Abstract This study explores whether L2-learning environment (i.e., ESL [English as a second language] vs. EFL [English as a foreign language]) and vowel type (i.e., L1-similar L2 vowels vs. L1-dissimilar L2 vowels) have an impact on L2 vowel acquisition with reference to the perception-production relationship. L1-Korean learners of English in the U.S. (ESL learners) and Korea (EFL learners) performed English vowel production, English vowel identification, and English-to-Korean vowel mapping tasks using bVt words (e.g., bit, but, bought ). The learners’ English vowel productions were assessed for intelligibility, their vowel identification accuracy was calculated, and their perceptual mapping patterns were used to measure L1-L2 vowel similarity. The results show that the ESL learners outperformed the EFL learners in both perception and production, indicating an effect of L2-learning environment on L2 vowel acquisition. However, both groups showed a significant correlation between perception and production. As for vowel type, the results indicate a significant correlation between perception and production for both L1-similar and L1-dissimilar L2 vowels for the ESL learners but with variations in accuracy rates among L1-dissimilar L2 vowels. For the EFL learners, a significant correlation was found between the two modalities only for L1-similar L2 vowels, while variations in accuracy rates were observed for both vowel types. The results suggest that close counterparts in the L1 may even lead to more successful L2 vowel acquisition relative to L2 vowels without such L1 analogues, especially for ESL learners. The results further suggest that the acquisition of different L2 vowels follows different trajectories, which is ascribed partly to the formation of L2 vowel categories.
Abstract This paper investigates the pronunciation of English borrowings in Polish, focussing on how orthography influences their phonological adaptation. Two complementary studies are presented. Study 1 examines the attested pronunciations of over 200 English borrowings in Polish extracted from Polish-language YouTube videos. Study 2 analyzes the output of two commercial Text-to-Speech systems (Google and Microsoft) when rendering 184 company names with English material embedded in Polish carrier sentences. The patterns revealed are as follows: (1) Some English vowels and consonants are consistently adapted via Polish grapheme-to-phoneme rules, especially KIT /ɪ/ and commA /ə/; (2) other vowels, such as TRAP and GOAT, show variability, alternating between phonological and graphemic routes; (3) most consonants, and some vowels, are rendered phonologically as Polish phonological equivalents on the basis of English grapheme-to-phoneme correspondences; (4) those English phonemes whose English grapheme-to-phoneme correspondences deviate most noticeably from Polish grapheme-to-phoneme correspondences are more likely to be adapted phonologically; (5) those whose donor correspondences are similar to recipient correspondences are more likely to follow the graphemic route. Notably, the Text-to-Speech output generally aligns very well with attested spoken data, suggesting that these systems reflect evolving community norms. The findings may indicate the emergence of endonormative patterns for English in Polish contexts.
Non-native language learners may experience difficulty acquiring processes that occur in naturalistic speech. A particularly challenging case is unstressed vowel reduction, both in production, for speakers whose L1 lacks reduction, and suppression, for learners whose L1 reduces unstressed vowels. The latter is the case for English learners of Spanish: vowel reduction is rife in English but mainly realised phonologically, while Spanish is traditionally considered to be largely free of reduction. The current study analyses Spanish vowel realisations in English learners' conversational speech, focusing on unstressed vowel reduction via shortening or centralisation. Vowels were extracted from the speech of ten British English and ten native Spanish speakers, elicited using a spot-the-difference picture task. While both Spanish and English speakers signalled absence of vowel stress for some vowels via durational shortening, only the English cohort exhibited centralisation of formant space. Counter-intuitively, unstressed vowel space area was equivalent for the two cohorts, and English speakers produced stressed vowels at more extreme locations than native speakers. A comparison of non-native Spanish vowels with native English productions from the same speakers suggests that L1 influence rather than hyper-articulation of stressed vowels accounts for English learners' realisations of Spanish stressed and unstressed vowels in naturalistic speech.
Abstract This paper examines Korean [SØV & SVO] structures and argues that traditional accounts of null objects – based on empty pronominals or argument ellipsis – do not fully capture the conditions under which such objects are licensed. These approaches leave important empirical and theoretical issues unresolved, particularly regarding the derivational mechanisms that determine the distribution of null objects. As an alternative, I propose that the missing object in [SØV & SVO] structures is derived via rightward Across-The-Board movement, a mechanism independently attested in clausal coordination, rather than through pronominalization or ellipsis. This movement-based approach provides a more unified and empirically robust account of [SØV & SVO] structures and offers precise insights into the syntactic principles that govern object licensing and rightward displacement.
Non-inherent adjectives have been defined as attributive-only adjectives that cannot characterize the noun they precede (Bolinger 1967. Adjectives in English: Attribution and predication. Lingua 18. 1-34: 1). These adjectives have been comprehensively analyzed in relation to nouns (Levi 1978. The syntax and semantics of complex nominals. London: Academic Press), but their relation with other word-classes, namely adverbs, remains unexplored. Corpus studies of non-inherent adjectives with adverbial meaning are scarce, and the evidence provided is limited to particular sets of adjectives (Ghesqui & egrave;re and Davidse 2011. The development of intensification scales in noun-intensifying uses of adjectives: Sources, paths and mechanisms of change. English Language and Linguistics 15(2). 251-277) or on these adjectives in specific registers (Pavl & iacute;& ccaron;kov & aacute; 2014. On the indirectness of non-inherent adjectives. In Olga Dontcheva-Navr & aacute;tilov & aacute; & Milada Walkov & aacute; (eds.), English matters IV: A collection of papers, 34-41. Pre & scaron;ov: Pre & scaron;ovsk & aacute; univerzita v Pre & scaron;ove). Thus, based on the analysis of 53,737 bigrams extracted by lemma from the British National Corpus and the Corpus of Contemporary American English, the present research provides evidence of non-inherent adjectives with adverbial meaning. The results present quantitative and qualitative data that demonstrate that non-inherent adjectives convey several adverbial meanings, including process, degree, or time location, among others. Qualitative data show the relevance of semantics in the ability of the adjective to develop non-inherent senses and the influence of other syntactic features of the noun phrase. The evidence provided adds to the description of non-inherent adjectives and shows that these lexical units could be relevant in a wider framework such as the adjective/adverb interface.
This research presents the development and evaluation of Convolutional Neural Network (CNN)-based models for Named Entity Linking (NEL) in Serbian to the Wikidata knowledge base. It introduces four novel model configurations derived from two underlying Named Entity Recognition (NER) systems: the 7-class SrpCNNER2 and a specialized SrpCNNER2-loc model for locations. Different training strategies for entity linking were implemented, including training on all entity types (srNEL-all), the top three most frequent types (srNEL-top3), locations only with NER (srNEL-loc), and a fully specialized location-only approach (srNEL-LOCfull). Models were trained on a corpus (73,493 sentences) comprising literary excerpts, legal documents, news articles, and synthetically generated sentences. The performance was evaluated on the sr-geography corpus derived from elementary school geography textbooks, using strict evaluation metrics and SrpCNNeL model as the baseline. The srNEL-all configuration, leveraging the comprehensive SrpCNNER2 NER base and trained on all entity types, achieves the highest performance on linking locations within the sr-geography corpus, attaining an F1 0.845, outperforming the SrpCNNeL baseline (F1 0.731) and models employing more specialized training regimes, such as srNEL-LOCfull (F1 0.411). The findings highlight the effectiveness of the proposed methods for Serbian NEL and underscore the benefits of incorporating broader entity context during training. While transformer-based dense-retrieval and RAG systems currently dominate multilingual entity linking research, this work prioritizes an efficient CNN-based architecture suitable for less-resource and morphologically rich languages, demonstrating competitive performance for Serbian. Although this system is faster and requires lower hardware requirements in terms of machine performance, the highest-performing model was compared with a transformer-based baseline, demonstrating its superiority over the CNN architecture. Future directions include further development of models for Serbian and exploration of hybrid architectures that combine efficiency with state-of-the-art performance.
Spoken language understanding (SLU) models are a core component of voice assistants (e.g., Alexa, Bixby, Google Assistant), but collecting extensive labeled data for target languages is challenging. In this paper, we introduce a data-centric pipeline to expand On-Device SLU to new languages by leveraging a large language model (LLM) for machine translation of slot-annotated English training data. The LLM is fine-tuned to preserve slot annotations during translation using an HTML tag-based slot marking strategy. Our approach is evaluated on the MultiATIS++ benchmark, a multilingual SLU dataset covering eight languages. In an On-device setting, we achieve a new state-of-the-art overall accuracy of 62.18 %, up from 55.11 % achieved by the best prior method, (HCL)-L-2. In an Edge scenario with a tiny SLU model (5MB, no pre-training), our translated data boosts overall accuracy from a baseline 5.31 % to 22.06 %. In contrast to mentioned baselines, our LLM-based translation requires no changes to the SLU model architecture and is slot-type independent, requiring no manual slot descriptions. This work demonstrates that state-of-the-art LLMs can serve as effective "slot translators" providing a scalable path to multilingual SLU without costly SLU data collection or architecture overhaul.
This study investigates affectedness in P-oblique constructions, understood as transitivity alternations with or without overt voice marking, in which the P-argument is demoted from a core to an oblique. The central question is to what extent such alternations are functionally motivated by reduced affectedness of the P-argument. While affectedness and individuation are widely recognized as key P-parameters of transitivity, their interaction has not been systematically examined cross-linguistically, leaving open the question of whether the well-attested correlation between higher individuation and greater affectedness extends to transitivity alternations. Beyond this issue, the study explores how verbal properties shape the affectedness in P-obliques, along with structural factors such as voice marking conditions. This perspective is relevant in view of claims that verb-uncoded P-obliques systematically correlate with lower affectedness, a view connected to the characterization of such constructions as conatives. Drawing on a genealogically balanced sample of 55 languages, 34 of which display P-oblique patterns, the study provides a broad comparative basis for clarifying the dynamics of affectedness and its relation to both P-related and V-related parameters of transitivity. The findings aim to refine typological accounts of parameter co-variation and to contribute to a better understanding of the functional motivations underlying transitivity alternations.
In this paper, we explored features of human-written (HW) and large language model-generated (LLMG) texts, specifically on a dataset of Polish student reports. Expanding upon the previous research with email datasets, we constructed a dataset containing 2,400 pairs of original student-written reports and their LLMG counterparts. Using a feature-based detection approach, we analyzed 220 distinct features, including perplexity, burstiness, error frequency, stylometric elements, and sentence-level characteristics. Our findings reveal that LLMG texts demonstrate lower perplexity and burstiness values, fewer errors, greater lexical diversity, and more consistent sentence lengths compared to HW texts. Among stylometric features, relative pronouns, stop words, and content word diversity emerged as notably discriminative. The preliminary classification results, achieved with the proposed detection model, confirm the viability of feature-based methods for reliably discriminating between HW and LLMG documents.
This study aims to provide a unified analysis of the chadian mei VP construction in Mandarin Chinese, which exhibits ambiguity between negative and positive polar readings. Unlike previous research that interprets this intriguing phenomenon from a desirability-based perspective, this study focuses on the distinctions between telicity and atelicity encoded in the VP structure. I propose that the disambiguation of this language-specific construction depends on the aspectual classification derived from the verbs and their complements. Specifically, when embedded in a telic situation, the chadian mei VP gives rise to an affirmative reading. In contrast, when the VP is atelic, the chadian mei VP semantically aligns with the chadian VP to denote negative polarity. Furthermore, shifts in situation type may occur due to interactions between other elements and the VP within the sentence, which accounts for the variability in the interpretation of the chadian mei VP construction at the sentential level.
This study presents the first steps undertaken in building a database of Croatian prepositions, a novel computational resource for Croatian aimed at describing relevant syntactic and semantic features of prepositions based on Croatian corpus data. Extant studies and approaches to Croatian prepositions, as well as other computational resources available, e.g., valency dictionaries, are examined with regard to their (dis)advantages in building the preposition database. Furthermore, this study focuses on proposing an annotation schema based on a preliminary investigation of 2000 corpus examples of 50 prepositions. It presents a preposition sense inventory built with the aim to a) unify sense descriptions currently dispersed across various case studies of Croatian prepositions, b) represent preposition senses in a manner that captures both specificities of prepositional polysemy and their (para)synonymy relations. Other language phenomena, such as complex constructions, are discussed as interacting with prepositional senses in various ways, thus illustrating the need for various levels of syntactic and semantic annotation of prepositional uses. Therefore, based on previous usage-based investigations of Croatian and the current corpus-based analysis of prepositions, the study proposes a roadmap for prepositional data annotation.
In this paper we provide a Nanosyntactic account of two verbalizing affixes in Dutch: the prefix ver- and the suffix -el. We show that they behave differently in their interaction with past tense and in their effect on argument structure. We derive these differences by assigning the two affixes a different lexical representation - a so-called L-tree - which in turn means that they are inserted into the syntactic derivation in very different ways. The analysis intends to be a first step towards a more general formalization of the systematic asymmetries between pre- and post-elements.
The Podlasie dialect of Polish exhibits a curious nominal case declension paradigm where the dative is expressed by a preposition and a genitive suffix (prep N-gen), which constitutes a typologically rare pattern. Moreover, the instrumental in this dialect, which is hierarchically above the dative, is marked with only a suffix (N-ins), unexpectedly dropping the preposition. The fact that a hierarchically higher case drops a preposition challenges a generalization made in Caha (2009. The Nanosyntax of case. CASTL/University of Troms & oslash; dissertation), which builds on Blake's (2001. Case, 2nd edn. Cambridge, UK: Cambridge University Press) hierarchy concerning the morphological marking of case. The paper explores how these two atypical patterns follow from the Nanosyntactic lexicalization procedure, offering an analysis that contributes to the discussion of 'pre' and 'post' marking in morphosyntax. The analysis also sheds light on why the pattern of the case declension attested in the Podlasie dialect is typologically rare.
This paper discusses the syntax of the ordinal superlative with a focus on Dutch. The ordinal superlative, for example the second highest mountain, is an under researched construction. Contrary to what has been hitherto believed, the ordinal superlative also exists in (non-standard) Dutch: de tweede hoogste berg. The ordinal and the superlative may occur in the opposite order as well, in which case they display a subordinated reading instead. I propose a syntactic analysis for the co-occurrence of ordinals with superlatives, with both modifiers in separate functional projections in the DP. The key is that the ordinal requires a superlative to be interpreted, but adjacency of the superlative to the ordinal does not suffice to force an ordinal superlative reading: an ordinal may be interpreted by context, in which case a covert superlative is merged. I show that a position internal to the superlative phrase cannot be proven with classic constituency tests, due to the DP-internal location of both modifiers.
In a great many languages something special is required for the expression of reflexivity. Cross-linguistically what is necessary varies from marking the object (e.g. pronoun + intensifier, a body-part noun, etc.), or using a special verb form (e.g. affix, special clitic, etc.). We discuss the limits of certain extant approaches to binding in accounting for this pattern. Going over the various cases we show that a key factor underlying this pattern is "local identity avoidance" If a DP were to bind a coargument straight away a logical form would result with two occurrences of the same variable in a local domain. Unless these can be kept apart, this would lead to a violation of a general principle of computation, not specific to language. We point towards possibly similar effects in other domains. We then discuss what underlies the variation in the ways languages express reflexivity. Universal is the need to find a strategy to avoid the consequences of our Inability to distinguish indistinguishables (IDI), but the specific choice will depend on the resources that are available in a particular language.
The paper contributes to a discussion of the syntactic expertise of ChatGPT. It focuses on its handling of the range of 'syntactic puzzles', such as 'onion sentences', 'garden paths', grammatical illusions and structural ambiguity. The parsing of such cases by humans dwells on the property of structure-dependence, which is not shared by the AI language generator - ChatGPT. The latter is believed to process sentences in a sequential, though not strictly linear fashion, not using either recursion or hierarchy, the two pillars of human structure dependence. The assessment of ChatGPT's syntactic expertise is based on two experiments in which the language model is asked to interpret and judge acceptability of English examples displaying such peculiar syntax. This is preceded by 'a training session' in which the model's understanding of the nuances of such cases is assessed. The outcome is then confronted with the results of two parallel questionnaire studies, with the same set of linguistic tasks, administered to a group of human respondents, untrained in linguistics, with a native-like proficiency in English.
Computational linguistic phylogenetics has so far relied heavily on cognate data. In contrast, the potential of morphosyntactic characters as a valuable source for phylogenetic analysis has been largely overlooked. We argue that morphosyntactic characters may conflate historical signal with the results of homoplasies, horizontal transfer, and universal tendencies, and must be scrutinized in terms of their propensity to change and borrowing, analogously to the curation of lexical data which produced the Swadesh lists. In this paper we make a start by evaluating a set of morphosyntactic characters based on the World Atlas of Language Structures using three methods: we (1) calculated Pearson correlation coefficients for each character against different language groupings, reflecting either shared ancestry (genera) or contact (geographical proximity); (2) counted the minimum number of mutations needed for the distribution of a character’s states on a cognate-based reference tree (parsimony score), testing whether they correctly reflect language change known from historical linguistics; and (3) ran a classic hill-climbing algorithm to determine which random subsets of characters produced a phylogeny closest to a reference tree. We conclude that these are useful tools, but expect that making the definitions of the characters more theoretically informed will produce a stronger historical signal.
In this paper we study the impact of augmenting spoken language corpora with domain-specific synthetic samples for the purpose of training a speech recognition system. Using both a conventional neural TTS system and a zero-shot one with voice cloning ability we generate speech corpora that vary in the number of voices. We compare speech recognition models trained with addition of different amounts of synthetic data generated using these two methods with a baseline model trained solely on voice recordings. We show that while the quality of voice-cloned dataset is lower, its increased multivoiceity makes it much more effective than the one with only a few voices synthesized with the use of a conventional neural TTS system. Furthermore, our experiments indicate that using low variability synthetic speech quickly leads to saturation in the quality of the ASR whereas high variability speech provides improvement even when increasing total amount of data used for training by 30
In this introduction we look at the evolving research area of Cognitive Translation and Interpreting Studies (CTIS) through the lens of articles which report on cutting edge research in a variety of multilectal mediated communicative events. Seven contributions in this issue expand the frame of Cognitive Translation and Interpreting Studies with detailed, empirically grounded accounts of how language and mind interact in mediated communication. The new empirical evidence challenges key concepts, such as cognitive effort or translation expertise, and sheds light on overlooked areas of translation reception and accessibility research. Although the studies vary in topic and methods, they converge on a more sophisticated view of cognition in multilectal mediated communication which underscores its complexity and dynamics. The findings point to two overarching trends in CTIS research: (1) conceptual progress and (2) methodological sophistication. The contributors to this special issue are all mid- and early-career researchers and the development of their research expertise mirrors that of CTIS as a community of practice committed to producing knowledge based on empirical evidence and enriched by meaningful collaborative exchanges with neighboring disciplines.
