Online debates provide critical insights into public opinion and societal trends, yet the unstructured nature of these discussions presents significant challenges for analysis. In this paper, we present SAGESSE, a novel argumentation parsing pipeline tailored to Reddit debates, leveraging the capabilities of large language models to structure and interpret complex online arguments. SAGESSE generates detailed argument maps that organize debates systematically, offering a clearer understanding of discourse dynamics. We have developed a web application where users can select controversial Reddit topics and visualize the corresponding argument maps generated from user comments. This tool has the potential to aid analysts, policymakers, and researchers in tracking debate progress, gauging public sentiment, and identifying influential arguments. The web application is available at https://modemos.epfl.ch/sagesse/ .

The growth of large language models underscores the need for parameter-efficient fine-tuning. Despite its popularity, LoRA encounters storage and computational challenges when deploying multiple task- or user-specific modules. To address this, we introduce LoRA-XS, a novel fine-tuning method backed by a theoretical derivation. LoRA-XS drastically reduces trainable parameters by incorporating a small, trainable weight matrix between frozen low-rank matrices derived from the Singular Value Decomposition of pre-trained weights. This design enables LoRA-XS to reduce storage requirements by over 100x in 7B models compared to LoRA. Additionally, unlike other methods, LoRA-XS imposes no lower bound on trainable parameters - it can scale from a single parameter per module to arbitrarily large values, adapting to any storage or computational constraint. Evaluations on GLUE, GSM8K, MATH, and commonsense reasoning benchmarks across different model scales reveal that LoRA-XS consistently outperforms or matches LoRA and VeRA in accuracy, offering unmatched parameter efficiency. Our ablation studies highlight the significance of singular vectors in transformer weights, establishing LoRA-XS as a powerful, storage-efficient solution for scaling and personalizing large language models.

Documents are fundamental to preserving and disseminating information, often incorporating complex layouts, tables, and charts that pose significant challenges for automatic document understanding (DU). While vision-language large models (VLLMs) have demonstrated improvements across various tasks, their effectiveness in processing long-context vision inputs remains unclear. This paper introduces WikiMixQA, a benchmark comprising 1,000 multiple-choice questions (MCQs) designed to evaluate cross-modal reasoning over tables and charts extracted from 4,000 Wikipedia pages spanning seven distinct topics. Unlike existing benchmarks, WikiMixQA emphasizes complex reasoning by requiring models to synthesize information from multiple modalities. We evaluate 12 state-of-the-art vision-language models, revealing that while proprietary models achieve 70

Stance detection, a key task in natural language processing, determines an author's viewpoint based on textual analysis. This study evaluates the evolution of stance detection methods, transitioning from early machine learning approaches to the groundbreaking BERT model, and eventually to modern Large Language Models (LLMs) such as ChatGPT, LLaMa-2, and Mistral-7B. While ChatGPT's closed-source nature and associated costs present challenges, the open-source models like LLaMa-2 and Mistral-7B offers an encouraging alternative. Initially, our research focused on fine-tuning ChatGPT, LLaMa-2, and Mistral-7B using several publicly available datasets. Subsequently, to provide a comprehensive comparison, we assess the performance of these models in zero-shot and few-shot learning scenarios. The results underscore the exceptional ability of LLMs in accurately detecting stance, with all tested models surpassing existing benchmarks. Notably, LLaMa-2 and Mistral-7B demonstrate remarkable efficiency and potential for stance detection, despite their smaller sizes compared to ChatGPT. This study emphasizes the potential of LLMs in stance detection and calls for more extensive research in this field.

Federated learning (FL) has recently emerged as a compelling machine learning paradigm, prioritizing the protection of privacy for training data. The increasing demand to address issues such as “the right to be forgotten” and combat data poisoning attacks highlights the importance of techniques, known as unlearning, which facilitate the removal of specific training data from trained FL models. Despite numerous unlearning methods proposed for centralized learning, they often prove inapplicable to FL due to fundamental differences in the operation of the two learning paradigms. Consequently, unlearning in FL remains in its early stages, presenting several challenges. Many existing unlearning solutions in FL require a costly retraining process, which can be burdensome for clients. Moreover, these methods are primarily validated through experiments, lacking theoretical assurances. In this study, we introduce Fast-FedUL, a tailored unlearning method for FL, which eliminates the need for retraining entirely. Through meticulous analysis of the target client’s influence on the global model in each round, we develop an algorithm to systematically remove the impact of the target client from the trained model. In addition to presenting empirical findings, we offer a theoretical analysis delineating the upper bound of our unlearned model and the exact retrained model (the one obtained through retraining using untargeted clients). Experimental results with backdoor attack scenarios indicate that Fast-FedUL effectively removes almost all traces of the target client (achieving a mere 0.01

Maximal subgraph mining is increasingly important in various domains, including bioinformatics, genomics, and chemistry, as it helps identify common characteristics among a set of graphs and enables their classification into different categories. Existing approaches for identifying maximal subgraphs typically rely on traversing a graph lattice. However, in practice, these approaches are limited to relatively small subgraphs due to the exponential growth of the search space and the NP-completeness of the underlying subgraph isomorphism test. In this work, we propose SCAMA , an approach that addresses these limitations by adopting a divide-and-conquer strategy for efficient mining of maximal subgraphs. Our approach involves initially partitioning a graph database into equivalence classes using bootstrapped backbones, which are tree-shaped frequent subgraphs. We then introduce a learning process based on a novel graph convolutional network (GCN) to extract maximal backbones for each equivalence class. A critical insight of our approach is that by estimating each maximal backbone directly in the embedding space, we can avoid the exponential traversal of the graph lattice. From the extracted maximal backbones, we construct the maximal frequent subgraphs. Furthermore, we outline how SCAMA can be extended to perform top- k largest frequent subgraph mining and how the discovered patterns facilitate graph classification. Our experimental results demonstrate the effectiveness of SCAMA in identifying almost perfectly maximal frequent subgraphs, while exhibiting approximately 10 times faster performance compared to the best baseline technique.

Mapping the technology landscape is crucial for market actors to take informed investment decisions. However, given the large amount of data on the Web and its subsequent information overload, manually retrieving information is a seemingly ineffective and incomplete approach. In this work, we propose an end-to-end recommendation based retrieval approach to support automatic retrieval of technologies and their associated companies from raw Web data. This is a two-task setup involving (i) technology classification of entities extracted from company corpus, and (ii) technology and company retrieval based on classified technologies. Our proposed framework approaches the first task by leveraging DistilBERT which is a state-of-the-art language model. For the retrieval task, we introduce a recommendation-based retrieval technique to simultaneously support retrieving related companies, technologies related to a specific company and companies relevant to a technology. To evaluate these tasks, we also construct a data set that includes company documents and entities extracted from these documents together with company categories and technology labels. Experiments show that our approach is able to return 4 times more relevant companies while outperforming traditional retrieval baseline in retrieving technologies.

Vision-Language Pre-training (VLP) has advanced the performance of many vision-language tasks, such as image-text retrieval, visual entailment, and visual reasoning. The pre-training mostly utilizes lexical databases and image queries in English. Previous work has demonstrated that the pre-training in English does not transfer well to other languages in a zero-shot setting. However, multilingual pre-trained language models (MPLM) have excelled at a variety of single-modal language tasks. In this paper, we propose a simple yet efficient approach to adapt VLP to unseen languages using MPLM. We utilize a cross-lingual contextualized token embeddings alignment approach to train text encoders for non-English languages. Our approach does not require image input and primarily uses machine translation, eliminating the need for target language data. Our evaluation across three distinct tasks (image-text retrieval, visual entailment, and natural language visual reasoning) demonstrates that this approach outperforms the state-of-the-art multilingual vision-language models without requiring large parallel corpora. Our code is available at https://github.com/Yasminekaroui/CliCoTea.

Social media is an important source of real-time imagery concerning world events. One subset of social media posts which may be of particular interest are those featuring firearms. These posts can give insight into weapon movements, troop activity and civilian safety. Object detection tools offer important opportunities for insight into these images. Unfortunately, these images can be visually complex, poorly lit and generally challenging for object detection models. We present an analysis of existing gun detection datasets, and find that these datasets to not effectively address the challenge of gun detection on real-life images. Following this, we present a novel object detection pipeline. We train our pipeline on a number of datasets including one created for this investigation made up of Twitter images of the Russo-Ukrainian War. We compare the performance of our model as trained on the different datasets to baseline numbers provided by original authors as well as a YOLO v5 benchmark. We find that our model outperforms the state-of-the-art benchmarks on contextually rich, real-life-derived imagery of firearms.

We present the first in-depth and large-scale study of misleading repurposing, in which a malicious user changes the identity of their social media account via, among other things, changes to the profile attributes in order to use the account for a new purpose while retaining their followers. We propose a definition for the behavior and a methodology that uses supervised learning on data mined from the Internet Archive's Twitter Stream Grab to flag repurposed accounts. We found over 100,000 accounts that may have been repurposed. We also characterize repurposed accounts and found that they are more likely to be repurposed after a period of inactivity and deleting old tweets. We also provide evidence that adversaries target accounts with high follower counts to repurpose, and some make them have high follower counts by participating in follow-back schemes. The results we present have implications for the security and integrity of social media platforms, for data science studies in how historical data is considered, and for society at large in how users can be deceived about the popularity of an opinion.

In this work, we study whether multilingual language models (MultiLMs) can transfer logical reasoning abilities to other languages when they are fine-tuned for reasoning in a different language. We evaluate the cross-lingual reasoning abilities of MultiLMs in two schemes: (1) where the language of the context and the question remain the same in the new languages that are tested (i.e., the reasoning is still monolingual, but the model must transfer the learned reasoning ability across languages), and (2) where the language of the context and the question is different (which we term code-switched reasoning). On two logical reasoning datasets, RuleTaker and LeapOfThought, we demonstrate that although MultiLMs can transfer reasoning ability across languages in a monolingual setting, they struggle to transfer reasoning abilities in a code-switched setting. Following this observation, we propose a novel attention mechanism that uses a dedicated set of parameters to encourage cross-lingual attention in code-switched sequences, which improves the reasoning performance by up to 14% and 4% on the RuleTaker and LeapOfThought datasets, respectively.

Reviewing papers for conferences is an important and hard task that brings several challenges. The Data Management community has been increasingly struggling with high reviewer load, low-quality reviews and low reviewer engagement, unethical reviewing practices as well as undeclared and under-declared conflicts of interest. In this article, we report the results of a survey we conducted to gather the opinion of the Data Management community regarding what could be done to address these challenges. We reached out to about 1,200 members of the data management community with relevant reviewing experience and collected 345 anonymous responses. We plan to follow up with a subsequent report, discussing in more depth particular proposals, inspired by the collective feedback from the community.

Recent transformer language models achieve outstanding results in many natural language processing (NLP) tasks. However, their enormous size often makes them impractical on memory-constrained devices, requiring practitioners to compress them to smaller networks. In this paper, we explore offline compression methods, meaning computationally-cheap approaches that do not require further fine-tuning of the compressed model. We challenge the classical matrix factorization methods by proposing a novel, better-performing autoencoder-based framework. We perform a comprehensive ablation study of our approach, examining its different aspects over a diverse set of evaluation settings. Moreover, we show that enabling collaboration between modules across layers by compressing certain modules together positively impacts the final model performance. Experiments on various NLP tasks demonstrate that our approach significantly outperforms commonly used factorization-based offline compression methods.

The COVID-19 pandemic has fueled the spread of misinformation on social media and the Web as a whole. The phenomenon dubbed `infodemic' has taken the challenges of information veracity and trust to new heights by massively introducing seemingly scientific and technical elements into misleading content. Despite the existing body of work on modeling and predicting misinformation, the coverage of very complex scientific topics with inherent uncertainty and an evolving set of findings, such as COVID-19, provides many new challenges that are not easily solved by existing tools. To address these issues, we introduce SciLander, a method for learning representations of news sources reporting on science-based topics. We extract four heterogeneous indicators for the sources; two generic indicators that capture (1) the copying of news stories between sources, and (2) the use of the same terms to mean different things (semantic shift), and two scientific indicators that capture (1) the usage of jargon and (2) the stance towards specific citations. We use these indicators as signals of source agreement, sampling pairs of positive (similar) and negative (dissimilar) samples, and combine them in a unified framework to train unsupervised news source embeddings with a triplet margin loss objective. We evaluate our method on a novel COVID-19 dataset containing nearly 1M news articles from 500 sources spanning a period of 18 months since the beginning of the pandemic in 2020. Our results show that the features learned by our model outperform state-of-the-art baseline methods on the task of news veracity classification. Furthermore, a clustering analysis suggests that the learned representations encode information about the reliability, political leaning, and partisanship bias of these sources.

Understanding narratives requires reasoning about the cause-and-effect relationships between events mentioned in the text. While existing foundation models yield impressive results in many NLP tasks requiring reasoning, it is unclear whether they understand the complexity of the underlying network of causal relationships of events in narratives. In this work, we present CRAB, a new Causal Reasoning Assessment Benchmark designed to evaluate causal understanding of events in real-world narratives. CRAB contains fine-grained, contextual causality annotations for ~2.7K pairs of real-world events that describe various newsworthy event timelines (e.g., the acquisition of Twitter by Elon Musk). Using CRAB, we measure the performance of several large language models, demonstrating that most systems achieve poor performance on the task. Motivated by classical causal principles, we also analyze the causal structures of groups of events in CRAB, and find that models perform worse on causal reasoning when events are derived from complex causal structures compared to simple linear causal chains. We make our dataset and code available to the research community.

Graph neural networks take node features and graph structure as input to build representations for nodes and graphs. While there are a lot of focus on GNN models, understanding the impact of node features and graph structure to GNN performance has received less attention. In this paper, we propose an explanation for the connection between features and structure: graphs can be constructed by connecting node features according to a latent function. While this hypothesis seems trivial, it has several important implications. First, it allows us to define graph families which we use to explain the transferability of GNN models. Second, it enables application of GNNs for featureless graphs by reconstructing node features from graph structure. Third, it predicts the existence of a latent function which can create graphs that when used with original features in a GNN outperform original graphs for a specific task. We propose a graph generative model to learn such function. Finally, our experiments confirm the hypothesis and these implications. (C) 2022 Elsevier B.V. All rights reserved.

Lots of machine learning models with applications in Multimedia Search are released as Open Source Software. However, integrating these models into an application is not always an easy task due to the lack of a consistent interface to run, train or distribute models. With MoZuMa, we aim at reducing this effort by providing a model zoo for image similarity, text-to-image retrieval, face recognition, object similarity search, video key-frames detection and multilingual text search implemented in a generic interface with a modular architecture. The code is released as Open Source Software at https://github.com/mozuma/mozuma.

Advances in deep neural network (DNN) architectures have enabled new prediction techniques for stock market data. Unlike other multivariate time-series data, stock markets show two unique characteristics: (i) multi-order dynamics, as stock prices are affected by strong non-pairwise correlations (e.g., within the same industry); and (ii) internal dynamics, as each individual stock shows some particular behaviour. Recent DNN-based methods capture multi-order dynamics using hypergraphs, but rely on the Fourier basis in the convolution, which is both inefficient and ineffective. In addition, they largely ignore internal dynamics by adopting the same model for each stock, which implies a severe information loss. In this paper, we propose a framework for stock movement prediction to overcome the above issues. Specifically, the framework includes temporal generative filters that implement a memory-based mechanism onto an LSTM network in an attempt to learn individual patterns per stock. Moreover, we employ hypergraph attentions to capture the non-pairwise correlations. Here, using the wavelet basis instead of the Fourier basis, enables us to simplify the message passing and focus on the localized convolution. Experiments with US market data over six years show that our framework outperforms state-of-the-art methods in terms of profit and stability. Our source code and data are available at https://github.com/thanhtrunghuynh93/estimate.

Active Learning (AL) is a powerful tool for learning with less labeled data, in particular, for specialized domains, like legal documents, where unlabeled data is abundant, but the annotation requires domain expertise and is thus expensive.Recent works have shown the effectiveness of AL strategies for pre-trained language models.However, most AL strategies require a set of labeled samples to start with, which is expensive to acquire.In addition, pre-trained language models have been shown unstable during fine-tuning with small datasets, and their embeddings are not semantically meaningful.In this work, we propose a pipeline for effectively using active learning with pretrained language models in the legal domain.To this end, we leverage the available unlabeled data in three phases.First, we continue pretraining the model to adapt it to the downstream task.Second, we use knowledge distillation to guide the model's embeddings to a semantically meaningful space.Finally, we propose a simple, yet effective, strategy to find the initial set of labeled samples with fewer actions compared to existing methods.Our experiments on Contract-NLI, adapted to the classification task, and LEDGAR benchmarks show that our approach outperforms standard AL strategies, and is more efficient.Furthermore, our pipeline reaches comparable results to the fully-supervised approach with a small performance gap, and dramatically reduced annotation cost.Code and the adapted data will be made available.

Contrary to celebrities who owe their popularity online to their activity offline, malicious users such as trolls have to gain fame on social media through the social media itself. The exact reasons that a certain user has become popular are often obscure especially when the popularity was gained illicitly through means such as fake amplification of content. In this paper, we develop a methodology for uncovering why an account has become popular and present an open source tool that encapsulates this methodology. This tool aims to aid others in uncovering malicious accounts which have artificially gained many followers and to distinguish such accounts from those which gained followers and popularity honestly.