Meme-based social abuse detection is challenging because harmful intent often relies on implicit cultural symbolism and subtle cross-modal incongruence. Prior approaches, from fusion-based methods to in-context learning with Large Vision-Language Models (LVLMs), have made progress but remain limited by three factors: i) cultural blindness (missing symbolic context), ii) boundary ambiguity (satire vs. abuse confusion), and iii) lack of interpretability (opaque model reasoning). We introduce CROSS-ALIGN+, a three-stage framework that systematically addresses these limitations: (1) Stage I mitigates cultural blindness by enriching multimodal representations with structured knowledge from ConceptNet, Wikidata, and Hatebase; (2) Stage II reduces boundary ambiguity through parameter-efficient LoRA adapters that sharpen decision boundaries; and (3) Stage III enhances interpretability by generating cascaded explanations. Extensive experiments on five benchmarks and eight LVLMs demonstrate that CROSS-ALIGN+ consistently outperforms state-of-the-art methods, achieving up to 17% relative F1 improvement while providing interpretable justifications for each decision.

Cultural context profoundly shapes how people interpret online content, yet vision-language models (VLMs) remain predominantly trained through Western or English-centric lenses. This limits their fairness and cross-cultural robustness in tasks like hateful meme detection. We introduce a systematic evaluation framework designed to diagnose and quantify the cross-cultural robustness of state-of-the-art VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection. Results show that the common ``translate-then-detect'' approach deteriorate performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Our findings reveal systematic convergence toward Western safety norms and provide actionable strategies to mitigate such bias, guiding the design of globally robust multimodal moderation systems.

The emergence of social media has led memes to become a powerful mode of communication, blending text, images, and emojis. However, this surge in meme usage has also seen a rise in offensive material. With manual content moderation proving impractical due to the sheer volume of data, there's a pressing need for automated methods to identify harmful memes. Yet, existing research predominantly targets high-resource languages such as English, neglecting low-resource ones like Nepali. To bridge this gap, we introduce the first Nepali meme dataset annotated for hate speech and sentiment. Our contributions are threefold: (1) We create and release NeMeme, a unique dataset featuring Nepali and code-mixed Nepali memes (combining Nepali and English). (2) We evaluate NeMeme using cutting-edge unimodal and multimodal models to establish initial performance benchmarks. (3) We introduce MemeNePAL, a novel multimodal framework employing prompt-assisted learning to effectively categorize Nepali memes. MemeNePAL overcomes the shortcomings of prior state-of-the-art (SOTA) techniques, which were designed for high-resource languages and struggle with Nepali's linguistic differences and cultural subtleties. This work not only promotes inclusivity in content moderation research but also aligns with UN Sustainable Development Goals such as promoting well-being, reducing inequalities, and fostering peace. We adhere to FAIR principles by making the dataset publicly available.

Sequential recommender systems (SRSs) aim to recommend the next items to well match users' preferences. In addition to recommendation accuracy, diversity is another critical aspect in evaluating SRSs. Recently, the emerging diffusion models (DMs) have been widely adopted in SRSs. Their employed learning-to-generate paradigm allows them to cover a much broader range of users' preferences and thus generate more diversified items. However, existing DM-based SRSs still face two significant gaps that prevent them from further improving the recommendation diversity: (1) they often rely on non-diversified users' preferences as guidance to direct the training of diffusion networks, restricting networks' ability to generate diverse items; and (2) they are based on a homogeneous diffusion inference mechanism to generate the next items and thus can only accommodate users' major preferences. Such a practice neglects users' heterogeneous preferences towards various types of items, further limiting recommendation diversity. To bridge these two critical gaps and to further unleash the potential of DMs in enhancing the recommendation diversity of SRSs, we propose a novel diversity-guided diffusion model for sequential recommendations, called DiffDiv for short. To be specific, first, a new diversity-aware guidance learning mechanism is devised to direct the training of DMs to effectively capture users' diversified preferences from their historical interactions. Then, a novel heterogeneous diffusion inference mechanism is designed to generate diversified items to accommodate users' heterogeneous preferences, further boosting the recommendation diversity. Extensive experiments on real-world datasets validate the effectiveness of DiffDiv in terms of both recommendation accuracy and diversity.

Frequent fluctuations of client nodes in highly dynamic mobile clusters can lead to significant changes in feature space distribution and data drift, posing substantial challenges to the robustness of existing federated learning (FL) strategies. To address these issues, we proposed a mobile cluster federated learning framework (MoCFL). MoCFL enhances feature aggregation by introducing an affinity matrix that quantifies the similarity between local feature extractors from different clients, addressing dynamic data distribution changes caused by frequent client churn and topology changes. Additionally, MoCFL integrates historical and current feature information when training the global classifier, effectively mitigating the catastrophic forgetting problem frequently encountered in mobile scenarios. This synergistic combination ensures that MoCFL maintains high performance and stability in dynamically changing mobile environments. Experimental results on the UNSW-NB15 dataset show that MoCFL excels in dynamic environments, demonstrating superior robustness and accuracy while maintaining reasonable training costs.

Memes serve as a powerful medium of expression in the digital age, shaping cultural discourse and conveying ideas succinctly and engagingly. However, their potential for social abuse highlights the importance of developing effective methods to detect harmful content within memes. Recent studies on memes have focused on transforming images into textual captions using large language models (LLMs). However, these approaches often result in non-informative captions. Furthermore, previous methods have only been tested on limited datasets, providing insufficient evidence of their robustness. To address these limitations, we present a multimodal, agent-based framework designed to generate informative visual descriptions of memes by asking insightful questions to improve visual descriptions in zero-shot visual question-answering settings. Specifically, we leverage an LLM as agents with distinct roles and a large multimodal model (LMM) as a vision expert. These agents first analyze the images and then ask informative questions related to potential social abuse in memes to obtain high-quality answers about the images. Through continuous discussion guided by instructional prompts, the agents gather high-quality information while repeatedly acquiring image data from the LMM, which helps detect social abuse in memes. Results on a dataset of 6,626 memes across 5 tasks show our framework surpasses state-of-the-art methods, demonstrating strong generalizability and improved detection of social abuse in memes.

User-generated content on social media platforms provides a valuable resource for developing automated computational methods to detect mental health issues online leading to suicidal thoughts automatically. Although current fully automated methods show promise, they may produce uncertain predictions, leading to flawed conclusions. To address this, we propose a novel model called DiGrI, or Distorted Greedy Approach for Human-Assisted Online Suicide Ideation Detection, which reformulates suicide ideation assessment as a selective, prioritized prediction problem. The model incorporates a novel multi-classifier distorted greedy model that is optimized to operate under various levels of automation and abstains from making uncertain predictions with theoretical guarantees. Our results show that DiGrI outperforms strong comparative models including large language models in detecting mental health issues on a publicly available Reddit dataset. We discuss the empirical and practical implications, including the ethical considerations of using DiGrI for online automatic suicide ideation detection involving humans, if it were to be translated for use in clinical and public health practice.

In the last 10 years, there has been a rise in the number of Arabic texts, which necessitates a more profound understanding of algorithms to efficiently understand and classify Arabic texts in many applications, like sentiment analysis. This paper presents a comprehensive review of recent developments in Arabic text classification (ATC) and Arabic text representation (ATR). We analyze the effectiveness of various models and techniques. Our review finds that while deep learning models, particularly transformer‐based architectures, are increasingly effective for ATC, challenges such as dialectal variations and insufficient labeled datasets remain key obstacles. However, developing suitable representation models and designing classification algorithms is still challenging for researchers, especially in Arabic. A basic introduction to ATC is provided in this survey, including preprocessing, representation, dimensionality reduction (DR), and classification with many evaluation metrics. In addition, the survey includes a qualitative and quantitative study of the ATC’s existing works. Finally, we conclude this work by exploring the limitations of the existing methods. We also mention the open challenges related to ATC, which help researchers identify new directions and challenges for ATC.

Federated learning facilitates collaborative training of machine learning models on data distributed across multiple locations, effectively addressing the privacy concerns by eliminating the need for data centralization—a critical consideration in medical image analysis. In healthcare applications like cardiovascular segmentation, datasets from individual sites often feature annotations for specific heart regions, leading to partial overlaps. To address this challenge, we present a two-step partial annotation framework for federated learning, featuring a hybrid 3D multi-encoding UNet enhanced with focal modulation layers in the second stage. This architecture enables specialized subnetworks to act as experts, extracting features tailored to specific regions of interest based on each client’s data. To further improve feature extraction and differentiation, we incorporate focal modulation blocks and apply regularization by introducing an auxiliary generic decoder during training. Comprehensive experiments on diverse cardiac MRI datasets demonstrate that our approach significantly outperforms centralized learning models.

Mental health concerns are a critical issue in today’s digital age, posing a threat to both individual and societal well-being and making the identification of at-risk individuals crucial. Analyzing an individual’s social media post history can offer insights into their mental health state and help identify the presence of suicidal ideation. However, the complexity of linguistic and temporal data, along with sparsity and time irregularities, poses a formidable challenge in machine learning. Previous methods in this domain either rely on Euclidean space for processing which does not adequately model the power-law properties of social media posts, or lose information due to the discretization of the time axis. To address these challenges, we propose a novel framework, THYMES, which leverages pre-trained encoders and a rich representation learning paradigm with hyperbolic learning to model power-law features for enhanced sequence modeling. We perform experiments on two datasets and demonstrate that THYMES outperforms previously proposed methods while maintaining classification fairness under heavy data imbalances. Additionally, we qualitatively analyze commonly misclassified samples to reveal the shortcomings of models in this domain.

The rising rate of mental health issues in the digital age underscores the critical need for proactive interventions to assess an individual’s well-being. This problem is further exacerbated by the social stigma surrounding the subject, which suppresses the willingness of victims to seek help. Social media can serve as an outlet for such individuals to express their negative emotions or thoughts of self-harm. The social media account of an individual can offer a plethora of valuable information that can be used to predict their mental health. By unifying principles of robust classifier training and selective classification, we propose a novel framework, SAFENet, to predict the suicide risk of users by using their historical social media posts. When the confidence of prediction is low or the individual is classified as a high-risk user, SAFENet delegates the analysis of the posts to a human evaluator for further intervention. Our experiments show that SAFENet outperforms existing state-of-the-art frameworks. We further qualitatively analyze predictions from SAFENet and demonstrate that it performs robustly on difficult samples that may cause contemporary methods to make errors. Our system addresses the urgent need for efficient and effective mental health intervention in the digital era.

Automatic Chinese patent approval prediction is an emerging and valuable task in patent analysis. However, it involves a rigorous and transparent decision-making process that includes patent comparison and examination to assess its innovation and correctness. This resultant necessity of decision evidentiality, coupled with intricate patent comprehension presents significant challenges and obstacles for the patent analysis community. Consequently, few existing studies are addressing this task. This paper presents the pioneering effort on this task using a retrieval-based classification approach. We propose a novel framework called DiSPat, which focuses on structural representation learning and disentanglement to predict the approval of Chinese patents and offer decision-making evidence. DiSPat comprises three main components: base reference retrieval to retrieve the Top-k most similar patents as a reference base; structural patent representation to exploit the inherent claim hierarchy in patents for learning a structural patent representation; disentangled representation learning to learn disentangled patent representations that enable the establishment of an evidential decision-making process. To ensure a thorough evaluation, we have meticulously constructed three datasets of Chinese patents. Extensive experiments on these datasets unequivocally demonstrate our DiSPat surpasses state-of-the-art baselines on patent approval prediction, while also exhibiting enhanced evidentiality.

Social media users use disease and symptoms words in different ways, including describing their personal health experiences figuratively or in other general discussions. The health mention classification (HMC) task aims to separate how people use terms, which is important in public health applications. Existing HMC studies address this problem using pretrained language models (PLMs). However, the remaining gaps in the area include the need for linguistic grounding, the requirement for large volumes of labelled data, and that solutions are often only tested on Twitter or Reddit, which provides limited evidence of the transportability of models. To address these gaps, we propose a novel method that uses a transformer-based PLM to obtain a contextual representation of target (disease or symptom) terms coupled with a contrastive loss to establish a larger gap between target terms' literal and figurative uses using linguistic theories. We introduce the use of a simple and effective approach for harvesting candidate instances from the broad corpus and generalising the proposed method using self-training to address the label scarcity challenge. Our experiments on publicly available health-mention datasets from Twitter (HMC2019) and Reddit (RHMD) demonstrate that our method outperforms the state-of-the-art HMC methods on both datasets for the HMC task. We further analyse the transferability and generalisability of our method and conclude with a discussion on the empirical and ethical considerations of our study.

Harmful memes detection is challenging due to the semantic gap between different modalities. Previous studies mainly focus on feature extraction and fusion to learn discriminative information from memes. However, they ignore the misalignment of the modalities caused by the modality gap and suffer from data scarcity, resulting in insufficient learning of fusion-based models. Recently, researchers transformed images into textual captions and used language models for predictions, resulting in non-informative image captions. To address these gaps, this paper proposes an instructions-based abstracting approach CapAlign, in zero-shot visual question-answering settings. Precisely, we prompt a large language model (LLM) to ask informative questions to a pre-trained vision-language model and use the dialogues to generate a high-quality image caption. Further, to align the generated caption with the textual content of a meme, we used an LLM with instructions to generate informative captions of the meme and then prepend it with the attributes of the visual content of a meme to a prompt-based LM for prediction. Experimental findings on two benchmark datasets show that our approach produces informative captions and outperforms state-of-the-art methods for detecting harmful memes.

The proliferation of social media platforms has fueled the rapid dissemination of fake news, posing threats to our real-life society. Existing methods use multimodal data or contextual information to enhance the detection of fake news by analyzing news content and/or its social context. However, these methods often overlook essential textual news content (articles) and heavily rely on sequential modeling and global attention to extract semantic information. These existing methods fail to handle the complex, subtle twists1 in news articles, such as syntax-semantics mismatches and prior biases, leading to lower performance and potential failure when modalities or social context are missing. To bridge these significant gaps, we propose a novel multi-hop syntax aware fake news detection (MSynFD) method, which incorporates complementary syntax information to deal with subtle twists in fake news. Specifically, we introduce a syntactical dependency graph and design a multi-hop subgraph aggregation mechanism to capture multi-hop syntax. It extends the effect of word perception, leading to effective noise filtering and adjacent relation enhancement. Subsequently, a sequential relative position-aware Transformer is designed to capture the sequential information, together with an elaborate keyword debiasing module to mitigate the prior bias. Extensive experimental results on two public benchmark datasets verify the effectiveness and superior performance of our proposed MSynFD over state-of-the-art detection models.

Topic modeling is a widely used technique for revealing underlying thematic structures within textual data. However, existing models have certain limitations, particularly when dealing with short text datasets that lack co-occurring words. Moreover, these models often neglect sentence-level semantics, focusing primarily on token-level semantics. In this paper, we propose PromptTopic, a novel topic modeling approach that harnesses the advanced language understanding of large language models (LLMs) to address these challenges. It involves extracting topics at the sentence level from individual documents, then aggregating and condensing these topics into a predefined quantity, ultimately providing coherent topics for texts of varying lengths. This approach eliminates the need for manual parameter tuning and improves the quality of extracted topics. We benchmark PromptTopic against the state-of-the-art baselines on three vastly diverse datasets, establishing its proficiency in discovering meaningful topics. Furthermore, qualitative analysis showcases PromptTopic’s ability to uncover relevant topics in multiple datasets.

Memes can be a useful way to spread information because they are funny, easy to share, and can spread quickly and reach further than other forms. With increased interest in COVID-19 vaccines, vaccination-related memes have grown in number and reach. Memes analysis can be difficult because they use sarcasm and often require contextual understanding. Previous research has shown promising results but could be improved by capturing global and local representations within memes to model contextual information. Further, the limited public availability of annotated vaccine critical memes datasets limit our ability to design computational methods to help design targeted interventions and boost vaccine uptake. To address these gaps, we present VaxMeme, which consists of 10,244 manually labelled memes. With VaxMeme, we propose a new multimodal framework designed to improve the memes' representation by learning the global and local representations of memes. The improved memes' representations are then fed to an attentive representation learning module to capture contextual information for classification using an optimised loss function. Experimental results show that our framework outperformed state-of-the-art methods with an F1-Score of 84.2%. We further analyse the transferability and generalisability of our framework and show that understanding both modalities is important to identify vaccine critical memes on Twitter. Finally, we discuss how understanding memes can be useful in designing shareable vaccination promotion, myth debunking memes and monitoring their uptake on social media platforms.

The creative nature of memes has made it possible for harmful content to spread quickly and widely on the internet. Harmful memes can range from spreading hate speech promoting violence, and causing emotional distress to individuals or communities. These memes are often designed to be misleading, manipulative, and controversial, making it challenging to detect and remove them from online platforms. Previous studies focused on how to fuse visual and language modalities to capture contextual information. However, meme analysis still severely suffers from data deficiency, resulting in insufficient learning of fusion modules. Further, using conventional pretrained encoders for text and images exhibits a greater semantic gap in feature spaces and leads to low performance. To address these gaps, this paper reformulates a harmful meme analysis as an auto-filling and presents a prompt-based approach to identify harmful memes. Specifically, we first transform multimodal data to a single (i.e., textual) modality by generating the captions and attributes of the visual data and then prepend the textual data in the prompt-based pre-trained language model. Experimental results on two benchmark harmful memes datasets demonstrate that our method outperformed state-of-the-art methods. We conclude with the transferability and robustness of our approach to identify creative harmful memes.

The creative web is all about combining different types of media to create a unique and engaging online experience. Multimodal data, such as text and images, is a key component in the creative web. Social media posts that incorporate both text descriptions and images offer a wealth of information and context. Text in social media posts typically relates to one topic, while images often convey information about multiple topics due to the richness of visual content. Despite this potential, many existing multimodal topic models do not take these criteria into account, resulting in poor quality topics being generated. Therefore, we proposed a Coherent Topic modeling for Multimodal Data (CTM-MM), which takes into account that text in social media posts typically relates to one topic, while images can contain information about multiple topics. Our experimental results show that CTM-MM outperforms traditional multimodal topic models in terms of classification and topic coherence.

Fashion recommendation (FR) has received increasing attention in the research of new types of recommender systems. Existing fashion recommender systems (FRSs) typically focus on clothing item suggestions for users in three scenarios: 1) how to best recommend fashion items preferred by users; 2) how to best compose a complete outfit, and 3) how to best complete a clothing ensemble. However, current FRSs often overlook an important aspect when making FR, that is, the compatibility of the clothing item or outfit recommendations is highly dependent on the scene context. To this end, we propose the scene-aware fashion recommender system (SAFRS), which uncovers a hitherto unexplored avenue where scene information is taken into account when constructing the FR model. More specifically, our SAFRS addresses this problem by encoding scene and outfit information in separation attention encoders and then fusing the resulting feature embeddings via a novel scene-aware compatibility score function. Extensive qualitative and quantitative experiments are conducted to show that our SAFRS model outperforms all baselines for every evaluated metric.

Matrix factorization (MF) is a popular method for collaborative filtering. Recently, more and more MF methods have been proposed to incorporate side information. However, most of them are vulnerable to changes in data or sub-models. Moreover, data often follows a Pareto distribution and such an imbalance of data leads to a biased global mean, affecting the prediction accuracy. To overcome these defects, we designed a Hierarchical Generalized Linear Model-based MF method (HGLMMF) which can leverage both the original and processed side information. More specifically, HGLMMF utilizes one portion of the side information to construct covariates for fixed effects and the other portion to model the cluster-specific effects to adjust the global-bias problem. In fact, a number of state-of-the-art MF models can be viewed as special cases of HGLMMF. The obtained prediction results from experiments prove that HGLMMF is highly competitive with state-of-the-art methods.

There are certain challenges in news recommender systems that arise due to changing users’ preferences over dynamically generated news articles. It is important to expose users to a variety of information. Diversity is required in a news recommender system not only so that users do not get bored of reading similar news but because so that they do not get trapped in information bubbles. We propose a deep neural network based on a two-tower architecture that learns news representation through a news item tower and users’ representations through a query tower. To learn diversity, we introduce a category loss function that aligns items’ representation of uneven news categories. Experimental results on two news datasets reveal that our proposed architecture is more effective compared to the state-of-the-art methods and achieves a balance between accuracy and diversity.

User-generated text on social media is a promising avenue for public health surveillance and has been actively explored for its feasibility in the early identification of depression. Existing methods in the identification of depression have shown promising results; however, these methods were all focused on treating the identification as a binary classification problem. To date, there has been little effort towards identifying users’ depression severity level and disregard the inherent ordinal nature across these fine-grain levels. This paper aims to make early identification of depression severity levels on social media data. To accomplish this, we built a new dataset based on the inherent ordinal nature over depression severity levels using clinical depression standards on Reddit posts. The posts were classified into 4 depression severity levels covering the clinical depression standards on social media. Accordingly, we reformulate the early identification of depression as an ordinal classification task over clinical depression standards such as Beck’s Depression Inventory and the Depressive Disorder Annotation scheme to identify depression severity levels. With these, we propose a hierarchical attention method optimized to factor in the increasing depression severity levels through a soft probability distribution. We experimented using two datasets (a public dataset having more than one post from each user and our built dataset with a single user post) using real-world Reddit posts that have been classified according to questionnaires built by clinical experts and demonstrated that our method outperforms state-of-the-art models. Finally, we conclude by analyzing the minimum number of posts required to identify depression severity level followed by a discussion of empirical and practical considerations of our study.

In a user-generated text such as on social media platforms and online forums, people often use disease or symptom terms in ways other than to describe their health. In data-driven public health surveillance, the health mention classification (HMC) task aims to identify posts where users are discussing health conditions rather than using disease and symptom terms for other reasons. Existing computational research typically only studies health mentions in Twitter, with limited coverage of disease or symptom terms, ignore user behavior information, and other ways people use disease or symptom terms. To advance the HMC research, we present a Reddit health mention dataset (RHMD), a new dataset of multi-domain Reddit data for the HMC. RHMD consists of 10,015 manually labeled Reddit posts that mention 15 common disease or symptom terms and are annotated with four labels: namely personal health mentions, non-personal health mentions, figurative health mentions, and hyperbolic health mentions. With RHMD, we propose HMCNET that combines a target keyword (disease or symptom term) identification and user behavior hierarchically to improve HMC. Experimental results demonstrate that the proposed approach outperforms state-of-the-art methods with an F1-Score of 0.75 (an increase of 11% over the state-of-the-art) and shows that our new dataset poses a strong challenge to the existing HMC methods.

The sentiment analysis of the social media-based short text (e.g., Twitter messages) is very valuable for many good reasons, explored increasingly in different communities such as text analysis, social media analysis, and recommendation. However, it is challenging as tweet-like social media text is often short, informal and noisy, and involves language ambiguity such as polysemy. The existing sentiment analysis approaches are mainly for document and clean textual data. Accordingly, we propose a Deep Intelligent Contextual Embedding (DICE), which enhances the tweet quality by handling noises within contexts, and then integrates four embeddings to involve polysemy in context, semantics, syntax, and sentiment knowledge of words in a tweet. DICE is then fed to a Bi-directional Long Short Term Memory (BiLSTM) network with attention to determine the sentiment of a tweet. The experimental results show that our model outperforms several baselines of both classic classifiers and combinations of various word embedding models in the sentiment analysis of airline-related tweets.