Comparable corpora have long served as a foundation for multilingual NLP, supporting transfer across languages in tasks such as classification, retrieval, translation, and argument mining. Yet in the era of multilingual transformers and generative models, a central question is no longer simply whether texts are comparable, but what kinds of internal representations and downstream behaviors that comparability actually enables. In this keynote, I argue that cross-lingual transfer is best understood as a continuum oscillating between shared semantic structures and language-specific realizations. Drawing on two complementary studies, I demonstrate how this tension manifests both in the data models learn from and in the representations they develop. The first case study investigates multilingual stance and argument mining using the new Russian LoveHate corpus alongside English debate data. The results indicate that translated or multilingual resources are useful but insufficient proxies for language-specific corpora: local topics, culturally situated argumentation patterns, and stance expression still shape model performance and generalization. The second case study presents a neuron-level analysis of multilingual emotion detection, showing that multilingual encoders such as XLM-R develop both polyglot neurons, which respond consistently across languages, and monolingual neurons, which remain tied to particular linguistic systems. This reveals that even successful cross-lingual emotion transfer depends on only partial internal alignment. Together, these findings suggest that multilingual NLP needs corpora that preserve culturally specific meaning while supporting robust transfer, as well as interpretability frameworks that can diagnose where multilingual systems genuinely share representations and where they merely approximate them. Comparable corpora are not just training material; they are essential to understand how cross-lingual generalization succeeds, where it breaks down, and how truly multilingual NLP can move beyond English-centric assumptions and conclusions.

Idiopathic Parkinson’s disease is associated with motor speech impairments collectively referred to as hypokinetic dysarthria, which can appear at early disease stages and remain challenging to assess objectively in clinical practice. Most automatic assessment studies rely on individual speech corpora analyzed in isolation, leaving open questions regarding their comparability and their suitability for joint use within unified classification frameworks. This study explicitly investigates the cross-corpus comparability of existing Parkinsonian speech datasets designed for hypokinetic dysarthria assessment. Rather than assuming their compatibility, we evaluate it empirically through the generalization performance of classification systems trained on single or multiple corpora. We examine which datasets can be effectively combined and whether multi-corpus training improves robustness across heterogeneous recording conditions and speech tasks. Four corpora are evaluated under intra-corpus, cross-corpus, and out-of-domain settings. Results demonstrate that multi-corpus training enhances robustness and generalization performance, while also revealing substantial differences in cross-dataset compatibility. These findings provide a clearer understanding of the degree of comparability between existing resources and offer practical guidelines for the design of future corpora and more generalizable tools for the automatic clinical assessment of Parkinsonian speech.

Semi-parallel text refers to versions of the same text that have to some extent been edited by authors, translators, or others. They are of relevance especially in the social sciences and in literary genres. In this paper, we consider the bilingual (English/German) variant of the problem. The philosopher Hannah Arendt, for example, wrote political essays that often exist in multiple versions and in both languages. She repeatedly modified her texts, added or deleted parts, and framed topics differently for target audiences. For researchers to explore the history of such material in detail, and at the same time at scale, automatic alignment (i.e., finding the best match of semantically similar sentences) is a very valuable preprocessing step. In this paper, we compare the performances of a range of methods for this task, based on computing semantic similarity. We present the results and conduct a qualitative error analysis to identify recurring sources of error.

Parallel and comparable corpora are the main linguistic resources to identify multilingual terminology using automatic term extraction tools. However, parallel corpora are available only for certain languages, domains and genres, and comparable corpora have some limitations when identifying corresponding terms. To implement a more efficient selection of multilingual terminology, we compared the performance of using specialised parallel and comparable corpora applied to languages with various forms of capital in linguistic resources. This paper presents a comparative study in corpus linguistics in which we automatically identify terms in Catalan, Spanish and English in legislation and administrative law using parallel corpora, comparable corpora and a combined methodology based on both typologies of corpora together with word embeddings. We observe that the combined methodology implemented obtains a higher number of term candidates than when working exclusively with parallel or comparable corpora. The evaluation of the results is performed using a terminological thesaurus as a gold standard. The new methodology presented in our study permits us to identify multilingual terminology in an efficient way, especially in Catalan-Spanish languages.

The paper examines the use of comparable corpora for contrastive research on the category of nominal number across three languages—English, Czech, and Greek. Two objectives are pursued: a cross-linguistic analysis of number and an assessment of the impact of automatic annotation on linguistic findings. For this study, corpora of comparable size and composition were compiled for the three languages from the Leipzig Corpora Collection. The data were automatically annotated using two open-access tools, Stanza and UDPipe, producing six datasets (two per language), each containing about 5 million sentences and 100 million tokens. Although derived from the same source, the paired datasets for each language differ in sentence and word segmentation, in the number of nouns identified, and in the number values assigned. These differences, nevertheless, do not appear to substantially affect the overall picture of number in the languages examined. The distribution of lemmas by the ratio of singular and plural forms challenges the view commonly presented in grammars that most nouns occur in both numbers and that singular-only and plural-only nouns are rare. However, a closer analysis of nouns assumed to have defective number indicates that answers to more nuanced questions vary depending on the annotation tool used.

We present a multi-register (web, news, and government texts), diachronic (2015-2024), comparable corpus annotated for lexical gender-inclusive language (gil) features in German and Spanish. Apart from rule-based annotations, we train a transformer-based classifier to resolve semantically ambiguous neutral expressions like epicenes to reliably annotate true human referents. In a sample study, we analyze register variation in the three registers in terms of gil features both contrastively and diachronically. We show that gil usage increases and varies diachronically in terms of register in both languages. German texts show a higher overall frequency and diversity of gil features than Spanish texts. However, across languages, registers behave similarly, with government text showing the strongest usage of gil followed by news and web texts, and web texts showing the strongest innovation in terms of features. The results of our study are valuable to linguistic areas such as human and machine translation, SLA, and contribute to register-conform gender inclusive NLP downstream tasks such as machine translation, summarization or textgeneration. From a diachronic point of view, our corpus and analyses are a valuable contribution to observing language change in the making.

This study introduces a comparable corpus of Spanish digital news (2017–2026) designed to analyze potential linguistic shifts coinciding with the widespread adoption of Generative AI. We propose an analytical framework structured across three levels: lexical statistics, semantic topology, and neural classification. By implementing a protocol of NER-masking, we isolate structural discourse markers from topical content to identify the stylistic patterns of the contemporary period. Our results suggest a measurable structural shift within the analyzed corpus, indicating a trend toward a more standardized professional register. While macro-statistical metrics like Shannon entropy remain stable —indicating statistical consistency— Zipf-Mandelbrot distributions and SVD mapping reveal a concentration of unique vocabulary into more predictable clusters. In this scenario, the 2023–2026 subcorpus exhibits a discernible topological displacement compared to the 2017–2021 baseline. The study identifies a ‘Gray Zone’ where highly structured technical reporting and hybridized production become indistinguishable, suggesting a structural stylistic convergence within this digital environment. These findings provide a methodological baseline for analyzing discursive stabilization in professional domains without assuming definitive authorship.

Text simplification plays a crucial role in improving the accessibility and comprehensibility of written information for diverse audiences, including language learners and readers with limited literacy. Despite its importance, large-scale, high-quality datasets for training and evaluating text simplification models remain scarce for languages other than English. This paper reports an experimental study on the collection and processing of crowd-sourced simplification data to construct a corpus suitable for both training and testing text simplification systems across multiple languages (Catalan, English, French, Italian and Spanish). We report mechanisms for sentence-level alignment from document-level data. The resulting dataset of the aligned sentence pairs is publicly available.

Cross-dialect bi-text mining relies on robust multilingual sentence representations to identify semantically equivalent sentence pairs across languages. While recent multilingual bi-encoder models achieve strong performance on standardized written languages, their behavior on dialectal varieties is largely unknown. In this study, we use Tatoeba to evaluate the performance of four widely-used bi-encoders on dialect-to-standard German translation retrieval, covering German documents and queries written in three dialects: Low German, Bavarian, and Alemannic. Motivated by the lack of resources, we examine the extent to which synthetic translations (from dictionaries and large language models; LLMs) can serve as weak supervision for dialect adaptation. Our results reveal that bi-encoders, when applied in a zero-shot setting, exhibit deficiencies in capturing semantic similarity between German and dialects, while fine-tuning on synthetic data substantially improves their retrieval effectiveness, with larger gains obtained from LLM-translated training data. We further analyze retrieval performance on Bavarian across varying dialect word proportions and observe a drop when dialect words make up more than 60% of the text.

The growing ability of large language models (LLMs) to process long-range context opens new perspectives for document-level machine translation (MT), especially in scholarly communication. In fact, translating scholarly texts requires to integrate both local and long-range contextual information to ensure the consistency and coherence across the full document. However, document-level parallel corpora for such text types remain scarce, limiting both evaluation and domain adaptation of MT systems for this task. To address this gap, we introduce ParaEPS (Earth and Planetary Sciences Bilingual Corpus) and ParaNLP (Natural Language Processing Bilingual Corpus), two new parallel corpora covering 14k abstracts and 103 full-length articles in two scientific domains to be used for fine-tuning and evaluation purposes. We compare the performance of eight MT systems on these test sets and find that fine-tuning on document-level data closes the gap between open systems based on Large Language Models (LLMs) and commercial systems. We also find that the performance of recent LLMs can worsen when translating full articles instead of translating them on a per paragraph basisfine-tuning. These experiments underscore the need for corpora such as ParaEPS and ParaNLP.

Government-issued travel advisories are used by citizens to get information about destination countries for tourism and other purposes such as temporary work stays or permanent relocation plans. However, qualitative evidence suggests that travel advisories may be influenced by considerations beyond current security situations. Systematic and rigorous quantitative analyses of advisories are scarce because relevant corpus data are not readily available and official government websites often provide practical obstacles. We validate a pipeline to generate a time-series cross-sectional dataset of government-issued travel advisories for three English-speaking issuing countries based on the Internet Archive’s Wayback Machine. Using official government data sources that are prohibited to be scraped and used for research, we illustrate that our approach provides (near-)complete coverage. The resulting corpus and code are intended to support downstream research on comparative risk communication, international relations, and text analysis using natural language processing methods.

To mitigate the prevalence of toxic language on digital social media, various NLP approaches have been proposed for automatic text detoxification. However, the potential of toxic expression lexicons as a comparable cross-lingual resource to guide this process remains largely unexplored. In this work, we investigate how such resources can be effectively used to inform multilingual language models about what should and should not be considered toxic. We evaluate four models under two settings—zero-shot prompting and fine-tuning—to assess the impact of incorporating toxic expressions in prompt instruction, including in cross-lingual transfer scenarios. Our results show that both zero-shot prompting and fine-tuning approaches benefit considerably from adding toxic expressions in prompt instructions during training and/or inference. Our findings demonstrate that comparable, lightweight, language-specific toxic expression lexicons constitute an effective mechanism for injecting explicit information about lexical toxicity into multilingual language models.