Nurkhan Laiyk, Daniil Orel, Ayana Mussabayeva, Maiya Goloburda, Kamila Kuishibekova, Liya Goloburda, Diana Turmakhan, Preslav Nakov, Yuxia Wang, Fajri Koto. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhuohan Xie, Daniil Orel, Rushil Thareja, Dhruv Sahnan, Hachem Madmoun, Fan Zhang, Debopriyo Banerjee, Georgi Nenkov Georgiev, Xueqing Peng, Lingfei Qian, Jimin Huang, Jinyan Su, Aaryamonvikram Singh, Rui Xing, Rania Elbadry, Chen Xu, Haonan Li, Fajri Koto, Ivan Koychev, Tanmoy Chakraborty, Yuxia Wang, Salem Lahlou, Veselin Stoyanov, Sophia Ananiadou, Preslav Nakov. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Potentially idiomatic expressions (PIEs) carry meanings inherently tied to the everyday experience of a given language community. As such, they constitute an interesting challenge for assessing the linguistic (and to some extent cultural) capabilities of NLP systems. In this paper, we present XMPIE, a parallel multilingual and multimodal dataset of potentially idiomatic expressions. The dataset, containing 34 languages and over ten thousand items, allows comparative analyses of idiomatic patterns among language-specific realisations and preferences in order to gather insights about shared cultural aspects. This parallel dataset allows evaluation of language model performance for a given PIE in different languages and whether idiomatic understanding in one language can be transferred to another. Moreover, the dataset supports the study of PIEs across textual and visual modalities, to measure to what extent PIE understanding in one modality transfers or implies in understanding in another modality (text vs. image). The data was created by language experts, with both textual and visual components crafted under multilingual guidelines, and each PIE is accompanied by five images representing a spectrum from idiomatic to literal meanings, including semantically related and random distractors. The result is a high-quality benchmark for evaluating multilingual and multimodal idiomatic language understanding.

Nurkhan Laiyk, Daniil Orel, Rituraj Joshi, Maiya Goloburda, Yuxia Wang, Preslav Nakov, Fajri Koto. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We present DroidCollection 1 2 , the most extensive open data suite for training and evaluating machine-generated code detectors, comprising over a million code samples, seven programming languages, outputs from 43 coding models, and three real-world coding domains.Alongside fully AI-generated examples, our collection includes human-AI co-authored code, as well as adversarial examples explicitly crafted to evade detection.Subsequently, we develop DroidDetect, a suite of encoderonly detectors trained using a multi-task objective over DroidCollection.Our experiments show that existing detectors' performance fails to generalise to diverse coding domains and programming languages outside of their narrow training data.We further demonstrate that while most detectors are easily compromised by humanising the output distributions using superficial prompting and alignment approaches, this problem can be easily amended by training on a small number of adversarial examples.Finally, we demonstrate the effectiveness of metric learning and uncertainty-based resampling as way to enhance detector training on possibly noisy distributions.

Automatic Speech Recognition (ASR) models can achieve human parity, but their performance degrades significantly when used in noisy industrial and urban environments. In this paper, we present monolingual and multilingual ASR models, which can perform effectively even in extreme noise conditions. Specifically, we first generated a large synthetic noise-augmented dataset for Kazakh and English speech, and then used this data to train mono- and multilingual ASR models using state-of-the-art deep learning architectures. To evaluate our models, we have compared them to models trained on original data. The results showed that our models outperform the original ones in terms of word error rate (WER). For the monolingual case, the average WER of our model is 25.1%, while the original model yields 39.9% WER. For the multilingual case, we created a noise-robust customization of the Whisper model, which has a significantly improved performance on Kazakh (for both original and noisy data) and resulted on average 10% performance improvement on noisy English audios. Finally, we tested our models in an industrial setting, which has proven that our models are robust to unseen types of noise and can be used for real-life applications.