Papers
My research spans Responsible AI and NLP, with a focus on making language technologies more transparent, equitable, and socially impactful. Core themes include explainability of model decisions, efficient cross-lingual knowledge transfer, and building inclusive multilingual datasets and systems. I apply these to socially critical domains — including hate-speech, sexism, and fake-news detection. Most recently, I focus on rigorous evaluation of LLMs, including their reasoning beyond high-resource languages. Beyond methodology, I contribute position papers and organise shared-task competitions. Several of these works have received significant recognition from the research community 🏆. The full, always up-to-date list is on Google Scholar.
From Specialization to Generalization: Instruction-tuned LLMs for Robust Harmful Content Mitigation
Instruction-tuned LLMs for harmful content mitigation: moving from task-specific models towards models that generalise robustly across harmful content tasks.
PluraMath: Extending Mathematical Reasoning Evaluation Beyond High-Resource Languages
A benchmark that extends the evaluation of mathematical reasoning in LLMs beyond high-resource languages.
Beyond Generic Responses: Target-Aware Strategies for Countering Hate Speech
Moving counter-speech generation beyond generic replies: strategies that take the target of the hate speech into account.
Explainable Semantic Textual Similarity via Dissimilar Span Detection
More in-depth, explainable evaluation of semantic textual similarity by detecting the spans in which two texts differ.
NLP for Social Good: A Survey of Challenges, Opportunities, and Responsible Deployment
Our massive position paper and survey of application opportunities of modern language technologies for social positive impact.
Self-Explaining Hate Speech Detection with Moral Rationales
This paper introduces SMRA (Supervised Moral Rationale Attention), a hate speech detection framework that moves beyond surface-level lexical patterns by grounding model attention in moral reasoning. Drawing on Moral Foundations Theory, it trains models to focus on morally relevant text spans using expert-annotated rationales as direct supervision, making predictions both more accurate and inherently interpretable.
Low-Resource, High-Impact: Building Corpora for Inclusive Language Technologies
This tutorial is designed for NLP practitioners, researchers, and developers working with multilingual and low-resource languages who seek to create more equitable and socially impactful language technologies.
CrossNews-UA: A Cross-lingual News Semantic Similarity Benchmark for Ukrainian, Polish, Russian, and English
This paper introduces a scalable crowdsourcing pipeline for cross-lingual fake news detection, along with CrossNews-UA, a new dataset pairing Ukrainian news with Polish, Russian, and English sources annotated for semantic similarity using a structured 4W framework.
Findings of the WMT 2025 Shared Task LLMs with Limited Resources for Slavic Languages: MT and QA
We present the findings of the WMT 2025 Shared Task LLMs with Limited Resources for Slavic Languages. This shared task focuses on training LLMs using limited data and compute resources for three Slavic languages: Upper Sorbian (hsb), Lower Sorbian (dsb), and Ukrainian (uk). We consider two tasks which are to be evaluated jointly: Machine Translation (MT) and Multiple-Choice Question Answering (QA).
EmoBench‑UA: A Benchmark Dataset for Emotion Detection in Ukrainian
In this work, we introduce EMOBENCH-UA, the first of its kind annotated dataset for emotion detection in Ukrainian texts.
Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval with Limited Labeled Data
This paper tackles hate speech detection in low-resource languages by combining cross-lingual transfer learning with nearest-neighbor retrieval. Starting from a small set of labeled examples in the target language, the method retrieves the most relevant examples from a large multilingual pool to augment training data. Tested across eight languages, the approach consistently outperforms models trained on target-language data alone, often beats the current state-of-the-art, and remains highly efficient — sometimes needing as few as 200 retrieved instances — while being easy to scale and adapt to new languages.
QiMP at SemEval-2025 Task 11: Optimizing Text-based Emotion Classification in English Beyond Traditional Methods
Our solution of emotion detection in English. The team was of bachelor students who overcame many internation teams and ended up in top third of the leaderboard.
DETOXIFY-IT: An Italian Parallel Dataset for Text Detoxification
Creation of the parallel text detoxification corpus for Italian
Can Prompting LLMs Unlock Hate Speech Detection across Languages? A Zero-shot and Few-shot Study
This paper evaluates multilingual LLMs (LLaMA, Aya, Qwen, BloomZ) for hate speech detection across eight non-English languages using zero-shot and few-shot prompting, comparing them against fine-tuned encoder models.
HatePRISM: Policies, Platforms, and Research Integration. Advancing NLP for Hate Speech Proactive Mitigation
Comprehensive examination of hate speech regulations and strategies from three angles: country regulations, social platform policies, and NLP research datasets. Reveals inconsistencies and proposes research directions for unified mitigation frameworks.
BRIGHTER: BRIdging the Gap in Human-Annotated Textual Emotion Recognition Datasets for 28 Languages
Multi-labeled, emotion-annotated datasets across 28 languages, primarily covering low-resource languages from Africa, Asia, Eastern Europe, and Latin America. Highlights challenges in cross-lingual emotion recognition and benchmarks LLM-based approaches. ACL Best Resource Paper 🏆
Multilingual and Explainable Text Detoxification with Parallel Corpora
Extends parallel text detoxification corpora to German, Chinese, Arabic, Hindi, and Amharic. Conducts the first automated, explainable analysis of toxic vs. non-toxic sentence features across 9 languages, and introduces a Chain-of-Thought–inspired detoxification method.
Cross-lingual Text Classification Transfer: The Case of Ukrainian
Explores cross-lingual knowledge transfer methods for Ukrainian NLP — covering toxicity, formality classification, and NLI — using multilingual encoders, translation systems, LLMs, and language adapters without manual data curation.
Overview of the Multilingual Text Detoxification Task at PAN 2025
We extend our previous Multilingual Text Detoxification (TextDetox) task to new languages—Italian, French, Hebrew, Hinglish, Japanese, and Tatar—suggesting participants to participate in multi-lingual and cross-lingual text detoxification challenges.
Overview of the Multilingual Text Detoxification Task at PAN 2024
Shared task overview for multilingual text detoxification. Presents the task setup, parallel corpora, evaluation protocol, and analysis of participating system submissions across multiple languages.
Crafting Tomorrow's Headlines: Neural News Generation and Detection in English, Turkish, Hungarian, and Persian
Benchmark dataset for neural news detection across four languages using multilingual generators (BloomZ, LLaMa-2, Mistral, Mixtral, GPT-4). Studies interpretability and robustness of machine-generated text detectors.
MultiParaDetox: Extending Text Detoxification with Parallel Data to New Languages
Extending parallel data creation pipeline for text style transfer and text detoxification easily to new languages.
Toxicity Classification in Ukrainian
Creation the first of its kind toxicity classification corpus for Ukrainian.
ChatGPT for good? On opportunities and challenges of large language models for education
This position paper examines the promise and pitfalls of applying large language models in education, highlighting benefits like personalized learning and content creation alongside challenges around bias, misuse, and the need for new digital literacies.
Adam-Smith at SemEval-2023 Task 4: Discovering Human Values in Arguments with Ensembles of Transformer-based Models
This paper presents the best-performing approach alias “Adam Smith” for the SemEval-2023 Task 4:“Identification of Human Values behind Arguments”. The goal of the task was to create systems that automatically identify the values within textual arguments.
AdamR at SemEval-2023 Task 10: Solving the Class Imbalance Problem in Sexism Detection with Ensemble Learning
Our team experimented with different ways to combat class imbalance throughout the tasks using data augmentation and loss alteration techniques. We tackled the challenge by utilising ensembles of Transformer models trained on different datasets, which are tested to find the balance between performance and interpretability.
Multiverse: Multilingual Evidence for Fake News Detection
This paper proposes Multiverse, a feature that leverages cross-lingual evidence to improve fake news detection beyond single-language approaches.
Detecting Text Formality: A Study of Text Classification Approaches
First systematic study of formality detection methods using statistical, neural, and Transformer-based approaches. Evaluated on GYAFC and X-FORMAL datasets across multiple languages, delivering best-performing public models.
Exploring Methods for Cross-lingual Text Style Transfer: The Case of Text Detoxification
Comprehensive comparison of cross-lingual transfer methods for text style transfer.
IFAN: An Explainability-Focused Interaction Framework for Humans and NLP Models
We propose IFAN, a framework for real-time explanation-based interaction and improvement of NLP classification models.
ParaDetox: Detoxification with Parallel Data
Introduces a novel pipeline for collecting parallel data for text detoxification. Presents over 10,000 non-toxic English paraphrases, demonstrating state-of-the-art results for seq2seq detoxification models.
Exploring Cross-lingual Text Detoxification with Large Multilingual Language Models
Investigates multilingual LLMs for cross-lingual text detoxification, exploring whether knowledge from high-resource languages can improve detoxification quality in low-resource settings.
RUSSE-2022: Findings of the First Russian Detoxification Shared Task Based on Parallel Corpora
The first of its kind text detoxification shared task.
Cross-lingual Evidence Improves Monolingual Fake News Detection
Demonstrates that cross-lingual evidence significantly improves monolingual fake news detection. True news repeats consistent facts across languages while misinformation diverges, providing a cross-lingual signal for verification.
Methods for Detoxification of Texts for the Russian Language
Introduces condBERT: a new method for text style transfer. Also, the first text detoxification methods for Russian, a morphologically rich language with limited resources. Presents evaluation benchmarks and baseline models for Russian-language detoxification.