arXiv subjects

Paolo Rosso

Publications and source records attributed to Paolo Rosso.

At least 19 recordsLinked to original sources

arXiv2026

Deconstructing Stereotypes: Scope-Conditioned Generation for Effective Multilingual Counterspeech

Counterspeech (CS) - direct responses that counter online Hate Speech (HS) using reasoning and alternative viewpoints - has emerged as an alternative to content removal. Current automatic CS generation methods, however, frequently produce generic, ineffective replies that fail to target the implicit stereotypes behind HS. To bridge this gap, we propose a novel scope-conditioned generation framework that explicitly integrates structured stereotype characteristics into Large Language Models prompts. We validate our approach on a novel, human-curated dataset annotated in English, Italian, and Spanish. Extensive evaluations show that stereotype-conditioned prompting substantially outperforms generic baselines across all three languages, obtaining significant gains in factuality, specificity, cogency, and effectiveness for both explicit and implicit implied stereotypes.

cs.CL↗

arXiv2026

Human Values in a Single Sentence: Moral Presence, Hierarchies, and Transformer Ensembles on the Schwartz Continuum

We study neural multi-label classification under severe label imbalance through sentence-level detection of the 19 refined Schwartz human values in 74k English news and manifesto sentences (ValueEval'24 corpus). Each sentence carries a roughly balanced moral-presence label and a 19-way value annotation. First, moral presence is learnable from single sentences: a DeBERTa-base classifier reaches positive-class $F_1 \approx 0.73$ at the default threshold, which calibration does not improve. Second, comparing direct multi-label detectors with presence-gated hierarchies under an 8 GB consumer-grade GPU budget, we find that gating does not improve over direct prediction, as gate recall becomes a bottleneck. Third, studying lightweight auxiliary signals and small ensembles, we isolate decision-threshold calibration as a decisive, often overlooked factor: a standard text-only baseline already matches the best official ValueEval'24 English run at the default threshold (macro-$F_1 = 0.282$ vs. $\approx 0.28$), and tuning the threshold on validation alone raises it to $0.315$, most of our overall gain. Lightweight features do not survive a paired per-seed test; a soft-voting ensemble reaches our best macro-$F_1 = 0.332$. Calibration is not architecture-specific: it reproduces on RoBERTa-base, where the gain is larger ($+0.043$). To our knowledge, this is the first systematic comparison of direct and presence-gated architectures, lightweight feature-augmented encoders, and instruction-tuned Large Language Models (LLMs) at sentence level; benchmarked 7-9B LLMs (zero-/few-shot and QLoRA) lag behind the supervised ensemble under the same budget. We provide empirical guidance for compute-efficient, value-aware NLP models.

cs.CL↗

arXiv2026

More Context, Larger Models, or Moral Knowledge? A Systematic Study of Schwartz Value Detection in Political Texts

Detecting Schwartz values in political texts is hard: cues are often implicit, and neighboring values differ by fine distinctions. Two remedies are widely assumed to help: more surrounding document text, and explicit moral knowledge. Knowledge-based retrieval has improved benchmarks elsewhere, but whether either transfers here is untested, because published systems vary context, knowledge, and model family at once. We separate these factors under matched conditions on the ValuesML/Touché ValueEval format. The input ranges from the target sentence to a local window to the full document. Retrieval is either absent or drawn from a curated moral knowledge base, injected by early, late, or cross-attention fusion. Supervised DeBERTa-v3 encoders are compared against zero-shot LLMs from 12B to 123B. More context is not uniformly better: full-document input improves the encoders by 2.5-3.8 macro-F1 points but does not consistently help the LLMs. Retrieved knowledge helps more reliably, improving every model family and context under early fusion. A control substituting random knowledge-base entries shows the families gain differently: LLMs from relevance, encoders from exposure to the value ontology. Neither larger encoders nor larger LLMs guarantee gains, and early fusion outperforms both trainable variants. Value-sensitive NLP should evaluate context, knowledge, and model family jointly.

cs.CL↗

arXiv2026

Zoom In Disparities in Healthcare LLM Q&A

Equitable access to reliable health information is vital when integrating AI into healthcare. Yet, information quality varies across languages, raising concerns about the reliability and consistency of multilingual Large Language Models (LLMs). We systematically examine cross-lingual disparities in pre-training source and factuality alignment in LLM answers for multilingual healthcare Q&A across English, German, Turkish, Chinese (Mandarin), and Italian. We (i) constructed Multilingual Wiki Health Care (MultiWikiHealthCare), a multilingual dataset from Wikipedia; (ii) analyzed cross-lingual healthcare coverage; (iii) assessed LLM response alignment with these references; and (iv) conducted a case study on factual alignment through the use of contextual information and Retrieval-Augmented Generation (RAG). Our findings reveal substantial cross-lingual disparities in both Wikipedia coverage and LLM factual alignment. Across LLMs, responses align more with English Wikipedia, even when the prompts are non-English. Providing contextual excerpts from non-English Wikipedia at inference time effectively shifts factual alignment toward culturally relevant knowledge. These results highlight practical pathways for building more equitable, multilingual AI systems for healthcare.

cs.CL↗

arXiv2026

Beyond Independent Labels: Schwartz-Geometry Decoding for Human Value Detection

Human value detection is commonly formulated as sentence-level multi-label classification over the 19 refined Schwartz values, typically predicted as independent labels. Schwartz theory, however, describes them as a circular motivational continuum, in which adjacent values are compatible and opposing values are in tension. We ask whether this structure can be operationalized as an explicit output-space geometry and used as a soft bias rather than a hard constraint. On a DeBERTa-v3-base classifier, we compare two ways of injecting it: training-time geometry-aware objectives and a post-hoc Schwartz-aware energy decoder that scores whole label sets jointly. Across five seeds, training-time geometry gives only limited gains-no larger for the true continuum than for a random ordering-whereas the decoder makes label sets more coherent with the continuum-on theory-aware coherence metrics we introduce-at no cost to Macro-F1 or Micro-F1 (held fixed by its selection rule). The gain is specific to the true Schwartz ordering: it does not appear for a random permutation or an empirical co-occurrence graph through the identical decoder. A bounded Qwen2.5-72B-Instruct diagnostic shows that supplying the continuum at inference shifts behavior but does not match supervised structured prediction. Theory-aware decoding thus offers a lightweight, controllable way to make value detection faithful to its label space.

cs.CL↗

arXiv2026

Do Schwartz Higher-Order Values Help Sentence-Level Human Value Detection? A Study of Hierarchical Gating and Calibration

Human value detection from single sentences is a sparse, imbalanced multi-label task. We study whether Schwartz higher-order (HO) categories help this setting on ValueEval'24 / ValuesML (74K English sentences) under a compute-frugal budget. Rather than proposing a new architecture, we compare direct supervised transformers, hard HO$\rightarrow$values pipelines, Presence$\rightarrow$HO$\rightarrow$values cascades, compact instruction-tuned large language models (LLMs), QLoRA, and low-cost upgrades such as threshold tuning and small ensembles. HO categories are learnable: the easiest bipolar pair, Growth vs. Self-Protection, reaches Macro-$F_1=0.58$. The most reliable gains come from calibration and ensembling: threshold tuning improves Social Focus vs. Personal Focus from $0.41$ to $0.57$ ($+0.16$), transformer soft voting lifts Growth from $0.286$ to $0.303$, and a Transformer+LLM hybrid reaches $0.353$ on Self-Protection. In contrast, hard hierarchical gating does not consistently improve the end task. Compact LLMs also underperform supervised encoders as stand-alone systems, although they sometimes add useful diversity in hybrid ensembles. Under this benchmark, the HO structure is more useful as an inductive bias than as a rigid routing rule.

cs.CL↗

arXiv2026

Human-Centered Multimodal Fusion for Sexism Detection in Memes with Eye-Tracking, Heart Rate, and EEG Signals

The automated detection of sexism in memes is a challenging task due to multimodal ambiguity, cultural nuance, and the use of humor to provide plausible deniability. Content-only models often fail to capture the complexity of human perception. To address this limitation, we introduce and validate a human-centered paradigm that augments standard content features with physiological data. We created a novel resource by recording Eye-Tracking (ET), Heart Rate (HR), and Electroencephalography (EEG) from 16 subjects (8 per experiment) while they viewed 3984 memes from the EXIST 2025 dataset. Our statistical analysis reveals significant physiological differences in how subjects process sexist versus non-sexist content. Sexist memes were associated with higher cognitive load, reflected in increased fixation counts and longer reaction times, as well as differences in EEG spectral power across the Alpha, Beta, and Gamma bands, suggesting more demanding neural processing. Building on these findings, we propose a multimodal fusion model that integrates physiological signals with enriched textual-visual features derived from a Vision-Language Model (VLM). Our final model achieves an AUC of 0.794 in binary sexism detection, a statistically significant 3.4% improvement over a strong VLM-based baseline. The fusion is particularly effective for nuanced cases, boosting the F1-score for the most challenging fine-grained category, Misogyny and Non-Sexual Violence, by 26.3%. These results show that physiological responses provide an objective signal of perception that enhances the accuracy and human-awareness of automated systems for countering online sexism.

cs.HC↗

arXiv2026

MEMEWEAVER: Inter-Meme Graph Reasoning for Sexism and Misogyny Detection

Women are twice as likely as men to face online harassment due to their gender. Despite recent advances in multimodal content moderation, most approaches still overlook the social dynamics behind this phenomenon, where perpetrators reinforce prejudices and group identity within like-minded communities. Graph-based methods offer a promising way to capture such interactions, yet existing solutions remain limited by heuristic graph construction, shallow modality fusion, and instance-level reasoning. In this work, we present MemeWeaver, an end-to-end trainable multimodal framework for detecting sexism and misogyny through a novel inter-meme graph reasoning mechanism. We systematically evaluate multiple visual--textual fusion strategies and show that our approach consistently outperforms state-of-the-art baselines on the MAMI and EXIST benchmarks, while achieving faster training convergence. Further analyses reveal that the learned graph structure captures semantically meaningful patterns, offering valuable insights into the relational nature of online hate.

cs.AI↗

arXiv2025

Transformers and Ensemble methods: A solution for Hate Speech Detection in Arabic languages

This paper describes our participation in the shared task of hate speech detection, which is one of the subtasks of the CERIST NLP Challenge 2022. Our experiments evaluate the performance of six transformer models and their combination using 2 ensemble approaches. The best results on the training set, in a five-fold cross validation scenario, were obtained by using the ensemble approach based on the majority vote. The evaluation of this approach on the test set resulted in an F1-score of 0.60 and an Accuracy of 0.86.

cs.CL↗

arXiv2025

Efficient Out-of-Scope Detection in Dialogue Systems via Uncertainty-Driven LLM Routing

Out-of-scope (OOS) intent detection is a critical challenge in task-oriented dialogue systems (TODS), as it ensures robustness to unseen and ambiguous queries. In this work, we propose a novel but simple modular framework that combines uncertainty modeling with fine-tuned large language models (LLMs) for efficient and accurate OOS detection. The first step applies uncertainty estimation to the output of an in-scope intent detection classifier, which is currently deployed in a real-world TODS handling tens of thousands of user interactions daily. The second step then leverages an emerging LLM-based approach, where a fine-tuned LLM is triggered to make a final decision on instances with high uncertainty. Unlike prior approaches, our method effectively balances computational efficiency and performance, combining traditional approaches with LLMs and yielding state-of-the-art results on key OOS detection benchmarks, including real-world OOS data acquired from a deployed TODS.

cs.CL↗

arXiv2025

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks about the DANA in Valencia

This study investigates the dissemination of disinformation on social media platforms during the DANA event (DANA is a Spanish acronym for Depresion Aislada en Niveles Altos, translating to high-altitude isolated depression) that resulted in extremely heavy rainfall and devastating floods in Valencia, Spain, on October 29, 2024. We created a novel dataset of 650 TikTok and X posts, which was manually annotated to differentiate between disinformation and trustworthy content. Additionally, a Few-Shot annotation approach with GPT-4o achieved substantial agreement (Cohen's kappa of 0.684) with manual labels. Emotion analysis revealed that disinformation on X is mainly associated with increased sadness and fear, while on TikTok, it correlates with higher levels of anger and disgust. Linguistic analysis using the LIWC dictionary showed that trustworthy content utilizes more articulate and factual language, whereas disinformation employs negations, perceptual words, and personal anecdotes to appear credible. Audio analysis of TikTok posts highlighted distinct patterns: trustworthy audios featured brighter tones and robotic or monotone narration, promoting clarity and credibility, while disinformation audios leveraged tonal variation, emotional depth, and manipulative musical elements to amplify engagement. In detection models, SVM+TF-IDF achieved the highest F1-Score, excelling with limited data. Incorporating audio features into roberta-large-bne improved both Accuracy and F1-Score, surpassing its text-only counterpart and SVM in Accuracy. GPT-4o Few-Shot also performed well, showcasing the potential of large language models for automated disinformation detection. These findings demonstrate the importance of leveraging both textual and audio features for improved disinformation detection on multimodal platforms like TikTok.

cs.CL↗

arXiv2025

Do LLMs Provide Consistent Answers to Health-Related Questions across Languages?

Equitable access to reliable health information is vital for public health, but the quality of online health resources varies by language, raising concerns about inconsistencies in Large Language Models (LLMs) for healthcare. In this study, we examine the consistency of responses provided by LLMs to health-related questions across English, German, Turkish, and Chinese. We largely expand the HealthFC dataset by categorizing health-related questions by disease type and broadening its multilingual scope with Turkish and Chinese translations. We reveal significant inconsistencies in responses that could spread healthcare misinformation. Our main contributions are 1) a multilingual health-related inquiry dataset with meta-information on disease categories, and 2) a novel prompt-based evaluation workflow that enables sub-dimensional comparisons between two languages through parsing. Our findings highlight key challenges in deploying LLM-based tools in multilingual contexts and emphasize the need for improved cross-lingual alignment to ensure accurate and equitable healthcare information.

cs.CL↗

arXiv2025

"Stupid robot, I want to speak to a human!" User Frustration Detection in Task-Oriented Dialog Systems

Detecting user frustration in modern-day task-oriented dialog (TOD) systems is imperative for maintaining overall user satisfaction, engagement, and retention. However, most recent research is focused on sentiment and emotion detection in academic settings, thus failing to fully encapsulate implications of real-world user data. To mitigate this gap, in this work, we focus on user frustration in a deployed TOD system, assessing the feasibility of out-of-the-box solutions for user frustration detection. Specifically, we compare the performance of our deployed keyword-based approach, open-source approaches to sentiment analysis, dialog breakdown detection methods, and emerging in-context learning LLM-based detection. Our analysis highlights the limitations of open-source methods for real-world frustration detection, while demonstrating the superior performance of the LLM-based approach, achieving a 16\% relative improvement in F1 score on an internal benchmark. Finally, we analyze advantages and limitations of our methods and provide an insight into user frustration detection task for industry practitioners.

cs.CL↗

arXiv2024

What distinguishes conspiracy from critical narratives? A computational analysis of oppositional discourse

The current prevalence of conspiracy theories on the internet is a significant issue, tackled by many computational approaches. However, these approaches fail to recognize the relevance of distinguishing between texts which contain a conspiracy theory and texts which are simply critical and oppose mainstream narratives. Furthermore, little attention is usually paid to the role of inter-group conflict in oppositional narratives. We contribute by proposing a novel topic-agnostic annotation scheme that differentiates between conspiracies and critical texts, and that defines span-level categories of inter-group conflict. We also contribute with the multilingual XAI-DisInfodemics corpus (English and Spanish), which contains a high-quality annotation of Telegram messages related to COVID-19 (5,000 messages per language). We also demonstrate the feasibility of an NLP-based automatization by performing a range of experiments that yield strong baseline solutions. Finally, we perform an analysis which demonstrates that the promotion of intergroup conflict and the presence of violence and anger are key aspects to distinguish between the two types of oppositional narratives, i.e., conspiracy vs. critical.

cs.CL↗

arXiv2024

RoCode: A Dataset for Measuring Code Intelligence from Problem Definitions in Romanian

Recently, large language models (LLMs) have become increasingly powerful and have become capable of solving a plethora of tasks through proper instructions in natural language. However, the vast majority of testing suites assume that the instructions are written in English, the de facto prompting language. Code intelligence and problem solving still remain a difficult task, even for the most advanced LLMs. Currently, there are no datasets to measure the generalization power for code-generation models in a language other than English. In this work, we present RoCode, a competitive programming dataset, consisting of 2,642 problems written in Romanian, 11k solutions in C, C++ and Python and comprehensive testing suites for each problem. The purpose of RoCode is to provide a benchmark for evaluating the code intelligence of language models trained on Romanian / multilingual text as well as a fine-tuning set for pretrained Romanian models. Through our results and review of related works, we argue for the need to develop code models for languages other than English.

cs.CL↗

arXiv2024

Toxic language detection: a systematic review of Arabic datasets

The detection of toxic language in the Arabic language has emerged as an active area of research in recent years, and reviewing the existing datasets employed for training the developed solutions has become a pressing need. This paper offers a comprehensive survey of Arabic datasets focused on online toxic language. We systematically gathered a total of 54 available datasets and their corresponding papers and conducted a thorough analysis, considering 18 criteria across four primary dimensions: availability details, content, annotation process, and reusability. This analysis enabled us to identify existing gaps and make recommendations for future research works. For the convenience of the research community, the list of the analysed datasets is maintained in a GitHub repository (https://github.com/Imene1/Arabic-toxic-language).

cs.CL↗

arXiv2024

Reading Between the Frames: Multi-Modal Depression Detection in Videos from Non-Verbal Cues

Depression, a prominent contributor to global disability, affects a substantial portion of the population. Efforts to detect depression from social media texts have been prevalent, yet only a few works explored depression detection from user-generated video content. In this work, we address this research gap by proposing a simple and flexible multi-modal temporal model capable of discerning non-verbal depression cues from diverse modalities in noisy, real-world videos. We show that, for in-the-wild videos, using additional high-level non-verbal cues is crucial to achieving good performance, and we extracted and processed audio speech embeddings, face emotion embeddings, face, body and hand landmarks, and gaze and blinking information. Through extensive experiments, we show that our model achieves state-of-the-art results on three key benchmark datasets for depression detection from video by a substantial margin. Our code is publicly available on GitHub.

cs.CV↗

arXiv2023

Vicinal Risk Minimization for Few-Shot Cross-lingual Transfer in Abusive Language Detection

Cross-lingual transfer learning from high-resource to medium and low-resource languages has shown encouraging results. However, the scarcity of resources in target languages remains a challenge. In this work, we resort to data augmentation and continual pre-training for domain adaptation to improve cross-lingual abusive language detection. For data augmentation, we analyze two existing techniques based on vicinal risk minimization and propose MIXAG, a novel data augmentation method which interpolates pairs of instances based on the angle of their representations. Our experiments involve seven languages typologically distinct from English and three different domains. The results reveal that the data augmentation strategies can enhance few-shot cross-lingual abusive language detection. Specifically, we observe that consistently in all target languages, MIXAG improves significantly in multidomain and multilingual environments. Finally, we show through an error analysis how the domain adaptation can favour the class of abusive texts (reducing false negatives), but at the same time, declines the precision of the abusive language detection model.

cs.CL↗