Current Abstract Meaning Representation (AMR) annotation guidelines, which largely tie argument structure to lexical rolesets, systematically misrepresent cases in which key semantic roles stem from clause-level structure rather than the verb, leaving these meanings either unnaturally attached, incorrect, or unexpressed. To address this limitation, we present CxGr-AMR, a novel extension of AMR that captures the semantics of various types of phrasal constructions, including argument structure constructions. We first examine how such cases are handled under current Standard-AMR guidelines and show that these analyses are often inadequate when constructionally contributed roles clash with those assigned by the verb. We then provide a theoretical grounding for our CxGr-AMR rolesets that lay out the relationship between the syntactic signatures of constructional slots and particular semantic roles associated with them. Finally, we develop an annotation-expert-in-the-loop pipeline for the semi-automatic annotation of sentences, and release a dataset containing 355 instances of phrasal constructions annotated with both Standard and CxGr-AMR.

To fully capture the meaning of a sentence, semantic representations should encode aspect, which describes the internal temporal structure of events. In graph-based meaning representation frameworks such as Uniform Meaning Representations (UMR), aspect lets one know how events unfold over time, including distinctions such as states, activities, and completed events. Despite its importance, aspect remains sparsely annotated across semantic meaning representation frameworks. This has, in turn, hindered not only current manual annotation, but also the development of automatic systems capable of predicting aspectual information. In this paper, we introduce a new dataset of English sentences annotated with UMR aspect labels over Abstract Meaning Representation (AMR) graphs that lack the feature. We describe the annotation scheme and guidelines used to label eventive predicates according to the UMR aspect lattice, as well as the annotation pipeline used to ensure consistency and quality across annotators through a multi-step adjudication process. To demonstrate the utility of our dataset for future automation, we perform simple baseline experiments using three modeling approaches. Our results establish initial benchmarks for automatic UMR aspect prediction and provide a foundation for integrating aspect into semantic meaning representations more broadly.

Idiomatic expressions, a subclass of multiword expressions (MWE), pose persistent challenges for semantic parsing, as their meanings often diverge from the compositional semantics of their constituent words and depend strongly on contextual cues. While Abstract Meaning Representation (AMR) parsers aim to capture sentence-level semantics in a structured graph form, existing datasets provide limited coverage of idiomatic language, constraining their ability to model such expressions accurately. To address this gap, we extended a subset of the MAGPIE dataset by constructing a corpus of potentially idiomatic expressions (PIE) annotated with their corresponding AMR graphs. The dataset includes both naturally occurring and synthetically generated sentences, covering idioms in literal and idiomatic contexts. We fine-tune a state-of-the-art AMR parser on this dataset and evaluate its capacity to generate context-sensitive graphs that correctly reflect idiomatic versus literal interpretations. Our results show that standard parsers often capture only literal meanings of such expressions, while fine-tuning on our dataset improves alignment with the intended interpretations.

Abstract Meaning Representation (AMR) is a graph-based semantic representation which captures the core elements of meaning of a text. AMR has been incorporated into a variety of downstream tasks, which rely heavily on the availability of gold-annotated AMR corpora. While the annotation process is fairly lightweight, annotator training is still required even for linguists due to the extensive nature of the annotation guidelines and comprehensive set of roles. Therefore, all corpus development projects for AMR (and extensions of AMR) require the dataset curators to first train annotators. In this paper, we develop an online AMR annotation training system called TrAinMR in order to ease this training process and thus motivate the development of additional AMR corpora. The two main components of TrAinMR are (1) a written tutorial covering the basics of AMR annotation, and (2) an interactive practice module with corrective feedback. To measure the effectiveness of this tool, we conduct two pilot studies with five human annotators each. We find that the majority of annotators state their understanding of AMR improved as a result of TrAinMR, and some annotators show a positive trend in SMATCH scores after completing the practice module.

Existing Persian named entity recognition (NER) research has focused predominantly on news and social media domains, leaving literary texts—with their distinct linguistic characteristics—virtually unexplored. This paper addresses this gap by developing a new literary NER corpus using the Persian translation of The Little Prince story and evaluating existing state-of-the-art Persian NER tools on this corpus, trained exclusively on news and social media corpora. Our analysis reveals significant performance degradation on literary text, identifying systematic errors related to narrative-specific entities, metaphorical language, and discourse structures that challenge conventional NER approaches.

Deverbal nouns pose challenges for semantic annotation frameworks that aim to represent event structures consistently across lexical categories. This paper examines problematic phenomena in the annotation of deverbal nouns in Czech and Latin within the Universal Meaning Representation (UMR) framework, addressing both manual graph construction and rule-based automatic conversion from existing resources. Current UMR guidelines lack operational criteria for deciding when a noun should be treated as an eventive concept, particularly in the absence of a PropBank-like lexicon with sufficient nominal coverage. We therefore propose practical annotation principles: deverbal nouns denoting events (such as učení ‘teaching’), results of events (řešení ‘solution’), or event participants (učitel ‘teacher’) should be related to underlying event concepts (represented as verbs in their particular senses, i.e., učit-001 ‘to teach’, vyřešit-001 ‘to solve’, and učit-001 ‘to teach’, respectively), while other deverbal nouns should remain unrelated to respective events (such as učebna ‘teaching room’). To reduce inter-annotator variation, we further suggest systematic strategies for selecting verbal labels, including the use of light-verb constructions, synonymous verbs, and a preference for imperfective verbs in Czech aspectual pairs. For automatic conversion, we outline a rule-based approach that combines multiple lexical resources and frequency-based heuristics to identify corresponding verb senses. Our findings provide guidelines for more consistent UMR annotation across languages.

This paper presents SAVI, a web-based interface for multilayer semantic annotation validation of Universal Semantic Representation (USR). USR encodes meaning across interdependent lexical, constructional, relational, discourse, and co-reference layers, making validation challenging using conventional annotation tools. SAVI addresses this limitation through structured tab-based layer separation, constraint-aware editing mechanisms, and role-based review workflows. The system integrates a multilingual concept dictionary to ensure sense-level consistency, along with a Hindi text-generation module and dependency-based visualization to support interpretation and correction. SAVI is implemented using a Flask backend, Flutter frontend, and PostgreSQL for structured data management. Evaluation results demonstrate effective governance of concept proposals and improved efficiency in multilayer USR correction, positioning SAVI as a structured validation framework for scalable semantic corpus development.

Sentence embedding techniques aim to encode key concepts of a sentence’s meaning in a vector space. However, the majority of evaluation approaches for sentence embedding quality rely on the use of additional classifiers or downstream tasks. These additional components make it unclear whether good results stem from the embedding itself or from the classifier’s behaviour. In this paper, we propose a novel method for evaluating the effectiveness of sentence embedding methods in capturing sentence-level concepts. Our approach is classifier-independent, allowing for an objective assessment of the model’s performance. The approach adopted in this study involves the systematic introduction of syntactic noise and semantic negations into sentences, with the subsequent quantification of their relative effects on the resulting embeddings. The visualisation of these effects is facilitated by Concept Separation Curves, which show the model’s capacity to differentiate between conceptual and surface-level variations. By leveraging data from multiple domains, employing both Dutch and English languages, and examining sentence lengths, this study offers a compelling demonstration that Concept Separation Curves provide an interpretable, reproducible, and cross-model approach for evaluating the conceptual stability of sentence embeddings. The open-source code and a live interactive demo are available upon acceptance.

In this paper, we present a method for interpreting Yarn structures as logical formulas in a modal first order logic with temporality. Yarn is a recent semantic formalism that aims to bridge the gap between graph-based and logic-based semantic representations, providing a flexible and expressive framework for capturing the meaning of natural language utterances. Our approach translates the elements of Yarn structures such as predicates, features, into corresponding logical constructs, allowing for an interpretation of the represented meaning. Given that Yarn allows ambiguous representations, we associate to each Yarn structure a set of possible interpretations. We account for a range of semantic phenomena, extending beyond ambiguity to capture aspects of dynamic quantification as well. This work contributes to the understanding of the expressive power of graphical semantic representations and their relationship to formal logic.

Large language and vision-language models (VLMs) struggle with a ‘compositionality gap’. They treat language as a sequence of tokens lacking any structure and thus rely on a large number of parameters making them computationally expensive. To address these issues, we propose CCG-VQC, a quantum framework that unifies statistical distributions with linguistic structure. Guided by Combinatory Categorial Grammar, our model maps syntactic rules into parametrised quantum circuits and models sentences as quantum states. We evaluate CCG-VQC on structural VLM benchmarks such as ARO and SVO-Swap. Our experiments show that CCG-VQC consistently outperforms a quantum bag-of-words model, as well as classical VLMs such as CLIP and OpenCLIP. CCG-VQC achieved 71.19% accuracy on ARO-Attribution, significantly outperforming the parameter-matched MicroCLIP, which struggled to surpass random chance with a maximum performance of 50.85%.

Semantic frame, role, and relation labels are important parts of symbolic meaning representations. The mainstream approach is to define large language-specific lexicons that map predicate senses to their frame and argument labels, and to use separate label inventories for modifiers. Maintaining lexicons is very labor-intensive and scales poorly to the multilingual case. There are schemas that aim to simplify the annotation task using a more coarse-grained inventory of frames and roles, but suffer from a lack of systematicity and clear definitions. We present a schema that 1) uses a small inventory of frames and roles for lexicon-free annotation, 2) systematizes the frame inventory by factoring out aspect and mode, 3) has a unified vocabulary for arguments and modifiers, and 4) is designed to be annotated atop Universal Dependencies syntactic annotation. We argue for the adoption of such a schema for multilingual annotation and demonstrate promising results in annotation experiments on German.

The paper presents the first shared task on parsing Uniform Meaning Representation (UMR), a graph-based framework for cross-linguistic semantic annotation of typologically diverse languages. The task requires systems to enrich plain text with sentence-level structure, node–token alignment, and document-level relations. It involves processing data for seven languages from four language families (Indo-European, Sino-Tibetan, Na-Dene, and Algic). Six languages have at least some training data; for one language, data is not available, leading to a zero-shot scenario. The training dataset as well as the gold-standard test set for all seven languages is released and made available for follow-up research. We present the task setup and evaluation methodology, using two graph matching approaches – a traditional, and an alignment-sensitive one, tailored specifically for UMR. Two participating systems are compared, each representing different modeling approaches. Results highlight the challenges of UMR parsing, particularly for alignment prediction and document-level semantics, and reveal substantial variation across languages and annotation conditions.

Uniform Meaning Representation (UMR) is a novel meaning representation formalism emanating from Abstract Meaning Representation (AMR). Since it is more complex than AMR, including document level annotation it is more difficult to create a parsing pipeline which can predict an UMR document from a set of consecutive sentences. The UMR Parsing Shared Task was created to compare different approaches. We decided to use a 2-step approach to predict sentence level and document level annotation. Since the available data was limited, we opted for a multilingual model, even though unlike AMR, in UMR the concepts of the meaning graph are not drawn from a single source, but from language dependend resources. Our final score was 19.35%, 0.08 points behind the best participant (19.43%).

We present the Sema system for the DMR 2026 shared task on parsing from natural language to . Our approach relies on parameter-efficient fine-tuning of Qwen3-4B with a multistage training procedure. We first train on a capped subset of the noisy training data, then continue training on the clean split, and finally fine-tune a dedicated stage for word-to-node alignment prediction. The system generates sentence-level graphs, selected document-level information, and alignments in separate steps, followed by rule-based post-processing to satisfy the official evaluation format. Results show that the approach is viable across several languages and exhibits promising transfer to Italian despite the absence of Italian data for fine-tuning , while very low-resource languages remain challenging.

We present ongoing work on annotating fine-grained semantic distinctions for circumstantial meanings, focusing on spatial expressions. We describe our theoretical background, and annotation process, as well as how we evaluate the results obtained. Using multiple independent annotations across the 3-million-token, genre-diverse Prague Dependency Treebank – Consolidated corpus of Czech data, we analyse inter-annotator agreement, recurrent disagreement patterns, and the limits of semantic categorization. Our results highlight the inherent vagueness of linguistic meaning. We also propose strategies for handling disagreement, such as weighted annotations, intermediate labels, and fuzzy labels that preserve annotation nuance. This work builds on the legacy of Petr Sgall and the Functional Generative Description theory that underpins the multi-layer form–meaning framework.

Uniform Meaning Representation (UMR) has cross-linguistic design principles that make it particularly well-suited as a semantic representation framework for capturing all language-specific phenomena. Despite its growing adoption, no UMR corpus currently exists for Persian. In this paper, we present the first version of a Persian UMR dataset created through a rule-based conversion of existing Persian AMR annotations from The Little Prince corpus, followed by manual mapping of split semantic roles from AMR to their finer-grained UMR counterparts. We report detailed statistics on the conversion, analyze the challenges of mapping Persian AMR structures into UMR, and provide illustrative examples. The resource is freely available and it lays the groundwork for subsequent enrichment of Persian UMR with additional semantic layers, including co-reference, named entities, and discourse relations.

We present a platform for developing compositional semantic annotations for formal syntactic representations that allows users to interact with and explore annotations and to track their progress and quality. For this, we provide several forms of visualizations and take inspiration from research in linguistic treebanking. Thus, we contribute to the development of formal semantic parsers and corresponding meaning banks. The system is designed with a regression testing paradigm in mind and provides support for NLI so that the created semantic resources can be developed and validated in a task-driven environment. We defend this paradigm in comparison to modern approaches to semantic parsing that are mainly evaluated on the basis of gold standard annotations.