Publications

The chapter surveys applicative constructions in the languages of the Northwest Caucasian family.

Recent research has reported that standard fine-tuning approaches can be unstable due to being prone to various sources of randomness, including but not limited to weight initialization, training data order, and hardware. Such brittleness can lead to different evaluation results, prediction confidences, and generalization inconsistency of the same models independently fine-tuned under the same experimental setup. Our paper explores this problem in natural language inference, a common task in benchmarking practices, and extends the ongoing research to the multilingual setting. We propose six novel textual entailment and broad-coverage diagnostic datasets for French, German, and Swedish. Our key findings are that the mBERT model demonstrates fine-tuning instability for categories that involve lexical semantics, logic, and predicate-argument structure and struggles to learn monotonicity, negation, numeracy, and symmetry. We also observe that using extra training data only in English can enhance the generalization performance and fine-tuning stability, which we attribute to the cross-lingual transfer capabilities. However, the ratio of particular features in the additional training data might rather hurt the performance for model instances. We are publicly releasing the datasets, hoping to foster the diagnostic investigation of LMs in a cross-lingual scenario, particularly in terms of benchmarking, which might promote a more holistic understanding of multilingualism in LMs and cross-lingual knowledge transfer.  

P(reposition)-stranding is typologically rare. Nevertheless, many languages exhibit phenomena that look like P-stranding (Campos 1991, Poplack, Zentz, and Dion 2012) or involve P-stranding under common theorizing (see Philippova 2014 and references therein). These studies argue that these are not instances of P-complement movement and provide alternative analyses. This squib addresses Russian prepositions that can be postposed to and apparently stranded by their dependents. They are proposed to be PPs rather than P-heads, with dative dependents adjoined similarly to external possessors. The analysis captures all idiosyncrasies of their nominal dependents and alleviates the need to posit exceptional P-stranding in Russian.

To avoid post-neurosurgical language deficits, intraoperative mapping of the language function in the brain can be complemented with preoperative mapping with functional magnetic resonance imaging (fMRI). The validity of an fMRI “language localizer” paradigm crucially depends on the choice of an optimal language task and baseline condition. This study presents a new fMRI “language localizer” in Russian using overt sentence completion, a task that comprehensively engages the language function by involving both production and comprehension at the word and sentence level. The paradigm was validated in 18 neurologically healthy volunteers who participated in two scanning sessions, for estimating test–retest reliability. For the first time, two baseline conditions for the sentence completion task were compared. At the group level, the paradigm significantly activated both anterior and posterior language-related regions. Individual-level analysis showed that activation was elicited most consistently in the inferior frontal regions, followed by posterior temporal regions and the angular gyrus. Test–retest reliability of activation location, as measured by Dice coefficients, was moderate and thus comparable to previous studies. Test–retest reliability was higher in the frontal than temporo-parietal region and with the most liberal statistical thresholding compared to two more conservative thresholding methods. Lateralization indices were expectedly left-hemispheric, with greater lateralization in the frontal than temporo-parietal region, and showed moderate test-retest reliability. Finally, the pseudoword baseline elicited more extensive and more reliable activation, although the syllable baseline appears more feasible for future clinical use. Overall, the study demonstrated the validity and reliability of the sentence completion task for mapping the language function in the brain. The paradigm needs further validation in a clinical sample of neurosurgical patients. Additionally, the study contributes to general evidence on test–retest reliability of fMRI.

Geopolitical interventions since the end of the 1980s—such as the collapse of the Soviet Union, a decline in the activities of state-owned coal companies, and governmental initiatives to increase tourism activities—have affected the community viability of two main settlements on Svalbard: Barentsburg and Longyearbyen. This paper explores how the residents of these settlements (with different cultural backgrounds) perceive the effects of socioeconomic transitions on community viability. The analysis of qualitative interviews with residents of Barentsburg (n = 62) and Longyearbyen (n = 36) reveals the residents’ perceptions of the pace of the transition and the changing community composition. New types of commercial activities, such as tourism, contribute to local value creation and socioeconomic development but come with concerns grounded in community fluctuation, environmental protection, economic prioritisation, and power relationships. Compared to Longyearbyen, Barentsburg has undergone relatively minor demographic and social changes and remains stable in terms of culture, language, and management practices. We conclude that the viability of Longyearbyen and Barentsburg during the transition was affected by community dynamics and fluctuations, social relationships within and between communities, and local institutional practices.

In this paper, we address the issue of reliability of quantitative data on multilingualism of the past obtained as recall data. More specifically, we investigate whether the interviewees’ assessments of the language repertoires of their late relatives (indirect data) provide results that are quantitatively similar to those obtained from the people of the same age range themselves (direct data). The empirical data we use come from an ongoing field study of traditional multilingualism in Daghestan (Russia). We trained machine learning models to see whether they can detect differences in indirect and direct data. We conclude that our indirect quantitative data on L2 other than Russian are essentially similar to direct, while there may be a small but systematic underestimation when reporting other’s knowledge of Russian.

This paper focuses on the noun phrase in Tanti Dargwa (East Caucasian) and presents evidence for the distinction between modifiers proper (adjectival phrases, participial relative clauses and non-genitive adnominal NPs) and determiner-like elements (demonstratives, indefinite pronouns, numerals and most quantity expressions) in this language. Crucially, this dichotomy, which presumably reflects the distinction between the determinative and descriptive components in the NP, is realized in Tanti Dargwa mostly morphologically – in the distribution of “attributive markers” and in the expression of number. Syntactically, in the most neutral constructions the order of elements other than the head is virtually free and does not display any scope-related effects, while the head occupies the final position. In addition, Tanti Dargwa shows marginal constructions (a right-periphery construction locating a modifier after the head and a construction showing quasi-incorporation of a modifier into the noun) which are restricted to modifiers. Tanti Dargwa data support the idea that the description/determination distinction is gradual rather than discrete, as there are elements that show behavior intermediate between modifiers proper and determiner-like elements: possessor NPs, contrastive modifiers and the expressions like ‘other’.

In this paper, I consider double causatives in Mehweb, a one village language spoken in Daghestan, Russia, and belonging to the Dargwa branch of East Caucasian. The capability of stacking two causative suffixes seems to be lexically restricted, and mapping onto verbal meanings that are typically P-labile in the languages of the family. Interestingly, the verbs allowing double causatives are not morphosyntactically labile in Mehweb, which is generally poor in labile verbs as compared to sister languages. I conclude that the ability to form double causatives is not a consequence of the morphosyntactic property of being labile; rather, both morphosyntactic properties follow from the same component of the lexical semantics of these verbs and ultimately from the properties of the situational concepts they convey. As a tentative functional explanation I suggest that the relevant property is the weakened status of the agentive participant.

We study the correlation between phylogenetic and geographic distances for the languages of the Andic branch of the East Caucasian (Nakh-Daghestanian) language family. For several alternative phylogenies, we find that geographic distances correlate with linguistic divergence. Notably, qualitative classifications show a better fit with geography than cognacy-based phylogenies. We interpret this result as follows: The better fit may be due to implicit geographic bias in qualitative classifications. We conclude that approaches to classification other than those based on cognacy run a risk to implicitly include geography and geography-related factors as one basis of genealogical classifications.

The study presents the first systematic comparison of the global reading processes via scanpath analysis in Russian-speaking children with and without reading difficulties. First, we compared basic eye-movement characteristics in reading sentences in two groups of children in grades 1 to 5 (N = 72 in high risk of developmental dyslexia group and N = 72 in the control group). Next, using the scanpath method, we investigated which global reading processes these children adopt to read the entire sentence and how these processes differ between the groups. Finally, we were interested in the timeframe of the change in the global reading processes from the 1st to the 5th grades for both groups. We found that the main difference in word-level measures between groups was the reading speed reflected in fixation durations. However, the examination of the five identified global reading processes revealed qualitative similarities in reading patterns between groups. Children in the control group progressed quickly and by the 4th grade engaged in an adult-like fluent reading process. The high-risk group started with the beginner reading process, then similar to first graders in the control group, engaged mostly in the intermediate and upper-intermediate reading processes in 2nd to 4th grades. They reach the advanced process in the 5th grade, the same pattern preferred by the control group second graders. Overall, the scanpath analysis reveals that although there are quantitative differences in the word-level eye-tracking measures between groups, qualitatively children in the high-risk group read on par with typically developing peers but with a 3-year reading delay.

Recent studies report that autoregressive language models can successfully solve many NLP tasks via zero- and few-shot learning paradigms, which opens up new possibilities for using the pre-trained language models. This paper introduces two autoregressive GPT-like models with 1.3 billion and 13 billion parameters trained on 60 languages from 25 language families using Wikipedia and Colossal Clean Crawled Corpus. We reproduce the GPT-3 architecture using GPT-2 sources and the sparse attention mechanism; Deepspeed and Megatron frameworks allow us to parallelize the training and inference steps effectively. The resulting models show performance on par with the recently released XGLM models by Facebook, covering more languages and enhancing NLP possibilities for low resource languages of CIS countries and Russian small nations. We detail the motivation for the choices of the architecture design, thoroughly describe the data preparation pipeline, and train five small versions of the model to choose the most optimal multilingual tokenization strategy. We measure the model perplexity in all covered languages and evaluate it on the wide spectre of multilingual tasks, including classification, generative, sequence labeling and knowledge probing. The models were evaluated with the zero-shot and few-shot methods. Furthermore, we compared the classification tasks with the state-of-the-art multilingual model XGLM. source code and the mGPT XL model are publicly released.

Unlike stroke, neurosurgical removal of left-hemisphere gliomas acts upon a reorganized language network and involves brain areas rarely damaged by stroke. We addressed whether this causes the profiles of neurosurgeryand stroke-induced language impairments to be distinct. K-means clustering of language assessment data (neurosurgery cohort: N = 88, stroke cohort: N = 95) identified similar profiles in both cohorts. But critically, a cluster of individuals with specific phonological deficits was only evident in the stroke but not in the neurosurgery cohort. Thus, phonological deficits are less clearly distinguished from other language deficits after glioma surgery compared to stroke. Furthermore, the correlations between language production and comprehension scores at different linguistic levels were more extensive in the neurosurgery than in the stroke cohort. Our findings suggest that neurosurgery-induced language impairments do not correspond to those caused by stroke, but rather manifest as a ‘moderate global aphasia’ – a generalized decline of language processing abilities.

We outlined in chapter 1 the goals of this survey of number across a diverse sample of languages: investigate the properties of number systems in some depth, while at the same time guaranteeing direct comparability between the analyses of systems that can be very different from each other. At the conclusion of this survey, it is appropriate to consider what picture emerges from it. We will structure our answer in three steps. First, in section 2, we briefly review the main results arising from the chapters in this volume, and we organize the typology of number values that emerge from them. Most of these considerations broadly confirm what is generally known (or assumed) about the expression, content, and variation space of number systems across natural languages. The next sections, from 3 to 7, discuss in greater depth and in a more analytical perspective some important themes, especially focusing on issues that can shed new light or contribute to current understanding of number systems. Finally, section 8 offers a wrap-up discussion and points to desiderata that arise from these studies.

The volume is devoted to the typology of the category of number in the world's languages.

The chapter provides a detailed description of the expression of number in West Circassian.

The paper describes expressions with the meaning ‘other’ in East Caucasian (Nakh-Daghestanian) languages. It is shown that four main strategies can be distinguished: i) the ‘one’-based strategy: ‘other’ includes the numeral ‘one’; ii) the demonstrative-based strategy: ‘other’ includes a demonstrative pronoun; iii) the mixed demonstrative-based + ‘one’-based strategy: ‘other’ includes both a demonstrative and the numeral ‘one’; and iv) the lexical strategy: ‘other’ is a dedicated adjective (pronoun), not necessarily derived from any other clearly discernable source.

Our study investigates clickbait – ambiguous misleading headlines, aimed at exploiting readers’ curiosity gap. Clickbait has two unequally conspicuous interpretations, where the first interpretation is false with respect to the article content, while the second interpretation is correct content-wise, but considerably less obvious. We aim at identifying linguistic structures and pragmatic mechanisms employed in creating Russian clickbait, as well as testing the effectiveness of clickbait headlines experimentally. To answer our research questions, we compiled, annotated and analyzed a corpus of Russian clickbait headlines, as well as designed and ran an online experiment. We demonstrate that the main device employed in creating clickbait is a false generalized conversational implicature (GCI), created by flouting Gricean maxims of Quantity, Manner, or Relevance, which is subsequently canceled by the content of the article. We identify language-specific contexts conducive to their occurrence, and show how journalists explore semantic, syntactic, and referential peculiarities of the Russian language to create GCIs inherent in clickbait. Finally, we demonstrate that not all clickbait headlines are effective: in order to attract clicks, clickbait must not be obviously false and must increase the potential appeal and importance of insufficiently engaging news pieces on topics that do not require specialized knowledge.  

Despite recent achievements in predicting personality traits and some other human psychological features with digital traces, prediction of subjective well-being (SWB) appears to be a relatively new task with few solutions. COVID-19 pandemic has added both a stronger need for rapid SWB screening and new opportunities for it, with online mental health applications gaining popularity and accumulating large and diverse user data. Nevertheless, the few existing works so far have aimed at predicting SWB, and have done so only in terms of Diener’s Satisfaction with Life Scale. None of them analyzes the scale developed by the World Health Organization, known as WHO-5 – a widely accepted tool for screening mental well-being and, specifically, for depression risk detection. Moreover, existing research is limited to English-speaking populations, and tend to use text, network and app usage types of data separately. In the current work, we cover these gaps by predicting both mentioned SWB scales on a sample of Russian mental health app users who represent a population with high risk of mental health problems. In doing so, we employ a unique combination of phone application usage data with private messaging and networking digital traces from VKontakte, the most popular social media platform in Russia. As a result, we predict Diener’s SWB scale with the state-of-the-art quality, introduce the first predictive models for WHO-5, with similar quality, and reach high accuracy in the prediction of clinically meaningful classes of the latter scale. Moreover, our feature analysis sheds light on the interrelated nature of the two studied scales: they are both characterized by negative sentiment expressed in text messages and by phone application usage in the morning hours, confirming some previous findings on subjective well-being manifestations. At the same time, SWB measured by Diener’s scale is reflected mostly in lexical features referring to social and affective interactions, while mental well-being is characterized by objective features that reflect physiological functioning, circadian rhythms and somatic conditions, thus saliently demonstrating the underlying theoretical differences between the two scales.

Enhanced Universal Dependencies (EUD) are enhanced graphs expressed on top of basic dependency trees. EUD support repre- sentation of deeper syntactic relations in constructions such as coordi- nation, gapping, relative clauses, and argument sharing through control and raising. The paper presents experiments on the EUD parsing of the low-resource Belarusian language, for which no corpora with enhanced annotations were available.

Models trained on the Universal Dependencies treebanks of two closely related Slavic languages, Russian and Ukrainian, were used to parse sen- tences translated from Belarusian. After that, EUD were projected to the original sentences, which gave us ELAS (Enhanced Labeled Attach- ment Score) 78.1% for both Russian and Ukrainian in evaluation. We also trained a model of one of the IWPT 2020 Shared Task participants on obtained the annotations in Belarusian and achieved ELAS 83.4%. The analysis shows that the most common mistakes of cross-lingual parsing are rooted in different theoretical perspectives and practice approaches to the annotation of particular types of clauses in the three Slavic treebanks. Russian and Ukrainian EUD transfer models tend to make mistakes when dealing with the predicate argument relations, which are hard to iden- tify without understanding the semantics of the sentence. The alignment method decreases the quality of the annotation by confusing tokens that occur in a sentence more than once.

In the literature on nominal tautologies, it is assumed that common knowledge is a crucial ingredient for their interpretation. This paper explores a different approach: we argue that invoking shared knowledge is at the same time too strict and too vague as a condition for the understanding of tautologies in context. More specifically, we claim that, on the one hand, the hearer’s previous knowledge about some specific set of properties of the entity referred to in the tautology is not always necessary: lack of previous knowledge can be repaired by accommodating new assumptions or compensated by providing additional explicit content in discourse. On the other hand, the hearer’s previous knowledge about some specific set of properties of the entity referred to in the tautology is not always sufficient: only permanent, classificatory properties can be evoked by a tautology; transitory states, by contrast, are systematically rejected, even if they constitute shared knowledge and are supported by the context. We provide evidence for our claims both from the corpus study, analysing examples of tautologies with proper names from COCA and web-based sources, and experimental study designed as a verification task, additionally measuring reaction times for replying to a given question.