corpus

This paper constructs EventTyp and EventTypGold corpora to study typicality and atypicality inferences in conversation, finding that humans rely on three cues—commonsense knowledge, utterance emphasis, and snark—while LLMs struggle due to lacking commonsense knowledge about event typicality.

论文提出 Wieszcz-XIX，一个涵盖 1800–1918 年波兰语的 31 亿词（67.5 亿 token）语料库，并从零训练了 47M 至 349M 参数的时间受限语言模型，展示其在时期词汇上相对现代波兰基座模型的交叉优势，同时发布语料、代码与权重。

YODAS v3 is a weakly-labeled speech corpus containing over 1.1 million hours of high-bandwidth, stereophonic, multilingual audio in 147 languages, designed to advance speech foundation models.

We introduce ISAAC, an open corpus of 527 million+ English-language Reddit posts for analyzing social group discourse, with a multi-step pipeline for annotation and analysis. It enables cross-category comparisons and temporal tracking of public attitudes, accessible via web apps and programming interfaces.

This paper explores using lightweight webcam-based eye-tracking features to improve keyphrase extraction from Chinese academic abstracts, introducing the CLIS-ET corpus and showing that fixation features enhance KPE performance.

This paper introduces PERCEPT, the first large-scale Persian–English code-mixed corpus annotated with Universal Dependencies POS tags, collected from X, Instagram, and Digikala. It presents an LLM-assisted annotation framework and analyzes code-mixing patterns across platforms.

Introduces the 英语台词社 website, which provides a searchable corpus of subtitles from over 54,000 movies and 110,000 episodes of American TV series, for practicing English listening, shadowing, and authentic expressions.

This paper introduces ICLE++, a new corpus of persuasive student essays annotated with both holistic and trait-specific scores, aimed at improving generalization and multi-trait scoring in automated essay scoring (AES) research.

This paper presents a human-in-the-loop workflow for creating a corpus of LLM-based simplifications of scientific summaries, using SciSummNet and GPT-4o-mini with non-expert and expert feedback to improve cross-disciplinary accessibility.

A new annotated corpus of persuasion techniques in Bulgarian, Polish, and Russian, covering parliamentary debates and social media, with 25 fine-grained techniques and baseline models for detection and classification.

Introduces CKTN, the first multilingual corpus and benchmark for Cham, Khmer, and Tay-Nung languages, addressing fragmentation in multilingual encoders with a script-aware adaptation recipe.

Presents an open, offline word-level digital reader of the Prasthānatrayī with Śaṅkara's Bhāṣya, featuring clickable word analysis, concordance, and a hybrid pipeline using rule-based and LLM-assisted methods.

Introduces RMISC, a large-scale real-world multivariate time series corpus with around 200 datasets and 142 billion time points, and demonstrates that pretraining time series foundation models on real-world multivariate data improves zero-shot generalization compared to synthetic data.

An 18-year-old Tunisian student introduces an open-source machine translation pipeline and parallel corpus for Tunisian Darija in Arabizi script, built from scratch with a small 15.6M-parameter Transformer and an honest baseline BLEU of 3.89, and calls for contributors to ethically expand the corpus.

This paper introduces MultiSynt/MT, a trillion-token multilingual parallel corpus created by translating English pre-training data into 36 languages. Experiments show that LLMs trained on this translated data achieve performance comparable to native data with fewer tokens, though some evaluation blind spots and cultural gaps remain.

This paper introduces SEFORA, a public corpus of instructor feedback on student essays, and UniMatch, a reference-based evaluation framework for assessing LLM-generated feedback. Experiments show that current LLMs struggle to match instructor feedback, achieving at most 0.4 F1.

Introduces MMEE, a multilingual multi-emotion emphasis corpus of 10,000 utterances across 7 languages and 34 emotions, and benchmarks emphasis detection models under various transfer settings, finding that multilingual training improves robustness while monolingual models show limited zero-shot transfer.

This paper introduces two new Czech corpora, Hlava Cor and Hlava AD, designed to study human label variation in coreference and discourse relations. The corpora feature multiple annotations and annotator explanations, achieving 60-65% inter-annotator agreement and revealing systematic differences in interpretation.

We present the second consolidated version of the Prague Dependency Treebank, a 4-million-token manual multilingual annotation resource covering morphology, syntax, semantics, coreference, and discourse, along with compatible lexicons.

Compiled the public quarterly reports, notes, and interviews of fund manager Zheng Xi into a structured corpus, and built it as a traceable skill across AI platforms for real data-driven investment research Q&A and fund analysis.