Search

Transfer learning is one of the prevailing approaches towards training language-specific BERT models. However, some languages have uncommon features that may prove to be challenging to more domain-general models but not domain-specific models. Comparing the performance of BERTić, a Bosnian-Croatian-Montenegrin-Serbian model, and Multilingual BERT on a Named-Entity Recognition (NER) task and Masked Language Modelling (MLM) task based around a rare phenomenon of indeclinable female foreign names in Serbian reveals how the different training approaches impacts their performance. Multilingual BERT is shown to perform better than BERTić in the NER task, but BERTić greatly exceeds in the MLM task. Thus, there are applications both for domain-general training and domain-specific training depending on the tasks at hand.

This workshop is the fourth issue of a series of workshops on automatic extraction of sociopolitical events from news, organized by the Emerging Market Welfare Project, with the support of the Joint Research Centre of the European Commission and with contributions from many other prominent scholars in this field. The purpose of this series of workshops is to foster research and development of reliable, valid, robust, and practical solutions for automatically detecting descriptions of sociopolitical events, such as protests, riots, wars and armed conflicts, in text streams. This year workshop contributors make use of the state-of-the-art NLP technologies, such as Deep Learning, Word Embeddings and Transformers and cover a wide range of topics from text classification to news bias detection. Around 40 teams have registered and 15 teams contributed to three tasks that are i) multilingual protest news detection, ii) fine-grained classification of socio-political events, and iii) discovering Black Lives Matter protest events. The workshop also highlights two keynote and four invited talks about various aspects of creating event data sets and multi- and cross-lingual machine learning in few- and zero-shot settings., European Union (EU); Horizon 2020; European Research Council (ERC);Starting Grant; Emerging Welfare

Pattern matching, acceptance, and parsing algorithms on node-labeled, ordered, ranked trees ('tree algorithms') are important for applications such as instruction selection and tree transformation/term rewriting. Many such algorithms have been developed. They often are based on results from such algorithms on words or generalizations thereof using finite (tree) automata. Regrettably no coherent, extensive toolkit of such algorithms and automata existed, complicating their use. Our toolkit FOREST FIRE contains many such algorithms and automata constructions. It is accompanied by the graphical user interface (GUI) FIRE WOOD. The toolkit and GUI provide a useful environment for experimenting with and comparing the algorithms. In this tool paper we give an overview of the toolkit and GUI, their context and design rationale, and mention some results obtained with them.

We describe a simple IR approach for linking news about events, detected by an event extraction system, to messages from Twitter (tweets). In particular, we explore several methods for creating event-specific queries for Twitter and provide a quantitative and qualitative evaluation of the relevance and usefulness of the information obtained from the tweets. We showed that methods based on utilization of word co-occurrence clustering, domain-specific keywords and named entity recognition improve the performance with respect to a basic approach.

Supervised word sense disambiguation (WSD) has been shown to achieve state-of-the-art results but at high annotation costs. Active learning can ameliorate that problem by allowing the model to dynamically choose the most informative word contexts for manual labeling. In this paper we investigate the use of active learning for Croatian WSD. We adopt a lexical sample approach and compile a corresponding sense- annotated dataset on which we evaluate our models. We carry out a detailed investigation of the different active learning setups, and show that labeling as few as 100 instances suffices to reach near- optimal performance.

Being able to identify that different mentions refer to the same entity is beneficial for applications such as question answering and text summarization. In this paper, we propose the first model for entity coreference resolution for Croatian. We enforce transitivity constraints with integer linear programming on top of pairwise decisions produced by the supervised mention-pair model. Experimental results show that the proposed model significantly outperforms two different rule-based baselines, reaching performance of 74.4% MUC score and 77.6% B3 score.

We propose a method to specify, in a modular way, complex systems formed by interacting agents. The method is based on the notion of view, that is a partial representation of the system, reflecting one of its specific aspects. By composing the different views, we get the overall system, described as a special kind of transition system. By means of a suitable logical language, we can express interesting properties of the system; model-checking techniques can then be used to assess their validity. Views can be specified using different languages or notations, provided they can be translated in so-called agent aware transition systems. The method is explained with the help of a simple, but non trivial example