Abstract: Despite disputable possibility of extension of analysis of social relations on Twitter to real life, Twitter discussions are stiU being under attention of scholars studying structures and meanings of news- and issue-based ad-hoc public discourse. One of the socially relevant aspects of Twitter studies is that of influencers - accounts that produce impact, either inside or outside Twitter. But there is still no agreement in the research community on how to defme and measure who is an inDuencer: either by 'absolute figUres' or by network analysis metrics; this issue is even rarely discussed. Politically, today's mediatized pub6c sphere where traditional media play the role of information hubs is highly uneven in terms of auess to opinion expression; it privileges institutional players, including political elites, corporations, and media themselves. Hopes that Twitter would provide a more equal space for public deliberation are still not proven weD enough. Using web crawling and manual assessment of Twitter ad-hoc discussion on the Biryulyovo bashings of 2013, we show that users who post or even get commented most do not make it to the positions of most 'central' users by network metrics. We also demonstrate that usen that rank high by betweenness and pagerank centn.lity form circles of reciprocal commenting that show the social cleavage wider than the discussion itselt

Abstract: Many web-based applications such as advertising or recommender systems often critically depend on the demographic information, which may be unavailable for new or anonymous users. We study the problem of predicting demographic information based on user-generated texts on a Russian-language dataset from a large social network. We evaluate the efficiency of age prediction algorithms based on word2vec word embeddings and conduct a comprehensive experimental evaluation, comparing these algorithms with each other and with classical baseline approaches.

Abstract: We present the types or multiword expressions included into the thesaurus or Russian language RuThes. Maoy of these expressions may look like compositiomd expressions but have specific relations that can be useful in appllcatlons. The rela· tion system or the RuThes thesaurus allows natural description of relations between an expression and its components if necessary. Transforming the RnThes knowledge into the Princeton WordNet structure for creating Russian wordnet (RuWordNet), we tronsfer also all the described expressions into the new resource and propose to automatically introduce additional relations for their better representation.

Abstract: Recent advances in deep leaming for natural language processing achieve and improve over state of the art results in many natural language processing tasks. One problem with neural network models, however, is that they require large datasets, including large labeled datasets for the corresponding problems. In this work, we suggest a dala augmentation method based on extending a given dataset with synonyms for the words appearing there. We apply this approach to the morphologically rich Russian language and show improvements for modem neural network NLP models on standard tasks such as sentiment analysis.

Abstract: In this work, we address the hard clustering problem. We study how well clustering algorithm efficacy measures (clustering validity indices) cao rellect the clustering quality. We use assessors' estimations for cluster partition adequacy as the ground truth and explain, why tbis is the only measure that cao be used in tbis quality. We compare different clustering validity indices and show that none of them can be the universal, relleeting quality for each cluster partition. To do so, we introduce four quality measures for CVI evaluation. Also, we suggest an approach for the best CVI predietion for a given dataset based on meta-lesrning.

Abstract: In the process of globalization, the number of English words in other languages has rapidly increased. In automatic speech recognition systems, spell-checking, tagging, and other software in the field of natural language processing, loan words are not easily recognized and should be evaluated separately. In this paper we present a corpora-based approach to the automatic detection of anglicisms in Russian social network texts. Proposed method is based on the idea of simultaneous scripting, phonetics, and semantics similarity of the original Latin word and its Cyrillic analogue. We used a set of transliteration, phonetic transcribing, and morphological analysis methods to find possible hypotheses and distributional semantic models to filter them. Resulting list of borrowings, gathered from approximately 20 million LiveJoumal texts, shows good intenection with manually eoUe