We present the Annotated Reddit Conversation Corpus (ARCC), an English-language dataset of online discussions annotated for Speech Acts and Functional Dependence Relations, designed to investigate how varying degrees of subjectivity influence conversational dynamics and interaction patterns. At the speech act level, we distinguish factual from opinion statements and further classify opinions along a five-degree scale of subjectivity. Functional Dependence Relations capture how segments relate to preceding ones. Analyses show that opinion-discussion contexts feature frequent inter-subjective opinions eliciting explicit agreement and disagreement, while information-exchange contexts exhibit less subjective opinions with responses like answers or requests for clarification. We further demonstrate that a transformer model can predict the subjectivity scale with promising performance. The corpus and annotation guidelines are made available to support future research on opinion expression and automated dialogue analysis.

The N+N compound class is the largest and the most productive class of compounds in German. A significant number of N+N compounds insert a so-called linking element from a large inventory. The linker choice is notoriously irregular; instead of rules, it is governed by a set of constraints that can only limit this choice based on morphological, phonological, sometimes semantic and lexical properties of the first constituent. While constraints on linking element choice in German nominal compounding are extensively researched and well-documented, no large-scale corpus study has ever been reported on the subject of their empirical application. The present work aims at filling in this gap by conducting an extensive corpus study on potential and actual applicability of these constraints. The study summarizes 64 constraints collected from the relevant literature and obtains applicability statistics for 39 of them over 280k+ German N+N compounds. The study both confirms most of the evidence from previous literature and suggests novel evidence on German nominal compounding. It additionally highlights the importance of structured linguistic data for large-scale empirical studies.

This study compares the impact of natural and synthetic data on training and evaluating large language models (LLMs), using the case of passive verb alternation in French and Italian. We use Blackbird Language Matrices (BLMs), structured datasets designed to probe linguistic knowledge of underlying patterns across sentence sets. We compare structured templates instantiated with natural sentences extracted from Universal Dependencies to structured templates of synthetic sentences. Experiments show that while models achieve ceiling performance when trained and tested on synthetic datasets, they do not reliably generalize to natural sentences. In contrast, models trained on natural data exhibit robust performance across both natural and synthetic test suites, demonstrating their superior ability to capture abstract linguistic patterns. These results corroborate the value of natural data and of structured set ups in linguistic evaluation for probing LLMs’ syntactic and semantic knowledge.

We test whether the subevent structure encoded in VerbNet-GL predicts entity identity change in procedural text, using only the verb’s lexical specification and no training data. From each VN verb class’s SEMANTICS block we extract an aspectual classification and an I/O count, yielding a predicted dynamic event topology (DET). On the observation side, ten large language models (LLMs) annotate per-entity dynamic object mode (DOM) labels over ∼3100 OpenPI steps, from which we derive observed DET for comparison. The VN-only predictor achieves 67.4% precision (F1 = 0.35) for transformation, showing that formal subevent structure carries genuine predictive signal only for the most common topology, but low performance for other topologies. For events VN predicts as having no result state, only 24% are confirmed as no-change by silver, indicating that the remaining outcomes arise from the argument side of the composition. These results provide empirical evidence that event semantics is distributed across predicate and argument: the VN supplies the subeventual skeleton, but is not sufficient to determine the final outcome.

In this paper we introduce an example-based method for exploring dependency treebanks that is based on principles of vector symbolic architectures. It leverages key properties of this framework to provide fast and flexible search capabilities, since all combinations of query parameters can be compared with a given parse tree in parallel via a single vector operation. The framework also allows for graded similarity and the natural integration of various kinds of information, such as word embeddings. After some background on the framework and an explanation of our implementation, we provide a few examples of the system’s output and draw comparisons to similar applications.

Neural machine translation (NMT) has become the predominant approach for automated translation, yet conventional models trained on extensive bilingual datasets exhibit critical limitations, including quadratic scaling of training data, sensitivity to out-of-distribution inputs, and a lack of interpretability. Inspired by the classical “translation pyramid” concept, which advocates for translation via a semantic pivot (interlingua), this work explores the integration of Abstract Meaning Representation (AMR) as a structured semantic intermediary to decouple translation into comprehension (source-to-AMR) and generation (AMR-to-target) phases. We conduct a pilot study using a strong AMR parser to create a multilingual silver-standard AMR corpus from the United Nations Parallel Corpus, training modular semantic understanding and generation components for each language. Experimental results demonstrate that our approach achieves an average improvement of 3% in robustness and over 15% in generalization compared to traditional Seq2Seq baselines. Analysis suggests that enhancing semantic parsing and generation accuracy could bridge the gap to conventional NMT systems. To our knowledge, this is the first work to integrate AMR as a semantic pivot in NMT, offering enhanced transparency, scalability, and robustness. This study underscores the potential of semantic-driven translation frameworks and provides a foundation for future research in interpretable, resource-efficient multilingual systems.

We introduce Gutenberg+, a temporally more faithful version of the Project Gutenberg (PG) corpus, one of the most widely used resources for diachronic text analysis. Despite its popularity, the PG corpus contains a major yet overlooked flaw: around 15% of its entries are collections (e.g., anthologies of books, letters, or poems) rather than atomic works, which distorts temporal analyses since such collections may span multiple decades. We present an automatic method to detect and split these collections into their constituent works, producing a finer-grained and temporally consistent corpus. We further re-annotate publication years using LLM-based retrieval-augmented generative methods, demonstrating the potential of LLMs to enhance structured linguistic resources. To illustrate the utility of Gutenberg+, we conduct a small-scale diachronic case study on negation, showing that our refined corpus captures more nuanced cross-linguistic variation than the original PG data. Finally, we release the corpus in UIMA format with full metadata and linguistic annotations, providing a standardized resource for future research on diachronic language change.

We report on a new lemmatisation system for Norwegian, which is a particularly challenging language with two written standards, Bokmål and Nynorsk, that both have a lot of optionality. Our system covers both varieties and consists of a neural model that classifies words into rewrite rule classes that produce their lemma, as well as a large-scale computational lexicon of Norwegian that gives all possible inflections of a large part of the Norwegian vocabulary. We test different ways of combining these components. When evaluated with pure string-matching against the lemmas in the gold data, all systems perform approximately at the same level (99.1-99.2% on Bokmål and 98.5-98.6% on Nynorsk), but detailed error analysis shows that the computational lexicon reduces the number of true errors by more than half (reaching 99.6% accuracy on Bokmål and 99.3% on Nynorsk), as opposed to “surface errors” like using a different, but equally acceptable spelling variant of the correct lemma.

Quantification is common in text, but it is underrepresented in AMR. Quantification also stresses the common conjunctive interpretation of AMR graphs, since universal quantification introduces scope-taking structure and variable binding that cannot be captured as a flat list of conjuncts. We propose an enriched AMR that supports quantificational meaning while keeping AMR’s graph backbone. At the predicate level, we add QuantML features, such as domain restriction, determinacy, distributivity, and involvement. At the discourse level, we add contextual constraints that encode scope and other discourse-sensitive conditions. The two levels follow the UMR architecture and are linked by shared identifiers. We map the enriched graphs to two-block logical forms: a minimal model of events and participants, plus a constraint block that relates them.

This paper presents a two-stage training approach to improve the performance of Slovene large language models on lexicographic question-answering tasks. We developed a comprehensive lexical pretraining corpus containing 356,294 Slovene word entries. We constructed the corpus by converting structured data from multiple lexicographic sources into markdown format. Additionally, we created a question-answering dataset with 10,485 QA pairs from diverse sources, including automatically generated questions, a linguistic advisory portal, and community forums. Using the Slovenian GaMS model (based on Gemma 2 9B) and GaMS 3 model (based on Gemma 3 12B), we performed continued pretraining on the lexical corpus, followed by instruction fine-tuning with our QA dataset combined with translated general-domain questions. We compared results to different model configurations. Our results demonstrate significant improvements (text similarity increasing from 0.226 to 0.542, BERTScore F1 of 0.915) in answering Slovene lexicographic questions, validating the effectiveness of domain-specific continued pretraining for low-resource languages.

Non-concatenative morphology poses a persistent challenge for NLP, yet structured quantitative resources for Amazigh (Berber) languages remain scarce. We present the first large-scale computational study of Tashlhiyt Berber plural formation, drawing on a richly annotated dataset of 1,185 noun paradigms with phonological, morphological and semantic features. We decompose the plural system into macro-level word-formation strategies and micro-level stem mutations, and evaluate predictability across ten target domains using linguistic feature models, N-gram baselines, and Bi-LSTM neural models. Results reveal a structured split: linguistic features decisively outperform neural models on systematic macro-level strategies (e.g., +44.5pp F1), while Bi-LSTMs better capture lexically idiosyncratic patterns. Rather than supporting a categorical rule/memory divide, this complementarity reveals gradient layers of regularity within a single morphological system. These findings demonstrate the value of linguistically informed annotation for probing morphological complexity in low-resource, typologically diverse languages. All data, code, and models are publicly available.

Empirical grammar research has become increasingly data-driven, but the systematic analysis of annotated corpora still requires substantial methodological and technical effort. We explore how agentic large language models (LLMs) can streamline this process by reasoning over annotated corpora and producing interpretable, data-grounded answers to linguistic questions. We introduce an agentic framework for corpus-grounded grammatical analysis that integrates concepts such as natural-language task interpretation, code generation, and data-driven reasoning. As a proof of concept, we apply it to Universal Dependencies (UD) corpora, testing it on multilingual grammatical tasks inspired by the World Atlas of Language Structures (WALS). The evaluation spans 13 word-order features and over 170 languages, assessing system performance across three complementary dimensions – dominant-order accuracy, order-coverage completeness, and distributional fidelity – which reflect how well the system generalizes, identifies, and quantifies word-order variations. The results demonstrate the feasibility of combining LLM reasoning with structured linguistic data, offering a first step toward interpretable, scalable automation of corpus-based grammatical inquiry.

Large language models have renewed interest in the role of structured linguistic data for applications that require controllable, interpretable, and pedagogically aligned language generation. This need is especially visible in intelligent language tutoring, where grammar cannot be modeled as a flat inventory of patterns alone, but must also capture their relations and functions they realize. We present the L2 Network, a machine-readable knowledge graph of CEFR A1-A2 English grammar that encodes formal patterns, functions, and typed relations between them. The resource is grounded in established pedagogical reference materials, combining form inventory and progression information from the English Grammar Profile with a functional layer derived from CEFR descriptors. We further report content validation of the form-function mappings through expert annotation, including agreement analysis and a consensus-filtered core release. The resulting graph provides an explicit schema for representing pedagogically relevant grammatical knowledge and supports downstream uses such as learner modeling, adaptive task selection, and controlled generation in dialogue-based ICALL systems.

We utilize neural word alignment to improve the quality of paraphrase databases in English and Japanese. For large-scale paraphrase acquisition, previous studies have employed a framework of bilingual pivoting based on word alignment on bilingual parallel corpora. Naturally, the quality of paraphrases acquired by bilingual pivoting depends on the performance of word alignment. Previous studies based on statistical word alignment have limitations in the quality of acquired paraphrases because they do not consider word meaning. This study employs a more sophisticated neural approach for word alignment in bilingual pivoting to enhance the quality of paraphrase acquisition. Experimental results revealed that our paraphrase databases outperformed existing ones in both internal and external evaluations.

Deep learning methods in natural language processing often rely on statistical methods to tokenize texts before vectorization. This segmentation produces lexical subunits offering great flexibility. However, the reuse of identical tokens across words with different meanings can favor representations based on surface form rather than on linguistic information, especially semantics. This mismatch between semantics and surface form can lead to undesirable effects in language processing. To limit the influence of form on the semantics of vector representations, we propose an intermediate representation based on syntactic parsing that is more compact and more faithful to word meaning.

We present the largest multilingual experiment to date on word-to-morph segmentation, covering 58 typologically diverse languages. We describe a newly compiled collection of linguistically annotated resources for the task, providing broad coverage and enabling systematic cross-lingual evaluation. Second, we train two neural models on surface morphological segmentation, achieving 81% average word accuracy on the original datasets, slightly outperforming previous methods. Experiments on custom test sets reveal substantial variation in performance, highlighting the need for further harmonization and more robust multilingual approaches.

To foster investigations of noun phrase (NP) inflection in German at scale, this paper introduces DiNoS (Distributional Noun Structure), a data-driven lexicon of NP heads, which includes statistical information on the dependents and the morphosyntactic features of their original in-context appearances. We make available the source code for the extraction of NPs from CoNLL-U treebanks, which includes rule-based heuristics to improve feature annotation coverage and ensures a homogeneous lemmatisation strategy across treebanks. While the resulting JSON-based lexicon is suitable for no-code interaction for non-experts, it is further supported by a toolkit for the automatic calculation of, and access to, various statistical overviews. In this paper, we present the heuristics employed to extract NP datasets from the German Universal Dependencies’ Hamburg Dependency and GSD treebanks. In addition, we provide a preview of the emerging DiNoS lexica’s properties and discuss some implications of noun and determiner word form ambiguity for NP complexity.

Metaphor, a figure of speech and a cognitive device, offers a powerful way to explain one conceptual domain in terms of another. Particularly successful metaphorical mappings are conventionalized through frequent use and lose their creative quality. They become sense extensions of polysemous words. Our dataset project captures such metaphors with ten regular polysemy patterns that manifest repetitively in the meaning structures of English words. Regular metaphor, unlike its counterpart regular metonymy, has not previously received a dedicated dataset, and we intend to close this gap. The dataset under construction features naturalistic sentences extracted from a general language corpus and is manually annotated with sense labels for metaphorically extended polysemes. Its intended use is to support linguistic, cognitive, and computational investigations into patterns of meaning in polysemy, while accounting for its complexity, regularity, continuity, and heterogeneity. We see neural language models as an excellent experimental ground for such research because they are able to show both distributional (continuous) and symbolic (discrete) behavior in language processing and representation. In this paper, we reflect on how these systems tally.

We present a new corpus of contemporary German sermons and describe the steps taken in its preparation. We apply a semi-automatic approach to sentence segmentation, tokenization, and lemmatization, utilizing annotation guidelines that are specialized to this domain. In the process of preparing these data, we find that state-of-the-art tools for these tasks still make problematic errors, especially with non-standard data, despite apparently very high performance on common benchmarks. We obtain test scores of F1 = 96.69 % for sentence segmentation, F1 = 99.99 % for tokenization, and acc = 64.00 % for lemmatization with our domain-adapted models and show that domain-adaptation improves performance over state-of-the-art models for the token and sentence segmentation tasks.

Data augmentation is a common fix in domains where training data is scarce or difficult to collect, such as specialized medical or any other domain specific applications. In question answering (QA), most studies report headline accuracy while saying little about the quality of the synthetic data. Here, quality goes beyond fluent rewording: augmented items must remain faithful to the supporting evidence and preserve the original answerability. We study three augmentation families lexical, syntactic, and semantic edits generated with LLaMA 3.1 70B, and analyze how these edits affect model behavior. To mirror low-resource settings, we focus on subsets of SQuADv2 (general) and PubMedQA (biomedical, domain specific). We report Exact Match (EM)/F1 alongside quality diagnostics, yielding a fuller picture than accuracy alone. Our results show that augmentation behaves differently across domains and scales. In SQuADv2, augmented variants maintain performance on par with baselines, showing that added diversity mostly does not harm model quality, whereas in PubMedQA semantic edits bring improvements under extreme scarcity and support stronger performance as supervision grows.

As previous research on annotator disagreement in discourse phenomena has shown, understanding text coherence varies considerably from one individual to another. To explore this phenomenon, we created two corpora with multiple annotations of Czech texts, accompanied by annotators’ explanations of their choices. The first corpus consists of 1,024 contexts annotated in parallel by three annotators. It captures differences in the identification of coreference across various text types and grammatical-semantic categories, including pronouns, full noun phrases, and anaphoric adverbials. The second corpus comprises 512 contexts, annotated in parallel by five annotators, and focuses on identifying discourse relations in attributive and non-attributive constructions. Both corpora achieve a comparable inter-annotator agreement of approximately 60–65%. For coreference annotation, agreement tends to be lower in cases where automatic coreference resolution models disagree, suggesting that when the models disagree, the examples tend to be more difficult or ambiguous for human annotators to interpret. The annotators’ comments, both for coreference and discourse relations, further reveal differences in interpretation, varying levels of confidence in text understanding, and individual reading strategies.