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.

Euphemisms are words or phrases used to soften or indirectly refer to taboo or sensitive topics. They pose interpretation challenges because the same expression may appear in different senses depending on context: literal, figurative but non-euphemistic, or euphemistic. For example, pull the plug may refer euphemistically to ending a patient’s life support, figuratively to canceling a project or funding, or literally to unplugging a device. Euphemisms also vary across languages and cultures in both their surface forms and the contexts in which they are conventionally used. Previous work introduced datasets for the computational study of euphemisms in five languages. We extend this line of work by introducing two new annotated datasets for euphemism detection in Polish and Ukrainian and by standardizing resources for all seven languages into a unified benchmark format that supports cross-lingual evaluation. Finally, we provide zero-shot and few-shot baselines using GPT-5-nano. We ran each configuration five times and report the average score, establishing reference scores for multilingual pragmatic understanding. In addition, we performed pilot tests using Qwen3-4B on the English and Chinese datasets.

Child-directed Speech (CDS) has been shown to better support language learning as training data for computational models. Artificially generated input aims at replicating the advantage of CDS by re-creating targeted linguistic properties. Recently, the use of questions in CDS has been suggested as a linguistic property that may entail an effective discourse structure for model training. However, previous work has shown inconsistent improvement over baseline using questions in training data. In this study, we propose a new question generation method that aligns both the generation prompts and sampling methods with properties of CDS. We show that prompt wording substantially changes whether synthetic questions match CDS on surface properties such as MLU and question type. Despite marked improvements over baseline, enhanced CDS-likeness does not translate into consistent downstream gains. Overall, our results show that the role of questions in training data is a topic worth looking further into.

Child-Directed Speech (CDS) holds unique linguistic properties that distinguish it from other types of textual corpora. Language models trained using CDS often obtain superior results compared with the same size of different types of data. Several studies have aimed at modifying non-CDS data to mimic its linguistic properties to match the hypothesized advantageous aspects of CDS. Here, we propose to adapt the non-CDS portions of the training data to include questions similar to CDS interaction. We modify the data by adding artificially generated questions to the data and methodically analyzing the change in performance using each modified dataset. Our results show that artificial question generation strongly depends on the properties of the original dataset. While the performance improves for question-related measures, the overall performance is negatively affected as a result of the reduced syntactic diversity.