http://hdl.handle.net/11356/1024 CLARIN.SI repository language resources and tools 2026-10-06T04:00:59Z http://hdl.handle.net/11356/2572 Monitor corpus of Slovene Trendi 2026-09 Kosem, Iztok; Čibej, Jaka; Dobrovoljc, Kaja; Erjavec, Tomaž; Ljubešić, Nikola; Ponikvar, Primož; Šinkec, Mihael; Krek, Simon The Trendi corpus is a monitor corpus of Slovenian. It contains news articles from 106 media websites, published by 64 publishers. Trendi 2026-09 covers the period from January 2019 to September 2026, complementing the Gigafida 2.2 reference corpus of written Slovene (http://hdl.handle.net/11356/2106). The contents of the Trendi corpus are obtained using the Jožef Stefan Institute Newsfeed service (http://newsfeed.ijs.si/). The texts have been annotated using the CLASSLA-Stanza pipeline (https://github.com/clarinsi/classla), including syntactic parsing according to the Universal Dependencies (https://universaldependencies.org/sl/) and Named Entities (https://nl.ijs.si/janes/wp-content/uploads/2017/09/SlovenianNER-eng-v1.1.pdf). An important addition are topics or thematical categories, which have been automatically assigned to each text. There are 13 categories altogether: Arts and culture, Crime and accidents, Economy, Environment, Health, Leisure, Politics and Law, Science and Technology, Society, Sports, Weather, Entertainment, and Education. The text classification uses the following models: Text classification model SloBERTa-Trendi-Topics 1.0 (http://hdl.handle.net/11356/1709), Text classification model fastText-Trendi-Topics 1.0 (http://hdl.handle.net/11356/1710), and the SloBERTa model (https://huggingface.co/cjvt/sloberta-trendi-topics). The corpus is currently not available as a downloadable dataset due to copyright restrictions but we hope to make at least some of it available in the near future. The corpus is accessible through CLARIN.SI concordancers. If you would like to use the dataset for research purposes, please contact Iztok Kosem (iztok.kosem@ijs.si). This version adds texts from September 2026. 2026-10-05T00:00:00Z http://hdl.handle.net/11356/2258 Slovenian Social Assistance Rights Terminology SSAR-TERM 1.0 Pollak, Senja; Jemec Tomazin, Mateja; Arhar Holdt, Špela; Thi Hong Tran, Hanh SSAR-TERM 1.0 is a manually validated Slovenian terminology resource for the domain of social assistance rights. The terms were automatically extracted from the SSAR 1.0 collection of legally binding and explanatory texts (http://hdl.handle.net/11356/1985) using a transformer-based term extraction model. The extracted term candidates were subsequently reviewed by a terminology specialist. Of the 1,000 extracted candidates, 983 were retained as valid terms and assigned to one of three domain categories: 1 – terms from the domain of social assistance rights (533) 2 – terms from closely related domains (358) 3 – terms from other domains (92) The columns provide automatically assigned part-of-speech (POS) labels, the extracted term form, the canonical (dictionary) form, which was manually verified and corrected where necessary, and the manually assigned domain label (1, 2, or 3; see above). The resource is intended primarily for research on terminology and terminography. If you use this resource, please cite: Senja POLLAK, Mateja JEMEC TOMAZIN, Eva PORI, Hanh Thi Hong TRAN, Špela ARHAR HOLDT (2026). Luščenje terminologije pravic socialnega varstva. In Proceedings of Language Technologies and Digital Humanities: Proceedings of the Conference. Ljubljana, September 17–18, 2026. 2026-09-15T00:00:00Z http://hdl.handle.net/11356/2484 Gail Valley dialect E2E Automatic Speech Recognition model Zila 1.0 Žgank, Andrej; Verdonik, Darinka; Boto, Sanja; Donaj, Gregor This resource provides automatic speech recognition (ASR) models for the Slovenian Gail Valley dialect, developed using the open-source WeNet 3.1 framework. The models were trained on 84.25 hours of speech from the Corpus of the Gail Valley dialect Zila 1.0 (http://hdl.handle.net/11356/2257). The system employs a hybrid CTC/attention architecture with a Conformer encoder and Transformer decoder. The log-Mel filterbank features were used for training. The final averaged model achieves a word error rate of 24.64% and a character error rate of 12.05%. Models are provided in WeNet and ONNX formats, supporting local, server-based, and GPU-accelerated inference and deployment. 2026-08-31T00:00:00Z http://hdl.handle.net/11356/2317 Multimodal document understanding dataset from "Ilustrirani Slovenec" Martinc, Matej; Dimitrievska, Mihaela This entry contains the Ilustrirani Slovenec Multimodal Document Understanding Dataset, a benchmark dataset designed for training and evaluating vision-language models on historical Slovenian newspaper pages. The dataset is derived from Ilustrirani Slovenec, a historical illustrated weekly supplement of the newspaper Slovenec, published between 1924 and 1932 and available through the Digital Library of Slovenia (dLib.si): https://www.dlib.si/details/URN:NBN:SI:spr-XWQWZFUW. A total of 410 newspaper issues were collected in PDF format and split into individual page images, resulting in 2,566 newspaper pages. Visual regions on the pages were automatically detected using Segment Anything Model 3 (SAM3; https://ai.meta.com/research/sam3/), using prompts targeting common visual elements found in historical newspapers, including photographs, portraits, illustrations, sketches, cartoons, maps, diagrams, advertisements, and posters. The resulting detections were post-processed to remove unsuitable, overlapping, and nested regions. In total, the detection pipeline identified 14,805 image regions. The detected image regions and the corresponding newspaper pages were subsequently processed using Gemini 3 Pro Preview (https://ai.google.dev/gemini-api/docs/models/gemini-3-pro-preview). For each page, the model generated structured annotations containing image-region bounding boxes, corresponding image captions where available, and page-level Slovenian text excluding image captions. The original Slovenian text was preserved without translation, normalization, or spelling correction. The generated annotations were transformed into six task-specific multimodal instruction-following datasets, each distributed with predefined training and test files: dataset1_train.jsonl and dataset1_test.jsonl contain the image region to caption task, where a bounding box corresponding to an image region is provided and the model must predict the associated image caption. dataset2_train.jsonl and dataset2_test.jsonl contain the caption to image region task, where an image caption is provided and the model must identify the corresponding image region by predicting its bounding-box coordinates. dataset3_train.jsonl and dataset3_test.jsonl contain the image region detection task, where the model must identify all image regions present on a newspaper page. dataset4_train.jsonl and dataset4_test.jsonl contain the image region detection and captioning task, where image regions must be detected and the corresponding caption generated for each detected region. dataset5_train.jsonl and dataset5_test.jsonl contain the page-level text extraction task, where all textual content on a newspaper page, including headlines, article text, and other textual elements, must be extracted while excluding image captions. dataset6_train.jsonl and dataset6_test.jsonl contain the full-page understanding task, requiring extraction of the complete structured content of a newspaper page, including page text, image-region locations, and corresponding image captions. Datasets 1 and 2 are constructed from individual image-caption pairs, while Datasets 3–6 are constructed at the page level. Of the 2,566 extracted newspaper pages, 2,556 were included in the final benchmark; 10 pages did not yield valid final annotations during the automatic processing pipeline. The benchmark is distributed with predefined training and test splits, comprising 2,300 training pages and 256 test pages. The same page-level partition is applied consistently across all six datasets so that no newspaper page appears in both the training and test partitions. Manual Quality Assessment To verify the quality of the automatically generated annotations, a subset of the test data was manually evaluated. Ten examples were randomly sampled from each of the six benchmark tasks, resulting in 60 evaluated benchmark instances. Each example was independently assessed by two annotators using two task-specific binary quality criteria, resulting in a total of 120 binary quality assessments. The two annotators agreed on 118 of the 120 assessments, corresponding to an observed agreement of 98.33%. Inter-annotator agreement was high, with Cohen's kappa of 0.8483 and Gwet's AC1 of 0.9813. Agreement was 100% for all evaluated criteria except joint localization and caption generation in Dataset 4 and structural hierarchy in Dataset 6, which both achieved 90% agreement. All manually reviewed examples that passed both task-specific quality criteria according to both annotators were additionally aggregated into manually_verified_minibench.jsonl. This file contains 53 manually verified examples across the six benchmark tasks and can be used as an additional high-confidence evaluation benchmark. Accessing the Corresponding Newspaper Pages The original Ilustrirani Slovenec PDF files and page images are not redistributed as part of this resource. The corresponding newspaper issues are available through the Digital Library of Slovenia (dLib.si): https://www.dlib.si/details/URN:NBN:SI:spr-XWQWZFUW. The Python script download_and_split_ilustrirani_slovenec.py is included for downloading the corresponding PDF issues and converting them into individual page images. The generated page images correspond to the image identifiers referenced in the six datasets and in the manually verified mini-benchmark. 2026-08-07T00:00:00Z http://hdl.handle.net/11356/2257 Corpus of the Gail Valley dialect Zila 1.0 Žgank, Andrej; Donaj, Gregor; Sereinig, Uši; Koren-Zwitter, Tatjana; Boto, Sanja; Zwitter-Grilc, Sabina; Kanzian, Martina; Hren, Karl; Vidinić, Jasna; Verdonik, Darinka The Zila Speech Corpus is a spoken language resource documenting the Slovenian Gail Valley dialect, an endangered Slovenian minority dialect spoken in Austria. The corpus contains 206 recordings in WAV format, with a total duration of approximately 100 hours and 7 minutes, produced by 74 anonymized speakers. It was created within the LINGUA project to support the development of automatic speech recognition for the Gail Valley dialect, as well as linguistic research, education, corpus search, and long-term dialect documentation and preservation. The corpus includes different types of speech material: private speech, public informal speech, public formal speech, and read speech. The largest part of the corpus consists of private recordings, while the remaining material documents public, institutional, and read speech contexts. The recordings cover topics such as autobiographical narratives, Gail Valley identity, folk tales, local customs and rituals, rural work and farming, family and village networks, local history, memories, and community life. The recordings in WAV format are accompanied by manually prepared, time-aligned transcriptions in Transcriber .trs format. The transcription follows an extended orthographic approach: dialectal morphological and syntactic features, hesitations, fillers, and self-corrections are preserved, while the transcription is not intended as a full phonetic transcription. Personal data in the transcripts are anonymized using the marker [anonim], and the corresponding sections in the audio are masked with a beep before publication. Beside the .trs format, the transcriptions in EXMARaLDA format (.exb, .exs) were also generated. Speaker metadata include anonymized speaker ID, gender, age, first language, bilingualism, dialect information, dialect retention, education, residence information, spontaneity of speech, and notes. Recording metadata include recording and transcription identifiers, language, title, date, source, recording location, genre, speech type, channel, number of speakers, recording device, recording setting, original format, audio quality, length, and notes. 2026-08-31T00:00:00Z http://hdl.handle.net/11356/2358 Slovenian Word Associations with Explanations WAX-sl Vrabl, Klara; Vintar, Špela; Brglez, Mojca WAX-sl is the first Slovenian word association dataset that pairs cue–association pairs with a short explanatory sentence and, for a subset of the data, a semantic relation label. It adapts the methodology of the English WAX (Word Association eXplanations) dataset (Liu et al., 2022) to Slovenian. The 100 cue words were randomly sampled from the Reference List of Slovenian Frequent Common Words (Arhar Holdt et al., 2020), balanced by part of speech (34 nouns, 33 verbs, 33 adjectives). Data were elicited in two phases from 9 paid annotators (students of linguistics/digital linguistics), each cue answered by 4 participants: in Phase 1 participants gave up to three free associations per cue, and in Phase 2 they wrote a short sentence (3–15 words) explaining the relation between cue and association. This produced 1,261 cue–association–explanation triples. A subset of 700 triples (covering all 100 cues) was further annotated with a semantic relation label using an extended version of the WAX relation taxonomy, comprising 21 fine-grained labels across five coarse categories (Concept-Properties, Situational, Taxonomic, Linguistic, None-of-the-Above). Each item was independently labelled by two of five trained annotators and adjudicated by a third curator; inter-annotator agreement (pairwise Cohen's κ) was 0.46. The remaining 561 triples were not selected for relation annotation and are marked NA in the Relation column. The dataset is released as a CSV file with the columns: ID (unique triple identifier), CUE (cue word), ASSOCIATION (association given in response to the cue), EXPLANATION (explanatory sentence linking cue and association), and RELATION (semantic relation label, or NA if not annotated). WAX-sl was created as part of the master's thesis "A multilingual comparison of human-generated and LLM-generated associations" (Vrabl, 2026), which also compares this human dataset against associations generated by three LLMs (Gemma 3, GaMS, Magistral Small). References: Holdt, Špela Arhar, Senja Pollak, Marko Robnik Šikonja, and Simon Krek. 2020. ‘Referenčni Seznam Pogostih Splošnih Besed Za Slovenščino’. Conference on Language Technologies & Digital Humanities. http://hdl.handle.net/11356/1346. Liu, Chunhua, Trevor Cohn, Simon De Deyne, and Lea Frermann. 2022. WAX: A New Dataset for Word Association eXplanations. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 106–120, Online only. Association for Computational Linguistics. https://aclanthology.org/2022.aacl-main.9/ Vrabl, Klara. 2026. A multilingual comparison of human-generated and LLM-generated associations. Master's thesis, Faculty of Arts, University of Ljubljana, http://hdl.handle.net/20.500.12556/RUL-184558. 2026-08-25T00:00:00Z http://hdl.handle.net/11356/2356 Document-level aspect-based sentiment analysis dataset AspectBench (HBS Subset) 1.0 Chatterjee, Nishan; Koloski, Boshko; Doucet, Antoine; Pollak, Senja; Purver, Matthew AspectBench 1.0 (HBS Subset) is an HBS language dataset (hbs, ISO 639-3) containing online news articles, mainly in Serbian and Croatian, for studying sentiment towards target entities like companies and brand names. It is a document-level aspect-based sentiment analysis dataset containing long articles and unseen or new target entities that are not encountered during training. The article text has been filtered to remove URLs. The dataset contains 82,714 distinct records covering 65 targets. The articles were published between 31 December 2012 and 30 December 2023. The held-out test set contains 12,407 records. The other 70,307 records form one training and validation pool. Three random splits of this pool are provided. Each split contains 57,901 training records and 12,406 validation records. Only the source-system internal identifier was removed. The article text, sentiment labels, aspect names, and aspect tags are unchanged. Each record is one document-aspect pair. It contains the full article headline and body, one target company or brand, the target's tracked keyword patterns, the number of matched mentions, publication metadata, and one sentiment label: negative (-1), neutral (0), or positive (1). A single article can occur in more than one record when it has more than one target. The literal <aspect>...</aspect> tags in the article mark the mentions that refer to that record's target. These tags are part of the annotation. Examples: { "uuid": "11855", "article": "bankama u srbiji bolje nego ikada ... neke banke poput sosijete zeneral banke i <aspect>erste banke</aspect> pohvalile su se da su postigle najbolji rezultat otkako posluju u srbiji. ...", "translation": "Banks in Serbia are doing better than ever ... Some banks, such as Societe Generale Bank and Erste Bank, highlighted that they achieved their best result since they began operating in Serbia. ...", "aspect": "Erste banka", "keywords": ["erst* bank*", "ерст* банк*"], "sentiment": 1, "language": "Serbo-Croatian", "published": "2017-05-02T22:00:00.000Z", "country": "RS" } { "uuid": "11854", "article": "bankama u srbiji bolje nego ikada ... <aspect>telenor banka</aspect> ponovila je los rezultat iz prve godine poslovanja, a strucnjaci kazu da je to ocekivano zbog ogromnih ulaganja i da bi odrzivost ovog modela poslovanja mogla da se oceni posle cetiri, pet godina. ...", "translation": "Banks in Serbia are doing better than ever ... Telenor Bank repeated the weak result from its first year of operation, while experts said this was expected because of substantial investments and that the sustainability of this business model could be assessed after four or five years. ...", "aspect": "Telenor banka", "keywords": ["telenor bank*", "теленор банк*"], "sentiment": -1, "language": "Serbo-Croatian", "published": "2017-05-02T22:00:00.000Z", "country": "RS" } Although both records originate from the same article text, their target aspects and corresponding sentiment labels differ. The first record is positive because the article reports that Erste banka achieved its best result since entering the Serbian market. The second record is negative because Telenor banka is discussed as repeating a weak business result. This illustrates why the task is formulated as document-level entity-targeted sentiment analysis: sentiment is predicted with respect to a specific target entity rather than assigned broadly to the document as a whole. The same training and validation pool is supplied under three splits. Therefore, the four partition files contain 223,328 record entries in total. This is not the number of distinct records. The target catalogue contains 65 targets and no dataset records. Files in the submitted archive: - hbs_train_val_0.json, hbs_train_val_1.json, and hbs_train_val_2.json: the three random train and validation splits of the same 70,307-record pool. - hbs_test.json: the held-out test set with 12,407 records. - hbs_aspects.json: the 65 targets, their forms, and keyword patterns. Use this file to identify seen and unseen targets when calculating Macro-F1, quadratic weighted kappa, or other scores. - statistics.json: exact counts for records, articles, labels, languages, characters, and tokens. - hbs-dataloader.ipynb: examples for loading and inspecting the data, displaying sentiment-label tables by target, and tallying seen, unseen, and train/validation-only targets. - README.md: the dataset description, file guide, record-field guide, and usage examples. - CITATION.cff: structured citation information, including the Frontiers article, GitHub repository, and Hugging Face model collection. - LICENSE-and-access.md: a plain-language summary of the access conditions. - release-validation.json: the machine-readable report of the schema, split, target-catalogue, and tag checks. - MANIFEST.sha256: SHA-256 checksums for all files in the archive. 2026-08-04T00:00:00Z http://hdl.handle.net/11356/2357 List of potentially non-standard vocabulary candidates MEZZANINE-NstdLex 1.1 Čibej, Jaka; Gantar, Kaja; Gantar, Polona; Šešet, Jure; Krek, Simon; Robida, Nejc MEZZANINE-NstdLex is a dataset containing 4,281 potentially non-standard vocabulary candidates from the Sloleks Morphological Lexicon of Slovene (collected from among the manually inspected entries of version 3.0; http://hdl.handle.net/11356/1745) and various corpora of spoken Slovene (e.g. GOS 1.1 http://hdl.handle.net/11356/1438; GOS-VL 4.2 http://hdl.handle.net/11356/1444; Artur 1.0 http://hdl.handle.net/11356/1772) and transcriptions (not publicly available) of Slovene university lectures used for the Online Notes project (https://www.cjvt.si/online-notes/). Most of the candidates were collected through manual analysis, with the exception of 1,232 candidate pairs in file "MEZZANINE-NstdLex__Sloleks3.0_levenshtein.tsv", which were extracted automatically by comparing pairs of entries using Levenshtein distance to extract potential non-standard pairs that differ in the presence of the letter "j" (e.g. "genialec" vs. "genijalec"). The candidates were manually annotated in terms of their non-standardness according to a custom typology (STD - standard vocabulary, NST - vocabulary that is non-standard in terms of register, OBL - vocabulary that is non-standard in terms of form, NSTg - vocabulary that is non-standard and typically spoken; combinations of these tags are also possible for borderline examples). The main purpose of the dataset is to provide an overview of different types of non-standard and spoken words in Slovene. The overview can be used as a basis for a robust, empirical development of lexicographic labels for Slovene language resources. In version 1.1, a total of 44 additional manually identified candidates were added in "MEZZANINE-NstdLex__Sloleks3.0_manual.tsv", and several missing or erroneous tags were updated in "MEZZANINE-NstdLex__Sloleks3.0_levenshtein.tsv". For more information on the structure of the dataset, please consult 00README.txt. 2026-08-21T00:00:00Z http://hdl.handle.net/11356/2220 List of headword candidates for the Dutch-Slovene Dictionary nl-sl-Lex 1.0 Grahek, Egidij; Srebnik, Anita; Čibej, Jaka nl-sl-Lex is a list of headword candidates that can be used as a basis to update the Dutch-Slovene Dictionary (Srebnik 2007). The list was extracted from the nlTenTen20 corpus (https://www.sketchengine.eu/nltenten-dutch-corpus/) using Sketch Engine (https://www.sketchengine.eu/) and custom Python scripts. For each part-of-speech (NOUN, VERB, ADJ, ADV; according to the Universal Dependencies POS-tagset: https://universaldependencies.org/u/pos/), lists of up to 1,000 lexeme candidates beginning with each letter of the Dutch alphabet (only lower-case candidates were considered) were extracted from nlTenTen20 (along with their absolute and relative frequencies), then merged, sorted by frequency in descending order, and cross-compared with several existing resources: • the Dutch-Slovene Dictionary (Srebnik 2007) • the list of headwords from the Dictionary of Contemporary Dutch (Algemeen Nederlands Woordenboek - ANW; https://anw.ivdnt.org/lemmalist) • the list of 5,000 most frequent Dutch words (Tiberius & Schoonheim 2014) • two frequency lists for Dutch - 0-2000 & 2000-5000 by Hazenberg & Hulstijn (de Boer et al. 2012) Version 1.0 contains 60,777 candidates and was compiled as a starting point for an analysis of the Dutch-Slovene Dictionary in order to determine priorities for future dictionary updates and facilitate linking with other resources such as WordNets. For more information on the structure of the list, please consult 00README.txt. 2026-06-01T00:00:00Z http://hdl.handle.net/11356/2337 English-Slovene sample of the ETHICS dataset ETHICS-EN-SL 1.0 Novak, Ela English-Slovene sample of the ETHICS dataset ETHICS-EN-SL 1.0 contains an English-Slovene sample derived from the ETHICS dataset (Hendrycks et al. 2021), prepared for a master’s thesis on language as a factor in the moral evaluation of large language models. The resource contains selected examples from five ETHICS subsets: commonsense moral judgment, deontology, justice, virtue ethics, and utilitarianism. The dataset contains 100 examples or scenario pairs per subset. The English files contain the selected examples from the original ETHICS dataset, while the Slovene files contain translated and partially localized versions of the same examples. The bilingual files align the English and Slovene versions row by row. For the Slovene part of the study, a human reference baseline was established on the basis of majority judgments by Slovene annotators. The resource is intended for research on multilingual moral evaluation, language effects in large language models, value alignment, dataset translation and localization, and the creation of localized evaluation resources for less-resourced languages. The data are distributed as UTF-8 encoded TSV files with header rows. The package also includes a README file with the directory structure and column descriptions. The resource is based on the ETHICS dataset introduced by Hendrycks et al. (2021): Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. “Aligning AI With Shared Human Values.” ICLR 2021. https://arxiv.org/abs/2008.02275 2026-08-13T00:00:00Z