Multilingual sentiment analysis remains challenging due to limited labeled data and strong cross-lingual variation. More specifically, the same expression may signal praise in one language but criticism in another, making sentiment detection highly context-dependent. At the same time, large language models (LLMs) often struggle to maintain stable classifications in low-resource and zero-shot settings. Therefore, we introduce MultiSent-RAG, a training-free retrieval-augmented framework that integrates structured sentiment corpora with unstructured multilingual evidence, and MultiSent-RAG-Cache, a semantic memory module that reuses prior label inferences based on embedding similarity. Our study uses 80,000 labeled instances across 12 languages, including low-resource and zero-shot settings. Retrieval augmentation leads to substantial improvements, with F1 gains of up to +0.65 points, and relative gains exceeding 70%–110% over strong LLM baselines. The semantic cache further reveals when memory reinforces consistent classifications and when it risks propagating errors. Beyond improving accuracy, our analysis reveals how retrieval strengthens contextual grounding across languages, while the semantic cache exposes a trade-off between classification stability and error propagation. These findings provide practical insights for designing robust and interpretable multilingual information processing systems.

Cong Huy Nguyen, Son Dinh Nguyen, Guanlin Li, Tuan Dung Nguyen, Aditya Narayan Sankaran, Mai Huy Thong, Thanh Trung Nguyen, Mai Hong Son, Reza Farahbakhsh, Phi Le Nguyen, Noel Crespi. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

International audience

We present a novel class of jailbreak adversarial attacks on LLMs, termed Task-in-Prompt (TIP) attacks. Our approach embeds sequence-to-sequence tasks (e.g., cipher decoding, riddles, code execution) into the model’s prompt to indirectly generate prohibited inputs. To systematically assess the effectiveness of these attacks, we introduce the PHRYGE benchmark. We demonstrate that our techniques successfully circumvent safeguards in six state-of-the-art language models, including GPT-4o and LLaMA 3.2. Our findings highlight critical weaknesses in current LLM safety alignment and underscore the urgent need for more sophisticated defence strategies.

Online abusive content detection, particularly in low-resource settings and within the audio modality, remains underexplored. We investigate the potential of pre-trained audio representations for detecting abusive language in low-resource languages, in this case, in Indian languages using Few Shot Learning (FSL). Leveraging powerful representations from models such as Wav2Vec and Whisper, we explore cross-lingual abuse detection using the ADIMA dataset with FSL. Our approach integrates these representations within the Model-Agnostic Meta-Learning (MAML) framework to classify abusive language in 10 languages. We experiment with various shot sizes (50-200) evaluating the impact of limited data on performance. Additionally, a feature visualization study was conducted to better understand model behaviour. This study highlights the generalization ability of pre-trained models in low-resource scenarios and offers valuable insights into detecting abusive language in multilingual contexts.

Recent studies improve the cross-lingual transfer learning by better aligning the internal representations within the multilingual model or exploring the information of the target language using self-training. However, the alignment-based methods exhibit intrinsic limitations such as non-transferable linguistic elements, while most of the self-training based methods ignore the useful information hidden in the low-confidence samples. To address this issue, we propose CoNLST (Contrastive Negative Learning and Self-Training) to leverage the information of low-confidence samples. Specifically, we extend the negative learning to the metric space by selecting negative pairs based on the complementary labels and then employ self-training to iteratively train the model to converge on the obtained clean pseudo-labels. We evaluate our approach on the widely-adopted cross-lingual benchmark XNLI. The experiment results show that our method improves upon the baseline models and can serve as a beneficial complement to the alignment-based methods.

The rise of emergence of social media platforms has fundamentally altered how people communicate, and among the results of these developments is an increase in online use of abusive content. Therefore, automatically detecting this content is essential for banning inappropriate information, and reducing toxicity and violence on social media platforms. The existing works on hate speech and offensive language detection produce promising results based on pre-trained transformer models, however, they considered only the analysis of abusive content features generated through annotated datasets. This paper addresses a multi-task joint learning approach which combines external emotional features extracted from another corpora in dealing with the imbalanced and scarcity of labeled datasets. Our analysis are using two well-known Transformer-based models, BERT and mBERT, where the later is used to address abusive content detection in multi-lingual scenarios. Our model jointly learns abusive content detection with emotional features by sharing representations through transformers' shared encoder. This approach increases data efficiency, reduce overfitting via shared representations, and ensure fast learning by leveraging auxiliary information. Our findings demonstrate that emotional knowledge helps to more reliably identify hate speech and offensive language across datasets. Our hate speech detection Multi-task model exhibited 3% performance improvement over baseline models, but the performance of multi-task models were not significant for offensive language detection task. More interestingly, in both tasks, multi-task models exhibits less false positive errors compared to single task scenario.

With the freedom of communication provided in online social media, hate speech has increasingly generated. This leads to cyber conflicts affecting social life at the individual and national levels. As a result, hateful content classification is becoming increasingly demanded for filtering hate content before being sent to the social networks. This paper focuses on classifying hate speech in social media using multiple deep models that are implemented by integrating recent transformer-based language models such as BERT, and neural networks. To improve the classification performances, we evaluated with several ensemble techniques, including soft voting, maximum value, hard voting and stacking. We used three publicly available Twitter datasets (Davidson, HatEval2019, OLID) that are generated to identify offensive languages. We fused all these datasets to generate a single dataset (DHO dataset), which is more balanced across different labels, to perform multi-label classification. Our experiments have been held on Davidson dataset and the DHO corpora. The later gave the best overall results, especially F1 macro score, even it required more resources (time execution and memory). The experiments have shown good results especially the ensemble models, where stacking gave F1 score of 97% on Davidson dataset and aggregating ensembles 77% on the DHO dataset.

Impersonators are playing an important role in the production and propagation of the content on Online Social Networks, notably on Instagram. These entities are nefarious fake accounts that intend to disguise a legitimate account by making similar profiles and then striking social media by fake content, which makes it considerably harder to understand which posts are genuinely produced. In this study, we focus on three important communities with legitimate verified accounts. Among them, we identify a collection of 2.2K impersonator profiles with nearly 10k generated posts, 68K comments, and 90K likes. Then, based on profile characteristics and user behaviours, we cluster them into two collections of `bot' and `fan'. In order to separate the impersonator-generated post from genuine content, we propose a Deep Neural Network architecture that measures `profiles' and `posts' features to predict the content type: `bot-generated', `fan-generated', or `genuine' content. Our study shed light into this interesting phenomena and provides interesting observation on bot-generated content that can help us to understand the role of impersonators in the production of fake content on Instagram.

Recent years have seen a new form of advertisement campaigns emerge: those involving so-called social media influencers. These influencers accept money in return for promoting products via their social media feeds. We gather a large-scale Instagram dataset covering thousands of accounts advertising products, and create a categorisation based on the number of users they reach. We then provide a detailed analysis of the types of products being advertised by these accounts, their potential reach, and the engagement they receive from their followers. Based on our findings, we train machine learning models to distinguish sponsored content from non-sponsored, and identify cases where people are generating sponsored posts without labelling them.

Impersonators on Online Social Networks such as Instagram are playing an important role in the propagation of the content. These entities are the type of nefarious fake accounts that intend to disguise a legitimate account by making similar profiles. In addition to having impersonated profiles, we observed a considerable engagement from these entities to the published posts of verified accounts. Toward that end, we concentrate on the engagement of impersonators in terms of active and passive engagements which is studied in three major communities including “Politician”, “News agency and “Sports star” on Instagram. Inside each community, four verified accounts have been selected. Based on implemented approach in our previous studies [1], we have collected 4. 8K comments, and 2. 6K likes across 566 posts created from 3. 8K impersonators during 7 months. Our study shed light into this interesting phenomena and provides a surprising observation that can help us to understand better how impersonators engaging themselves inside Instagram in terms of writing Comments and leaving Likes.

Social networking sites (SNSs) facilitate the sharing of ideas and information through different types of feedback including publishing posts, leaving comments and other type of reactions. However, some comments or feedback on SNSs are inconsiderate and offensive, and sometimes this type of feedback has a very negative effect on a target user. The phenomenon known as flaming goes hand-in-hand with this type of posting that can trigger almost instantly on SNSs. Most popular users such as celebrities, politicians and news media are the major victims of the flaming behaviors and so detecting these types of events will be useful and appreciated. Flaming event can be monitored and identified by analyzing negative comments received on a post. Thus, our main objective of this study is to identify a way to detect flaming events in SNS using a sentiment prediction method. We use a deep Neural Network (NN) model that can identity sentiments of variable length sentences and classifies the sentiment of SNSs content (both comments and posts) to discover flaming events. Our deep NN model uses Word 2Vec and FastText word embedding methods as its training to explore which method is the most appropriate. The labeled dataset for training the deep NN is generated using an enhanced lexicon based approach. Our deep NN model classifies the sentiment of a sentence into five classes: Very Positive, Positive, Neutral, Negative and Very Negative. To detect flaming incidents, we focus only on the comments classified into the Negative and Very Negative classes. As a use-case, we try to explore the flaming phenomena in the news media domain and therefore we focused on news items posted by three popular news media on Facebook (BBCNews, CNN and FoxNews) to train and test the model. The experimental results show that flaming events can be detected with our proposed approach, and we explored main characteristics that trigger a flaming event and topics discussed in the flaming posts.

Fake accounts and Impersonators on Online Social Networks such as Instagram are turning difficulties for society. This has attended to an increasing interest in detecting fake profiles and investigating their behaviours. Questions like who are impersonators? what are their characteristics? and are they bots? will arise. To answer, we begin this research by collecting data from three important communities on Instagram including “Politician”, “News agency”, and “Sports star”. Inside each community, four verified top accounts are picked. Based on the users who reacted to their published posts, we detect 4K impersonators [1]. Then we employed well-known clustering methods to distribute impersonators into separated clusters to observe obscure behaviours and unusual profile characteristics. We also studied the cross-group analysis of clusters inside each community to explore engagements. Finally, we conclude the study by providing a complete investigation of the bot-like cluster.

There is an ever-growing number of users who duplicate the social media accounts of celebrities or generally impersonate their presence on online social media as well as Instagram. Of course, this has led to an increasing interest in detecting fake profiles and investigating their behaviour. We begin this research by targeting a few famous politicians, including Donald J. Trump, Barack Obama, and Emmanuel Macron and collecting their activity for the period of 3 months using a specifically-designed crawler across Instagram. We then experimented with several profile characteristics such as username, display name, biography, and profile picture to identify impersonator among 1,5M unique users. Using publicly crawled data, our model was able to distinguish crowds of impersonators and political bots. We continued by providing an analysis of the characteristics and behaviour of these impersonators. Finally, we conclude the analysis by classifying impersonators into four different categories.

Localization, as a crucial service for sensor networks, is an energy-demanding process for both indoor and outdoor scenarios. GPS-based localization schemes are infeasible in remote, indoor areas, and it is not a cost-effective solution for large-scale networks. Single mobile-beacon architecture is recently considered to localize sensor networks with the aim of removing numerous GPS-equipped nodes. The critical issue for the mobile beacon-assisted localization is to preserve the consumed power to increase the lifetime. This paper presents a novel power control scheme, namely “Z-power,” for mobile beacon traveling along a predefined path. The proposed scheme takes the advantage of deterministic path traveled by the single beacon to efficiently adjust the transmission power. Based on the extensive results, the proposed power control scheme could successfully improve the beacon and sensors energy consumption about 25.37% and 34.09%, respectively. A significant energy-accuracy tradeoff was achieved using Z-power, which could successfully keep the same level of accuracy while providing lower energy consumption. Another group of results collected when obstacle-handling algorithm was applied at the presence of obstacles. In this scenario, Z-power improves energy consumption and localization accuracy with the same level of success.

Recent research has shown a substantial active presence of bots in online social networks (OSNs). In this article, we perform a comparative analysis of the usage and impact of bots and humans on Twitter—one of the largest OSNs in the world. We collect a large-scale Twitter dataset and define various metrics based on tweet metadata. Using a human annotation task, we assign “bot” and “human” ground-truth labels to the dataset and compare the annotations against an online bot detection tool for evaluation. We then ask a series of questions to discern important behavioural characteristics of bots and humans using metrics within and among four popularity groups. From the comparative analysis, we draw clear differences and interesting similarities between the two entities.

The rising popularity of social media has radically changed the way news content is propagated, including interactive attempts with new dimensions. To date, traditional news media such as newspapers, television and radio have already adapted their activities to the online news media by utilizing social media, blogs, websites etc. This paper provides some insight into the social media presence of worldwide popular news media outlets. Despite the fact that these large news media propagate content via social media environments to a large extent and very little is known about the news item producers, providers and consumers in the news media community in social media. To better understand these interactions, this work aims to analyze news items in two large social media, Twitter and Facebook. Towards that end, we collected all published posts on Twitter and Facebook from 48 news media to perform descriptive and predictive analyses using the dataset of 152K tweets and 80K Facebook posts. We explored a set of news media that originate content by themselves in social media, those who distribute their news items to other news media and those who consume news content from other news media and/or share replicas. We propose a predictive model to increase news media popularity among readers based on the number of posts, number of followers and number of interactions performed within the news media community. The results manifested that, news media should disperse their own content and they should publish first in social media in order to become a popular news media and receive more attractions to their news items from news readers.

By the massive available people data in social media, many digital service providers exploit widely this information to improve their services by predicting future requirements of their customers. This prediction mainly needs to study users' previous behavior and interactions and identify their preferences to provide rigorous recommendations that fulfill their requirements more favorably. Meanwhile, experiments show the prediction methods which exploit representation learning instead of traditional hand-crafted features accomplish better results and more precise predictions. In this study, we take advantage of representation learning method to predict user's future interactions by extracting users embeddings from their reactions history and exploit them in predicting future reactions. In this approach, users embeddings are used in a neural network designed with one-hidden layer and a softmax function in the end layer in order to predict users reactions. The proposed method is evaluated when user embeddings come from two different sources; users reactions history and random walks on the user network. The performance of the method has been evaluated by using a large Flickr dataset including more than 2M users and 11M users reactions sequences. The results show outperforming of the prediction method when it uses the history of user reactions to derive user embeddings.

The current research in opinion propagation is largely based on content analysis of social interactions of users on a network. However, the social power of users in propagating the opinion is not considered. In this paper, we study the impact of influential users on propagating the opinion on singed networks and propose an opinion propagation model considering influential users (OPIU). In particular, we identified two major factors involved in a user opinion propagation: (i) the influential users' effect, induced by the presence of a highly confident individual in the network, and (ii) the neighbors' effect, caused by the presence of the users who have a link with the current user. To assess the performance of the proposed method, we applied it to two large datasets of Epinions signed and Etsy unsigned networks. The results are compared with similar opinion propagation algorithm which indicates the influential users have a significant impact on propagation and considering them can effectively improve the propagation extend.

Recent research has shown a substantial active presence of bots in online social networks (OSNs). In this paper we utilise our previous work (Stweeler) to comparatively analyse the usage and impact of bots and humans on Twitter, one of the largest OSNs in the world. We collect a large-scale Twitter dataset and define various metrics based on tweet metadata. Using a human annotation task we assign 'bot' and 'human' ground-truth labels to the dataset, and compare the annotations against an online bot detection tool for evaluation. We then ask a series of questions to discern important behavioural characteristics of bots and humans using metrics within and among four popularity groups. From the comparative analysis we draw differences and interesting similarities between the two entities, thus paving the way for reliable classification of bots, and studying automated political infiltration and advertisement campaigns.

Content originality detection is an interesting research topic in large-scale scenarios especially in social media where anyone has the ability to produce and disseminate content in different forms through their profiles and activities. What is missing in these communication sites is to be able to identify original content producers as some users spread information copied from other users without indicating its original producer, or where they found it. This paper provides a conceptualized approach for content originality detection and illustrates the efficiency of the model when applying it to a Twitter dataset. This approach amalgamates user's linguistic features and their online circadian behaviors to identify accurately the content originator for a given text. The proposed approach is evaluated using an F1-measure and the results indicate an accuracy of 95% or higher for all test scenarios. While achieving high accuracy in the test results, our approach, as a usecase, was applied in the context of news agencies popular worldwide to identify news producers and consumers by analyzing their Tweets. We investigated intra and inter news flows among several major news agencies considered in our dataset. Our results show that this proposed approach can distinguish News Story Tellers from News Propagators in the news agencies community as well as provide information that helps to understand the flow patterns between different news groups.

International audience

Nowadays, a big part of people rely on available content in social media in their decisions (e.g., reviews and feedback on a topic or product). The possibility that anybody can leave a review provides a golden opportunity for spammers to write spam reviews about products and services for different interests. Identifying these spammers and the spam content is a hot topic of research, and although a considerable number of studies have been done recently toward this end, but so far the methodologies put forth still barely detect spam reviews, and none of them show the importance of each extracted feature type. In this paper, we propose a novel framework, namedNetSpam, which utilizes spam features for modeling review data sets as heterogeneous information networks to map spam detection procedure into a classification problem in such networks. Using the importance of spam features helps us to obtain better results in terms of different metrics experimented on real-world review data sets from Yelp and Amazon Web sites. The results show thatNetSpamoutperforms the existing methods and among four categories of features, including review-behavioral, user-behavioral, review-linguistic, and user-linguistic, the first type of features performs better than the other categories.

The data generated by the devices and existing infrastructure in the Internet of Things (IoT) should be shared among applications. However, data sharing in the IoT can only reach its full potential when multiple participants contribute their data, for example when people are able to use their smartphone sensors for this purpose. We believe that each step, from sensing the data to the actionable knowledge, requires trust-enabled mechanisms to facilitate data exchange, such as data perception trust, trustworthy data mining, and reasoning with trust related policies. The absence of trust could affect the acceptance of sharing data in smart cities. In this study, we focus on data usage transparency and accountability and propose a trust model for data sharing in smart cities, including system architecture for trust-based data sharing, data semantic and abstraction models, and a mechanism to enhance transparency and accountability for data usage. We apply semantic technology and defeasible reasoning with trust data usage policies. We built a prototype based on an air pollution monitoring use case and utilized it to evaluate the performance of our solution.

Online Social Networks (OSNs) are being intensively used by professional users (e.g., companies, politician, athletes, celebrities, etc) in order to interact with a huge amount of regular OSN users with different purposes (marketing campaigns, customer feedback, public reputation, etc). Hence, due to the large catalog of existing OSNs, professional users usually count with OSN accounts in different systems. In this context an interesting question is whether professional users publish the same information across their OSN accounts, or actually they use different OSNs in a different manner. We define as cross-posting activity the action of publishing the same information in two or more OSNs. In this paper we aim at characterizing the cross-posting activity of professional OSN users across three major OSNs, Facebook, Twitter and Google+. To achieve this goal we perform a large-scale measurement-based analysis across more than 2M posts collected from 616 professional users with active accounts in the three referred OSNs.

Data in smart cities is commonly generated by a large variety of participants including institutional actors, equipment manufacturers, network operators, infrastructure providers, service providers, and end users. This data potentially undergoes several transformations such as aggregation and/or composition before finally being consumed. In this context of sharing data between diverse consumers, it is essential to provide the data producers the means by which they can exercise control over how and by whom the data is used. To date, usage control has received attention in the domains of the web and social networks, in terms of confidentiality, privacy and access control aspects. However, it has not yet been fully applied in a rigorous manner in the context of smart cites. In this paper we study usage control with the goal to address the problem of providing stakeholders more control over their data and enforcing accountable management of such data. We first propose a new data usage policy, called DUPO, which captures the diversity of obligations and constraints resulting from the usage control requirements for smart cities. Next, we apply a defeasible logic based approach on DUPO to formally define rule language, solve rule conflicts, and elaborate reasoning. We then introduce the data handling mechanism, which provides useful functionality to process consumer's request, ensuring the accountability of the policy enforcement, and traceability of the data usage. To this end we benefit from SPINdle reasoner to implement the proposed usage control module covered main functionalities of the mechanism.

Link prediction for new users who have not created any link is a fundamental problem in Online Social Networks (OSNs). It can be used to recommend friends for new users to start building their social networks. The existing studies use cross-platform approaches to predict a new user's links on a certain OSN by porting his existing links from other OSNs. However, it cannot work when OSNs are not willing to share their data or users do not want to connect different OSN accounts. In this paper, we use a single-platform approach to carry out the link prediction. We explore the users' profile attributes (e.g., workplace, high school and hometown) which can be easily obtained during the new users' sign up procedure. Based on the limited available information from the new user, along with the attributes and links from existing users, we extract three types of social features: basic feature, derived feature and latent relation feature. We propose a link prediction model using these social features based on Support Vector Machines. Eventually, we rely on a large Facebook data set consisting of 479,000 users to evaluate our proposed model. The result reveals that our model outperforms the baselines by achieving the AUC value of 0.83; it also demonstrates that each of the proposed social features contribute significantly to the prediction model.

Facebook, the most popular Online social network is a virtual environment where users share information and are in contact with friends. Apart from many useful aspects, there is a large amount of personal and sensitive information publicly available that is accessible to external entities/users. In this paper we study the public exposure of Facebook profile attributes to understand what type of attributes are considered more sensitive by Facebook users in terms of privacy, and thus are rarely disclosed, and which attributes are available in most Facebook profiles. Furthermore, we also analyze the public exposure of Facebook users by accounting the number of attributes that users make publicly available on average. To complete our analysis we have crawled the profile information of 479K randomly selected Facebook users. Finally, in order to demonstrate the utility of the publicly available information in Facebook profiles we show in this paper three case studies. The first one carries out a gender-based analysis to understand whether men or women share more or less information. The second case study depicts the age distribution of Facebook users. The last case study uses data inferred from Facebook profiles to map the distribution of worldwide population across cities according to its size.

Location-Based services with social networks improve users' experience and enrich people's social live. However, location information is often inadequate due to privacy and security concerns. We seek to infer users' ‘Current City’ on Facebook for coarse location based applications. We first extract users' multiple explicit and implicit location attributes, and analyze correlations of these attributes from two perspective: user-centric and user-friends. We observe that both user-centric and user-friends location attributes tightly correlate to a user's Current City (e.g., 60% of users stay in their hometown, 60% of users live in the same city as 50% of their friends). Based on extensive analysis and observations on location attributes correlations, we have constructed a Current City Prediction model (CCP) using artificial neural network (ANN) learning frameworks. The experimental results indicate that we achieve accuracy levels of 84% for city-level prediction and 98% for country-level which are increases of 9% and 18%, respectively than what is possible with Tweecalization.

During recent years, a few countries have put in place online antipiracy laws and there has been some major enforcement actions against violators. This raises the question that to what extent antipiracy actions have been effective in deterring online piracy? This is a challenging issue to explore because of the difficulty to capture user behavior, and to identify the subtle effect of various underlying (and potentially opposing) causes. In this paper, we tackle this question by examining the impact of two major antipiracy actions, the closure of Megaupload and the implementation of the French antipiracy law, on publishers in the largest BitTorrent portal who are major providers of copyrighted content online. We capture snapshots of BitTorrent publishers at proper times relative to the targeted antipiracy event and use the trends in the number and the level of activity of these publishers to assess their reaction to these events. Our investigation illustrates the importance of examining the impact of antipiracy events on different groups of publishers and provides valuable insights on the effect of selected major antipiracy actions on publishers' behavior.

Content management appears as an essential requirement in order to deploy enhanced P2P networks. In P2P networks, the content that is requested by the query node could be located at different locations/nodes; therefore an efficient search mechanism is required. The proposed search algorithm in this paper, called socP2P, relies on peers' social relationships (friendships, shared interests, shared background and experiences) to improve the content discovery compared to similar solutions. With socP2P nodes can improve searches by using knowledge gained by `overheard information' during their stay in the network. In addition, our algorithm exploits peers' common interests, friendships, and capability of memorizing experiences (received and routed queries) by them. Simulation results show that socP2P is able to achieve a high success rate, low delay and low overhead. We have verified that our algorithm is not only useful in finding popular contents in the network but also good enough to locate rare files. The obtained results reflect that exploiting social information in P2P networks leads to a more efficient content search mechanism.