Phishing scams on Ethereum have expanded with the surge of the platform, posing substantial challenges due to the sheer similarity in user behaviours and sparse temporal instances. Current methods often fail to tackle these concerns and overlook the temporal sequence of transactions, resulting in suboptimal performance. In this paper, we aim to address these gaps by focusing on the alignment of two aspects: (1) User-specific local temporal behavior, and (2) Divergences from global activity patterns of the network. Hence, we introduce CATALOG (CApturing joint TemporAl dependencies from LOcal and Global user behaviour), a novel representation learning model that jointly captures the local and global user behviours and their correlations by leveraging a dual cross-attention mechanism paired with a bi-directional Masked Language Modelling (MLM) transformer. Our proposed model simultaneously learns from local behavioral shifts, global market trends, and contextually enriched embeddings, effectively distinguishing phishing from non-phishing users while addressing existing research gaps. Extensive experiments on real-world Ethereum transaction data show that our framework improves phishing detection by 7-8% in the F1-Score along with demonstrating the generalization to Ethereum versions 1.0 and 2.0.

Zero-Shot Stance Detection (ZSSD) identifies an author’s stance towards unseen targets. Existing works have mainly focused on contrastive, meta, adversarial learning, or data augmentation but face issues like data scarcity, generalizability , and lack of coherence between text and targets. Moreover, stance detection must be interpretable to ensure transparency. Recent works with large language models (LLMs) aim to enhance unseen target knowledge or generate explanations but often rely excessively on explicit reasoning or provide coarse explanations, overlooking implicit cues and complicating interpretation. To address these challenges, we propose a novel interpretable multi-stage ZSSD framework. Stage 1 decodes explanations (rationales) justifying the stance while Stage 2 provides the final stance label, thus providing inherent interpretability in predicting stances. Extensive experiments prove that our approach outperforms other baselines with an average improvement in F1 scores of 27.99% with LLMs and 23.60% without LLMs for SemEval and 14.62% with LLMs and 25.24% without LLMs for VAST datasets for the ZSSD task, benefiting from the proposed pipeline architecture and interpretable design. Furthermore, to mitigate the harmful effects of offensive content and promote a more respectful online environment, we integrate an intervention module that leverages the contextual insights derived from our ZSSD framework with the ethics-based text generation power of LLMs to develop interventions. Automatic and human evaluation of LLM-generated interventions based on various proposed criteria provide insights into how LLMs perceive similar information from different perspectives, which can help foster morally sound and respectful online discourse.

In this work, we aim to understand the general public perception of societal issues related to the current climate crisis and the COVID-19 pandemic on Twitter (X). Social media discussions on such matters often lead to misleading information, resulting in delays in initiatives proposed by governments or policymakers. Hence, we focus on extracting relevant information from the conversations on climate change and COVID that could be useful for authorities to curb the spread of potentially biased information by proposing the classification tasks of relevance detection (RD) and information categorization (IC). We first curate the datasets for the RD and IC tasks for the climate domain and extend the COVID-19 benchmark attention-worthy Twitter dataset for the IC task through manual annotation. We initially conduct experiments with LLMs and observe that LLMs can extract the relevant information in zero and few-shot settings based on multi-perspective reasoning in the form of cognitive empathy and ethical standards, but still perform worse than fine-tuned small language models. Based on the initial findings, we conclude that LLMs may not be the best extractor of relevant information, but induce cognitive empathy and ethical reasonings that can intuitively guide supervised models. To achieve this idea, we develop a cognitive empathy and ethical reasoning-based multi-tasking pipelined network for RD and IC tasks. Our proposed approach provides valuable insights that could be useful in real-world scenarios for governments, policymakers, and other researchers to decode the overall public outlook on societal issues.

Climate change has become one of the biggest challenges of our time. Social media platforms such as Twitter play an important role in raising public awareness and spreading knowledge about the dangers of the current climate crisis. With the increasing number of campaigns and communication about climate change through social media, the information could create more awareness and reach the general public and policy makers. However, these Twitter communications lead to polarization of beliefs, opinion-dominated ideologies, and often a split into two communities of climate change deniers and believers. In this paper, we propose a framework that helps identify denier statements on Twitter and thus classifies the stance of the tweet into one of the two attitudes towards climate change (denier/believer). The sentimental aspects of Twitter data on climate change are deeply rooted in general public attitudes toward climate change. Therefore, our work focuses on learning two closely related tasks: Stance Detection and Sentiment Analysis of climate change tweets. We propose a multi-task framework that performs stance detection (primary task) and sentiment analysis (auxiliary task) simultaneously. The proposed model incorporates the feature-specific and shared-specific attention frameworks to fuse multiple features and learn the generalized features for both tasks. The experimental results show that the proposed framework increases the performance of the primary task, i.e., stance detection by benefiting from the auxiliary task, i.e., sentiment analysis compared to its uni-modal and single-task variants.

In this work, we address the United Nations Sustainable Development Goal 13: Climate Action by focusing on identifying public attitudes toward climate change on social media platforms such as Twitter. Climate change is threatening the health of the planet and humanity. Public engagement is critical to address climate change. However, climate change conversations on Twitter tend to polarize beliefs, leading to misinformation and fake news that influence public attitudes, often dividing them into climate change believers and deniers. Our paper proposes an approach to classify the attitude of climate change tweets (believe/deny/ambiguous) to identify denier statements on Twitter. Most existing approaches for detecting stances and classifying climate change tweets either overlook deniers’ tweets or do not have a suitable architecture. The relevant literature suggests that emotions and higher levels of toxicity are prevalent in climate change Twitter conversations, leading to a delay in appropriate climate action. Therefore, our work focuses on learning stance detection (main task) while exploiting the auxiliary tasks of recognizing emotions and offensive utterances. We propose a multimodal multitasking framework MEMOCLiC that captures the input data using different embedding techniques and attention frameworks, and then incorporates the learned emotional and offensive expressions to obtain an overall representation of the features relevant to the stance of the input tweet. Extensive experiments conducted on a novel curated climate change dataset and two benchmark stance detection datasets (SemEval-2016 and ClimateStance-2022) demonstrate the effectiveness of our approach.

With the increasing use of Twitter for encouraging users to instigate violent behavior with hate and racial content, it becomes necessary to investigate the uniqueness in the dynamics of the spread of tweets made during violent communal incidents and the challenges they pose in early identification of potential viral content. In this article, we study the spread of the tweets made during several violent communal incidents along four major dimensions — the underlying follower network of the users, their structural and engagement characteristics, the cascades, and the cognitive aspects of the content, each of which plays a vital role in the spread of content. Using large public and collected data, we compare these features with tweets related to other subjects from several major domains, such as non-violent political events, celebrities, and technology, that contribute to a large fraction of the viral content over Twitter. We discover that while the spread of cascades and the users involved may provide strong early evidence of the viral content for several domains, the early phases of the spread of viral tweets related to violent communal incidents are characterized by cascades with protracted growth involving fringe or low-importance users, which would possibly make early prediction difficult. Our findings indicate that an interplay of certain network and cascade properties, together with the cognitive characteristics of tweets and the behavioral patterns of the engaging users, may provide stronger early indicators of the virality of this content.

Identifying tweets related to infrastructure damage during a crisis event is an important problem. However, the unavailability of labeled data during the early stages of a crisis event poses major challenge in training suitable models. Several domain adaptation strategies have been proposed for text classification that can be used to train models using available source data of previous crisis events and apply on a target data related to a current event. However, these approaches are insufficient to handle the distribution drift in the source and target data along with the class imbalance in the target data. In this paper we introduce an Ensemble learning approach with a Decoupled Adversarial (EnDeA) model to classify infrastructure damage tweets in a target tweet dataset. EnDeA is an ensemble of three different models two of which separately learn the event invariant and specific features of a target data from a set of source and target data. The third model which is an adversarial model helps to improve the prediction accuracy of both models. Unlike the existing approaches that also identify the domain invariant and specific properties of target data for sentiment classification, our method works for short texts and can better handle the distribution drift and class imbalance problem. We rigorously investigate the performance of the proposed approach using multiple public datasets and compare it with several state-of-the-art baselines. We discover that EnDeA outperforms these baselines with around 20% improvement in the 1 scores.