the different countries with regard to LT, including the current state of play in indus...

European Research Infrastructures

What is exactly Research Infrastructure? How can language technologies serve as research infrastr... more

CroSentiNews 2.0: A Sentence-Level News Sentiment Corpus

arXiv (Cornell University), May 14, 2023

This article presents a sentence-level sentiment dataset for the Croatian news domain. In additio... more This article presents a sentence-level sentiment dataset for the Croatian news domain. In addition to the 3K annotated texts already present, our dataset contains 14.5K annotated sentence occurrences that have been tagged with 5 classes. We provide baseline scores in addition to the annotation process and inter-annotator agreement.

Building the Croatian-English Parallel Corpus

Language Resources and Evaluation, May 1, 2000

Language Report Croatian

Cognitive technologies, 2023

This chapter presents a summary of the Language Report on Croatian (Tadić 2022) on general featur... more This chapter presents a summary of the Language Report on Croatian (Tadić 2022) on general features of the language and the level of technological support it receives since the previous report ). The chapter includes information about the typological and structural features of Croatian, its status and usage in the digital sphere and its support through Language Technologies. The Croatian language belongs to the West-South Slavic subgroup of the Balto-Slavic branch of the Indo-European linguistic family. It is the native language of over 5 million speakers. Croatian consists of the dialects and standard national language of the Croats, and is the official language of just under 4 million people in Croatia. Along with Bosnian and Serbian, it is one of the three official languages in Bosnia and Herzegovina, where it is spoken by about 400,000 people. Croatian is also spoken by national minorities in Croatia as well as by autochthonous Croatian minorities in Serbia, Montenegro, Slovenia, Hungary, Austria, Slovakia and Italy. Croatian is also used abroad. The largest Croatian diaspora is located in Germany, followed by the US, Canada and Australia. In 2013 Croatian became the 24th official EU language. According to the 2011 census, 90.42% of the country's inhabitants are ethnic Croats, with Croatian the native language for 95.6%. Croatian is the main language used and taught in schools. The literacy ratio in Croatia is 99.2%. Croatian was written with three scripts (Glagolitic, Cyrillic, Latin), and the Latin script became dominant in the 16th century. It was standardised after 1835, when the Croatian Latin alphabet settled on its modern-day form. The phoneme inventory of the Croatian standard language consists of 6 vowels and 25 consonants. Croatian differentiates ten parts of speech, five of which inflect (nouns, adjectives, numbers (partially), pronouns, verbs) and four do not (prepositions, conjunctions, particles, exclamations), while some adverbs inflect only in com-

RECSA: Resource for Evaluating Cross-lingual Semantic Annotation

by Daša Berović, Marko Tadić, and M. Srebačić

Hrvatske glagolske imenice i njihove istovrijednice u talijanskome jeziku – korpusna analiza

by Maslina Ljubičić and Marko Tadić

National Language Technology Platform (NLTP): The Final Stage

The final stage and the demo of the National Language Technology Platform (NLTP) developed within... more The final stage and the demo of the National Language Technology Platform (NLTP) developed within the CEF action of the same name is presented in this paper. The action aims at combining the most advanced language technology tools and solutions in a new state-of-the-art, artificial-intelligencedriven, web-based national platform for language technology oriented primarily towards users from public administrations of partner states. The Platform combines into a single framework the CAT tools, the TMs usage and management, the terminology management, several different MT engines and other language technology modules.

Croatian Film Review Dataset (Cro-FiReDa): A Sentiment Annotated Dataset of Film Reviews

Proceedings of the 9th Workshop on Slavic Natural Language Processing 2023 (SlavicNLP 2023)

This paper introduces Cro-FiReDa, a sentimentannotated dataset for Croatian in the domain of movi... more This paper introduces Cro-FiReDa, a sentimentannotated dataset for Croatian in the domain of movie reviews. The dataset, which contains over 10,000 sentences, has been annotated at the sentence level. In addition to presenting the overall annotation process, we also present benchmark results based on the transformerbased fine-tuning approach.

Corpus-based Syntactic Typological Methods for Dependency Parsing Improvement

This article presents a comparative analysis of four different syntactic typological approaches a... more This article presents a comparative analysis of four different syntactic typological approaches applied to 20 different languages to determine the most effective one to be used for the improvement of dependency parsing results via corpora combination. We evaluated these strategies by calculating the correlation between the language distances and the empirical LAS results obtained when languages were combined in pairs. From the results, it was possible to observe that the best method is based on the extraction of word order patterns which happen inside subtrees of the syntactic structure of the sentences.

Corpus-based Syntactic Typological Methods for Dependency Parsing Improvement

Zenodo (CERN European Organization for Nuclear Research), May 2, 2023

This article presents a comparative analysis of four different syntactic typological approaches a... more This article presents a comparative analysis of four different syntactic typological approaches applied to 20 different languages to determine the most effective one to be used for the improvement of dependency parsing results via corpora combination. We evaluated these strategies by calculating the correlation between the language distances and the empirical LAS results obtained when languages were combined in pairs. From the results, it was possible to observe that the best method is based on the extraction of word order patterns which happen inside subtrees of the syntactic structure of the sentences.

Analysis of Corpus-based Word-Order Typological Methods

Zenodo (CERN European Organization for Nuclear Research), Dec 1, 2022

This article presents a comparative analysis of four different syntactic typological approaches a... more This article presents a comparative analysis of four different syntactic typological approaches applied to 20 different languages. We compared three specific quantitative methods, using parallel CoNLL-U corpora, to the classification obtained via syntactic features provided by a typological database (lang2vec). First, we analyzed the Marsagram linear approach which consists of extracting the frequency word-order patterns regarding the position of components inside syntactic nodes. The second approach considers the relative position of heads and dependents, and the third is based simply on the relative position of verbs and objects. From the results, it was possible to observe that each method provides different language clusters which can be compared to the classic genealogical classification (the lang2vec and the head and dependent methods being the closest). As different word-order phenomena are considered in these specific typological strategies, each one provides a different angle of analysis to be applied according to the precise needs of the researchers.

Multilingual Comparative Analysis of Deep-Learning Dependency Parsing Results Using Parallel Corpora

Zenodo (CERN European Organization for Nuclear Research), Jun 1, 2022

This article presents a comparative analysis of dependency parsing results for a set of 16 langua... more This article presents a comparative analysis of dependency parsing results for a set of 16 languages, coming from a large variety of linguistic families and genera, whose parallel corpora were used to train a deep-learning tool. Results are analyzed in comparison to an innovative way of classifying languages concerning the head directionality parameter used to perform a quantitative syntactic typological classification of languages. It has been shown that, despite using parallel corpora, there is a large discrepancy in terms of LAS results. The obtained results show that this heterogeneity is mainly due to differences in the syntactic structure of the selected languages, where Indo-European ones, especially Romance languages, have the best scores. It has been observed that the differences in the size of the representation of each language in the language model used by the deep-learning tool also play a major role in the dependency parsing efficacy. Other factors, such as the number of dependency parsing labels may also have an influence on results with more complex labeling systems such as the Polish language.

Proceedings of the Seventh International Conference Formal Approaches to South Slavic and Balkan languages

The article examines the argument structures of a group of intransitive verbs of non-directed mov... more The article examines the argument structures of a group of intransitive verbs of non-directed movement and their formal description in the Bulgarian FrameNet (Koeva, 2008; Koeva & Dekova, 2008). The verbs to be discussed include the Bulgarian verbs блуждая 'roam', бродя 'rove', лутам се 'wander', скитам 'ramble,' and шляя се 'stroll/loaf about'. They all share a common meaning of non-directed movement and they also display corresponding argument structures. We argue that the proper formal description of the semantic features of a verb is crucial in predicting the verb's syntactic behavior. This is particularly evident when describing a group of semantically related verbs. 1 The financial support is granted under Contract No. BG051PO001-3.3.04/27 of 28 August 2009 within the Operation Support to the development of PhD students, post-doctoral students, postgraduate students and young scientists of the General Directorate Structural Funds and International Educational Programmes with the Ministry of Education, Youth and Science. 2 The authors would like to thank the anonymous reviewer of the paper, whose remarks helped us to improve this article. on both semantic and syntactic levels. Each argument structure consists of two arguments, that is, an external argument (the subject) and an internal argument (a PP with a locative function). 2.1 Syntactic-semantic specification of the external argument The external argument can be expressed by a noun (or an NP) which denotes an animated object, a person or an animal. An interesting finding is that the first argument does not have to be the conscious doer of the action represented by the examined verbs. This refers especially to arguments expressed by nouns denoting an animal, as animals are known to act not consciously but according to their instincts, which is also evident in the corpus data (see the examples in (1) below). (1) a) Според последното преброяване в "Мазалат" бродят 150 диви свине, 120 сърни, 100 елена и 30 мечки (BulNC); 'According to the last census 150 wild-boars, 120 doe, 100 deer, and 30 bears rove in "Mazalat." b) Тук може да се прекарат незабравими дни и през лятото, когато планината е по-достъпна и из нея човек може да скита със седмици (BulNC); 'Unforgettable days can be spent here, when the mountain is more accessible and one can ramble there for weeks. c) А между тях се лутат рой врани, орли, псета. (BulNC); 'A flight of crows, eagles, dogs wander among them. d) Разбира се, сега кучетата се шляят по целия остров…(BulNC). 'Of course, now the dogs loaf about on the entire island… It is important to mention that the corpus data provided evidence that human subjects may also be regarded as acting unconsciously, as seen in the example in (2) below. (2) Хората блуждаеха по улиците безмълвни и унесени като зомбита. (BulNC); 'The people roamed the streets silent and rapt as zombies.' The presence of the comparative phrase като зомбита ('as zombies') does prove the possible lack of conscious control over the action performed by the external argument of the sentence in (2), that is Хората ('The people'). On the syntactic level, the subject does not have to be explicitly mentioned. This is one of the fundamental characteristics of Bulgarian as a Balkan language and therefore we do not intend to discuss it any further in this paper. 2.2 Syntactic-semantic specification of the internal argument The actions denoted by the verbs блуждая 'roam', бродя 'rove', лутам се 'wander', скитам 'ramble,' and шляя се 'stroll/loaf about' imply also a location where the specified movement takes place. Hence, we can expect to have a syntactic component with a locative function expressed overtly whenever these verbs are used. This participant (the internal argument or the complement) is semantically related to a comparatively large area. That is, the actions denoted by the verbs at hand usually cannot be executed in a small region, as for example 2 m 2. Instead, they all entail a non-directed movement in larger space, 4 as illustrated by the phrases in italics in the examples in (1) and (2) above. Therefore, whenever present the internal argument inevitably denotes some kind of a space, a PP with a locative function on the syntactic level. The PP is introduced by one of the following spacial prepositions: из 'through', по 'on', в 'in', сред 'among', между 'between', край 'along', покрай 'along', около 'around', зад 'behind', пред 'in front of'. However, only the first three prepositions are discussed presently, being the most interesting ones on the syntax-semantic interface. used without the internal argument, while the verb скитам 'ramble' and most of all the verb бродя 'rove' allow for this syntactic drop out only in cases of iteration or habitual action 5. (3) a) Струваше си да блуждая две хиляди години, за да науча това (BulNC) 'It was worth roaming for two thousand years to learn this' b) … един слуга ми каза да не се шляя вечерта, … (BulNC) '… a servant told me not to stroll about in the evening, …' c) Лутал се безпризорен 13 дни. (BulNC) '(He) has wandered homeless for 13 days.' d) Хубаво беше само това, дето сега не скиташе много нощем. (BulNC) 'The only good thing was that now he didn't ramble much at nights.' e) Най-често бродел неуморно пешком, като се подпирал на жезъл … (BulNC) 'Most often (he) roved tirelessly on foot, resting against a scepter …' 5 This syntactic change triggered by the meaning of the verb whenever a semantic component of iteration or habitual action is present has received wide attention both in English and Bulgarian literature. Therefore, we do not feel it is necessary to discuss it in this paper. 6 Cf. Koeva & Dekova (2008) for a detailed explanation on the formal descriton used in the Bulgarian FramenNet. 7 However, по 'on' bears no association to any boundaries as the preposition из 'through' does. (GCBL, 1983:436) The locative argument of the verb пребродя 'traverse / go through' is often elucidated by the words всичко 'all' or цялото 'the whole', which essentially repeat the meaning introduced by the prefix пре-, as shown in the examples in (7). (7) a) Вече си мислех, че ще трябва да пребродя цялото имение, за да те открия (BulNC); 'I've already thought that I'd have to cross the estate all over to find you .' b) Бях готов да пребродя …, всички улици, … (BulNC); 'I was ready to travel across all streets.' c) Като преброди така цялата мера, той отново се върна по същия път … (BulNC); 'When (he) traveled like this across the whole common land, he returned back on the same road …' d) Дори да пребродите целия този обширен континент, … (BulNC); 'Even if you travel across this whole vast continent, ...' The corpus data show that the locative argument in the syntactic realization of the verb избродя 'travel all over' also includes the semantics of всичко 'all' or цялото 'the whole', as demonstrated in the examples in (8) below. (8) a) … и изброди цялата околия от край до край … (BulNC); '… and (he/she) crossed the whole region from one end to another …' b) Сега ми се налагаше да живея в тази империалистическа държава, да я избродя от край до край, … (BulNC); 'Now I wished that I lived in this imperialistic country, to cross it from one end to another …'.' c) Не се и съмнявам, …, че може да избродите всичките мотели в страната, … (BulNC); 'I don't even doubt it, …, that (you) can go through all the motels in the country, ….'

The Optimization of Portuguese Named-Entity Recognition and Classification by Combining Local Grammars and Conditional Random Fields Trained with a Parsed Corpus

Springer eBooks, 2021

This article presents the results of the study concerning namedentity recognition and classificat... more This article presents the results of the study concerning namedentity recognition and classification for Portuguese focusing on temporal expressions. We have used Conditional Random Fields (CRF) probabilistic method and features coming from automatically annotated parsed corpus and local grammars. We were able to notice that Part-of-Speech (PoS) tags are the most relevant information coming from a parsed corpus to be used as feature for this task. No positive synergy emerges from the association of these tags with other linguistic information from the parsed corpus. NooJ local grammar, created to recognize "Time" category entities (without detailing types and subtypes) provides information that surpasses PoS tags as a feature for CRF training in terms of precision and recall. The combination of PoS and NooJ annotations does not bring any advantage.

Why do we Need Croatian Wordnet

Racunalno pohranjeni rjecnici i semanticke baze podataka, odnosno semanticke mreže predstavljaju ... more Racunalno pohranjeni rjecnici i semanticke baze podataka, odnosno semanticke mreže predstavljaju temelj racunalne analize leksickoga znacenja. Takve mreže za hrvatski uopce jos nisu razvijane metodologijom uporabljenom kod vecine europskih jezika, a rijec je o WordNetu. Rad prikazuje preliminarne postupke pri izradbi hrvatskog Wordneta u Zavodu za lingvistiku Filozofskog fakulteta u Zagrebu na primjeru ontologije za uskodefinirano podrucje i obrade znacenjskih odnosa među njezinim elementima.

Arabic Treebank: from Phrase-Structure Trees to Dependency Trees

Language Resources and Evaluation, May 1, 2012

The aim here is to create a dependency treebank from a phrase-structure treebank for Arabic. Arab... more The aim here is to create a dependency treebank from a phrase-structure treebank for Arabic. Arabic has a number of characteristics,described below, which make it particularly challenging to any natural language processing (NLP) applications. We describe an encouraging semi-automatic technique for converting phrase-structure trees to dependency trees by using a head percolation table.One of the most signi?cant challenges here is the determination of the head of each subtree. We therefore examined different versionsof the head percolation table to ?nd the best priority list for each entry in the table. Given that there is no absolute measure of the?correctness? of a conversion of a phrase structure tree to dependency form, we tested the various transformations by seeing how well astate-of-the-art dependency parser learnt the generalisations that were embodied by the converted trees.

Interacting Croatian NERC System and Intex/NooJ Environment

In this contribution, we present design and implementation details of an early version of Croatia... more In this contribution, we present design and implementation details of an early version of Croatian finite state transducer engine called NercFst. The engine currently implements a small subset of Intex/NooJ finite state transducer functionality developed for the purpose of deriving a standalone module for named entity recognition and classification (NERC) system applicable to Croatian texts, previously created as a module in Intex. We also provide some general notes on the Intex module for Croatian NERC and notes on porting the module from Intex to NooJ. Current NercFst engine functionality overview is given in more detail along with some upcoming export features for NooJ which are currently under development with a purpose of supporting portability to various other open source finite state transducer libraries by exporting systems designed and implemented within Intex or NooJ linguistic development environment.

Dani e-infrastrukture Srce DEI 2018 : izazovi (ne)otvorenosti : knjiga sažetaka

Dani e-infrastrukture - Srce DEI 2018: Izazovi (ne)otvorenosti, 2018

Identification of translational equivalents in Croatian-English Parallel Corpus

U radu se ispituju mogucnosti identificiranja prijevodnih ekvivalenta u Hrvatsko-engleskom parale... more U radu se ispituju mogucnosti identificiranja prijevodnih ekvivalenta u Hrvatsko-engleskom paralelnom korpusu sravnjenom (aligned) na razini recenica koji je sastavljen u Zavodu za lingvistiku Filozofskog fakulteta u Zagrebu. Pozornost je u pocetku usmjerena na identifikaciju prijevodnih ekvivalenata od po jedne rijeci u svakom jeziku. To je postignuto generiranjem svih mogucih prijevodnih parova rijeci unutar svake recenice sravnjene u odnosu 1:1. Na tako generirane parove primijenjen je izracun statisticke mjere uzajamne obavijesnosti (mutual information) koja daje podatak od statisticki relevantnim supojavljivanjima rijeci u takvu paru. Parovi s visokom vrijednoscu uzajamne obavijesnosti predstavljaju dobre kandidate za prijevodne ekvivalente. U nastavku se rada provodi identifikacija jedinica sastavljenih od vise rijeci (multi-word units) u izvornim (hrvatskim) tekstovima i to primjenom iste statisticke mjere. Identican se postupak rabi za ustanovljavanje relevantnog supojavljiv...

The European Language Technology Landscape in 2020: Language-Centric and Human-Centric AI for Cross-Cultural Communication in Multilingual Europe

Multilingualism is a cultural cornerstone of Europe and firmly anchored in the European treaties ... more Multilingualism is a cultural cornerstone of Europe and firmly anchored in the European treaties including full language equality. However, language barriers impacting business, cross-lingual and cross-cultural communication are still omnipresent. Language Technologies (LTs) are a powerful means to break down these barriers. While the last decade has seen various initiatives that created a multitude of approaches and technologies tailored to Europe's specific needs, there is still an immense level of fragmentation. At the same time, AI has become an increasingly important concept in the European Information and Communication Technology area. For a few years now, AI – including many opportunities, synergies but also misconceptions – has been overshadowing every other topic. We present an overview of the European LT landscape, describing funding programmes, activities, actions and challenges in the different countries with regard to LT, including the current state of play in indus...

European Research Infrastructures

What is exactly Research Infrastructure? How can language technologies serve as research infrastr... more