Euphemisms present a persistent challenge for NLP because their interpretation depends on pragmatic inference, social norms, and contextual cues rather than surface meaning alone. Although Potentially Euphemistic Terms (PET)-based resources have been developed for several languages, Japanese euphemisms remain computationally unexplored despite their close interaction with honorifics, register variation, and orthographic choice. We introduce JP-PET, the first PET-based dataset for Japanese euphemism classification, comprising 1,672 annotated sentences across 101 PETs and ten semantic domains with register metadata. We evaluate two Japanese monolingual transformer models (Rinna RoBERTa and Tohoku BERT) and the multilingual XLM-R under three controlled PET-level data splits that isolate lexical familiarity and generalization to unseen euphemisms. While models achieve strong performance when PETs are shared between training and test data, performance drops substantially under PET-disjoint conditions, indicating reliance on lexical familiarity. Error analysis reveals systematic challenges in politically conventionalized expressions, metaphor-based euphemisms, and orthographic mitigation strategies. JP-PET provides the first benchmark for studying pragmatic meaning in Japanese NLP.

Pragmatic competence poses a persistent challenge for large language models, as it requires context-dependent inference beyond literal meaning. This study examines whether few-shot prompting can reliably steer LLMs toward appropriate interpretations of indirect speech acts under small-data conditions. Focusing on Italian, we evaluate three LLMs on a small dataset that captures pragmatic ambiguity through graded plausibility judgments. We compare a zero-shot baseline with multiple few-shot prompting configurations that vary in the number and composition of demonstrations, as well as in the presence of explicit pragmatic guidance. Results show that few-shot prompting does not yield robust or monotonic improvements overall. While performance improves substantially for conventionalized indirect speech acts, gains for non-conventionalized indirect speech acts are unstable and limited. In contrast, introducing explicit pragmatic reasoning along with demonstrations through guided chain-of-thought prompting appears more promising. Overall, these findings highlight the limits of example-based steering for pragmatic inference and suggest that explicitly modeling pragmatic reasoning may be a more effective direction in small-data settings.

In potentially idiomatic expressions (PIEs), the same surface form may receive either a literal or an idiomatic interpretation depending on context, making automatic literal–idiomatic disambiguation challenging. This is acute for Croatian, where annotated data and locally runnable generative models are limited. We present a study of Croatian PIE literal–idiomatic disambiguation examining how structured lexicographic knowledge can improve open-weight, decoder-only LLMs without fine-tuning. Using a new expert-annotated concordance dataset – CroPIEs, we compare baseline prompting to inference-time knowledge injection via retrieval-augmented generation (RAG) from a Croatian phraseological dictionary. We isolate the contribution of three knowledge types: definitional knowledge (structured meanings), contextual knowledge as curated prototypical usage examples, and their combination. Results show consistent improvements in macro-F1 for both GaMS-2B-Instruct and GaMS-9B-Instruct models. Definitional knowledge is generally more stable than examples alone, while examples can be effective but less consistent across expressions. The strongest and most reliable gains are obtained when definitions and examples are combined, indicating a synergistic effect between explicit meaning descriptions and contextual cues. Per-class analyses show that injected lexicographic evidence mitigates baseline biases between Literal and Idiomatic predictions, improving decision balance in a low-resource setting with small data of compact, expert-curated lexicographic evidence injected at inference time.

Metaphor identification remains challenging in specialized and low-resource domains, where large annotated datasets are unavailable and general-domain models often fail to transfer effectively. In this paper, we evaluate FLAVORS-AECC, a Spanish dataset of oncological discourse that provides transparent, instance-level annotations of basic meaning (BM) and contextual meaning (CM) following the Metaphor Identification Procedure (MIP). We test the state-of-the-art Contrast-WSD model under two splits: a random split and a lemma-based split to control for lexical memorization. We compare three configurations: (i) a control model with no meaning information, (ii) manually curated basic meanings, and (iii) first dictionary entry as an approximation of basic meaning. Results show that explicitly modeling meaning contrast substantially improves performance in low-resource settings (from below 0.30 to above 0.50 F1). However, contrary to expectations, manually annotated BM does not consistently outperform first dictionary entries, suggesting that definition length rather than theoretical fidelity may introduce noise. We also find that models perform best on cases with high annotator agreement and that verbs remain the most challenging part of speech. Overall, our findings highlight the importance of linguistically grounded modeling for metaphor detection in specialized domains.

Motifs are non-commonplace, recurring narrative elements, often found originally in folk stories and also in modern news, literature, and propaganda. Expressions of motifs in text can be most straightforwardly classified as simple or complex. Simple motif expressions are easy to detect because they almost always appear in a single sentence using the same words as the motif definition itself. However, complex motifs are strongly non-literal and often spread across multiple sentences, thus requiring more context to understand. We propose a baseline system to detect complex motif expressions that have challenged prior work. We used an annotated corpus that identified 992 complex motif expressions of 155 different motifs for training and testing. We tested five different generative approaches that included varying amounts of context: a single sentence baseline (from prior work); a window of 3 or 5 sentences; the entire story; or the entire story with the target sentence identified. We fine-tuned four off-the-shelf open-source LLMs using LoRA under these conditions. Somewhat surprisingly, we report a negative result: our experiments show that in our generative setup more context does not improve the detection of complex motifs. We speculate on why this might be so and identify directions for future research.

We analyze differences between human-written and automatically generated metaphors. Using two syntactically standardized datasets containing novel metaphors from poetry and science communication, we generate new figurative expressions with LLMs that describe the same concepts as human-written texts. Using crowdsourcing, we conduct extensive annotation across multiple dimensions (e.g., writing quality and creativity) and ask annotators to judge whether the metaphor was generated automatically. For the poetry set, we also asked annotators for the emotions conveyed by the metaphor. We find that, consistent with prior work, the authorship of scientific metaphors is difficult to determine. However, our results reveal that human-written poetic metaphors stand out by their capacity to convey emotion. We also analyze which types of metaphors are merely perceived as human. Finally, we show that, while human annotators cannot distinguish human from machine metaphors, automated approaches achieve high accuracy in identifying human writers, which suggests substantial differences in text structure.

Large Language Models (LLMs) can easily generate fluent text, but assessing whether they truly understand metaphors requires moving beyond English-centric datasets and binary token classification tasks. To test if current state-of-the-art models perform genuine structural alignment and analogical reasoning rather than just echoing statistical token co-occurrence, we introduce a new monolingual Spanish Natural Language Inference (NLI) dataset specifically built for metaphor interpretation. Using a Model-in-the-Loop approach, we reconstruct the literal truth conditions of metaphors sourced from science texts. Before human experts curated the data, we performed an ablation study—evaluated via BERTScore and Cross-Entropy—to test whether explicit symbolic scaffolding improves analogical reasoning. While automated evaluations suggested that forcing models to follow explicit metaphorical rules diminished their fluency and increased text surprisal, human evaluation revealed the opposite: this explicit guidance produced far more accurate and strictly literal outputs. This reveals a limitation in how we evaluate NLU: automated metrics consistently penalize the cognitive ‘heavy lifting’ required to resolve a metaphor, simply because they are built to reward surface-level statistical fluency. By releasing this resource, we aim to shift the focus from surface-level generation to real cognitive alignment and metaphorical understanding in Spanish NLU.

Metonymy, often considered as a figurative trope, is a frequently occurring linguistic phenomenon in which an entity is replaced by a semantically related entity. Named entities are commonly used to refer to associated concepts. For instance, in the sentence India signed a treaty, the geographical name India stands metonymically for the government rather than the physical location. This study develops a hybrid architecture to classify literal and metonymic usages of named entities in Marathi language using small data. The approach integrates Pustejovsky’s Generative Lexicon framework with linguistic features, including part-of-speech tags, named entity labels, and lemmas. The model is evaluated on 890 sentences and achieved F1 scores of 66.98% and 71.97% for literal and metonymic instances, respectively. The study highlights the effectiveness of the features in capturing metonymic contexts, though precision remains a target for improvement. Ablation results confirm that the Formal and Constitutive Qualia roles are the most critical components for detecting metonymic shifts, while the Telic role introduces modest noise in the present corpus. This experiment shows the scope for developing hybrid models for learning non-literal language using small data, which could be beneficial for less-explored and low-resource languages.

This work explores the connection between (non-)literalness and plausibility at the example of subject-verb-object events in English. We design a systematic setup of plausible and implausible event triples in combination with abstract and concrete constituent categories. Our analysis of human and LLM-generated judgments and example contexts reveals substantial differences between assessments of plausibility. While humans excel at nuanced detection and contextualization of (non-)literal vs. implausible events, LLM results reveal only shallow contextualization patterns with a bias to trade implausibility for non-literal, plausible interpretations.

In recent years, automatic metaphor detection has received considerable attention within NLP. However, the largest share of research, including most datasets annotated for metaphor, has concentrated on English and a limited set of genres. Automatic metaphor detection for a genre like religious online communication, which is particularly rich in metaphor, remains understudied, in particular since annotated data for this genre is lacking in the first place. This paper aims to close these gaps by offering a novel dataset of posts from German online forums annotated for metaphor, which opens up new research opportunities for automatic metaphor detection for German. Moreover, we present an in-depth exploration in which we evaluate the suitability of different strategies to overcome the relative lack of training data for this task by comparing cross-lingual and cross-genre transfer strategies with the use of LLM prompting. We find that fine-tuning encoder-only language models outperforms the prompting-based approach, that different architectures based on contextual embeddings indeed exhibit considerable differences in their behavior and that smaller in-genre data may be preferable for certain use cases over fine-tuning on larger datasets from different genres.

We introduce MetaphOrig, a small dataset comprising two genre-specific collections of spatial descriptions for the study of linguistic creativity and Non-Literal Expressions (NLEs). The sentence-level spatial descriptions were extracted from two distinct genre- and time-specific source corpora. Both source corpora comprise German texts: literary prose from the 18th to 20th century (KOLIMO) and factual travel reports from the 21st century (Wikivoyage). Along with the spatial descriptions, the dataset contains sentence-level originality ratings obtained through crowdsourcing and from four different LLMs (GPT-5, Qwen2.5-32B-Instruct, Mistral-Small-3.2-24B-Instruct, and Llama-3.2-3B), and word-level metaphor annotations. We provide the MetaphOrig datasets, including all annotations, to the community. The subsets can be used for further research on linguistic creativity or metaphor, either in one specific textual domain or comparatively across the two domains. We conduct an illustrative study on the dataset, treating originality as a proxy of textual creativity. In both subsets, we investigate potential correlations between sentence-level originality ratings and the density of metaphorical expressions within each sentence. We find the correlation to be present only in the KOLIMO subset. A comparison of human and LLM originality ratings shows that this pattern holds for both types of ratings.