Misinformation detection is a critical task that can benefit significantly from the integration of external knowledge, much like manual fact-checking. In this work, we propose a novel method for representing textual documents that facilitates the incorporation of information from a knowledge base. Our approach, Text Encoding with Graph (TEG), processes documents by extracting structured information in the form of a graph and encoding both the text and the graph for classification purposes. Through extensive experiments, we demonstrate that this hybrid representation enhances misinformation detection performance compared to using language models alone. Furthermore, we introduce TEGRA, an extension of our framework that integrates domain-specific knowledge, further enhancing classification accuracy in most cases.

Among news disorders, propagandist news are particularly insidious, because they tend to mix oriented messages with factual reports intended to look like reliable news. To detect propaganda, extant approaches based on Language Models such as BERT are promising but often overfit their training datasets, due to biases in data collection. To enhance classification robustness and improve generalization to new sources, we propose a neurosymbolic approach combining non-contextual text embeddings (fastText) with symbolic conceptual features such as genre, topic, and persuasion techniques. Results show improvements over equivalent text-only methods, and ablation studies as well as explainability analyses confirm the benefits of the added features. Keywords: Information disorder, Fake news, Propaganda, Classification, Topic modeling, Hybrid method, Neurosymbolic model, Ablation, Robustness

This paper presents FactNET, a framework for human-centered granular fact-checking. Each claim to check is decomposed into sub claims that are fact-checked independently by using an LLM’s internal knowledge. The user interacts with the framework through an innovative interface, displaying the complex claim in the form of a graph, allowing the input of human knowledge on the topic, and evaluating the trust given in the generated evidence. A video demonstrating the system is available at TO_BE_PUBLISHED (in submission documents for review phase).

Résumé L’encodage de texte est aujourd’hui basé sur de larges modèles de langue entièrement neuronaux, utilisés comme des boîtes noires . Afin de répondre au besoin d’explicabilité, nous proposons CATS 1 (Cognitive Attention To Syntax), unecouchepouvantêtreutiliséedanslesréseauxdeneuronesquipermetd’introduireduraisonnementsyntaxiquepourl’encodageetlaclassificationdetextes . Abstract Today, text encoding relies mostly on foundation models, used as black boxes . To bring more transparency, we propose CATS (Cognitive Attention To Syntax), a layer that incorporates syntactic reasoning in neural networks for text classification.

Most approaches for text encoding rely on the attention mechanism, at the core of the transformers architecture and large language models. The understanding of this mechanism is still limited and present inconvenients such as lack of interpretability, large requirements of data and low generalization. Based on current understanding of the attention mechanism, we propose CATS (Cognitive Attention To Syntax), a neurosymbolic attention encoding approach based on the syntactic understanding of texts. This approach has on-par to better performance compared to classical attention and displays expected advantages of neurosymbolic AI such as better functioning with little data and better explainability. This layer has been tested on the task of misinformation detection but is general and could be used in any task involving natural language processing.

Maxime Prieur, Souhir Gahbiche, Guillaume Gadek, Sylvain Gatepaille, Kilian Vasnier, Valerian Justine. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track). 2023.

In this paper, we combine two independent detection methods for identifying fake news: the algorithm VAGO uses semantic rules combined with NLP techniques to measure vagueness and subjectivity in texts, while the classifier FAKE-CLF relies on Convolutional Neural Network classification and supervised deep learning to classify texts as biased or legitimate. We compare the results of the two methods on four corpora. We find a positive correlation between the vagueness and subjectivity measures obtained by VAGO, and the classification o f t ext a s b iased by FAKE-CLF. The comparison yields mutual benefits: VAGO helps explain the results of FAKE-CLF. Conversely FAKE-CLF helps us corroborate and expand VAGO’s database. The use of two complementary techniques (rule-based vs data-driven) proves a fruitful approach for the challenging problem of identifying fake news.

Situation Awareness area deals with an ever-growing amount of data to be processed. Decision makers need new tools to swiftly assess a situation, in spite of the huge amount of information to interpret. This reality is even truer in Crisis Analysis as military and rescue domains. Automated systems are absolutely necessary for decision makers as they enable them to save a valuable time while making decisions quickly in order to solve problems. In this paper, we will first focus on a model of a dynamic environment designed for a military use case based on the NATO military doctrine. The scenario we have chosen is that of an army attacking the borders of a country. Intelligence services identify specific zones to observe and understand the enemy’s intentions. The model provides a topographical structure representing the different zones and the potential targets. The second part of the following work is an automated threat propagation model which provides a heat map of the threat in case of an enemy attack. Thanks to the threat spreading representation, the system is able to propose the k-best strategies of the assailant to identify which target the enemy is likely to attack to set up an efficient and fast counter-measure.

Situation Awareness area deals with an ever-growing amount of data to be processed. Decision makers need new tools to swiftly assess a situation, in spite of the huge amount of information to interpret. This reality is even truer in Crisis Analysis such as military and rescue domains. Considering the speed with which information is acquired, it is crucial to propose an efficient decision process. Automated systems are absolutely necessary for decision makers as they enable them to save a valuable time while making decisions quickly in order to solve problems. In this paper, we present an improvement of an automated threat propagation model through a dynamic environment. This model proposes a heat map of the potential threat of an enemy attacking different points. For a human operator, knowing the path and the threat in the upcoming moments rather than the final objective of the enemy is crucial to deploy his sensors or even establish early counter-measures. This work takes place in a military use case based on the NATO military doctrine. The proposed scenario describes an army attacking the borders of a country. This enemy tries to make the most of its troops while they are breaking through the defensive lines. Intelligence services identify specific zones to observe and understand the enemy’s intentions. The model provides a topographical structure that can be read like a graph structure representing the different observable zones that can be observed along with their potential targets.

Currently, maritime surveillance operators have to monitor by hand the massive amount of data at their disposal to spot the events of interest, thus limiting their capabilities. Maritime data comes from various and heterogeneous sources that can be merged into a dynamic attributed knowledge graph which represents an evolving maritime situation. Using this graph, the automation of alert rising comes through a link prediction task: given some labels from expert knowledge, are there similar situations of interest elsewhere in the graph? In this article, we review link prediction techniques for situation awareness in a maritime context, and draw conclusions on how the addition of attributes in a dynamic graph model could improve results on this task.

Currently, maritime surveillance operators have to monitor by hand the massive amount of data at their disposal to spot the events of interest, thus limiting their capabilities. Maritime data comes from various and heterogeneous sources, that can be merged into a dynamic attributed knowledge graph which represents an evolving maritime situation. Using this graph, the automation of alert rising comes through a link prediction task: given some labels from expert knowledge, are there similar situations of interest elsewhere in the graph? In this article, we review link prediction techniques for situation awareness in a maritime context, and draw conclusions on how the addition of attributes in a dynamic graph model could improve results on this task.

Most Situation Awareness applications in Information Fusion try to evaluate a dynamic environment by a passive approach combining heterogeneous information. However in a crisis situation, the decision has to be made efficiently as the world quickly evolves. A consequence is the difficulty to get the information in a fast and efficient way with an acceptable confidence. Another problem is the processing of a significant amount of heterogeneous information in near real time. To address this issue, we propose a multi-level Information Fusion framework based on Dynamic Bayesian Networks (DBN) with an active perception approach. The contribution of this model primarily lies in its capability of handling both Hard & Soft sensors and resulting information. And secondly in the identification of the most valuable DBN variables which maximize the information gains in the next step. These valuable variables allow to infer states on a sub-DBN to reduce the complexity of calculation. In this top-down approach, we seek which variables can provide the most valuable information to select automatically the right sensors and choose correct actions to optimally observe these variables. We finally propose an illustration with a basic maritime scenario.

Situation assessment is at the core of many critical tasks in the civilian and military domains: border monitoring, surveillance of areas and facilities, entity tracking and identification, all require accurate and up-to-day descriptions of the course of events. For all those applications, situations to be built are complex, dynamic and uncertain and their assessment is based on the integration of diverse sources, including sensors and their row values, images, observations, tactical information and knowledge expressed by domain experts or synthesized through discovery techniques. This paper presents a method to combine soft and sensor data to create enhanced situation assessment for a track-and-detect application. First we create a situation of entities and relationships by using only hard data provided by sensors and then we enrich this situation thanks to soft data, in the form of succinct or more complex observation reports. The system relies on semantic mediation to combine observations and sensor data by using ontologies as a common ground creating a bridge between two complementary yet incomplete representations of the world. The result is an augmented situation, having more precise, accurate or complete descriptions of entities and which is easier to analyze. This enhanced assessment allows for the situation to be understood and processed in a meaningful way by decision makers.

Threat assessment is a high-level information fusion task which supports a decision maker in achieving a proper level of situational awareness. In this paper, we propose an approach to threat assessment that fuses both game-theoretic behavior modeling and adversarial intention recognition. In adversarial situations, where two opposing forces are competing to achieve conflicting goals in a shared environment, the behavior of the observed adversary will not only depend on the goal it is trying to achieve, but also on the decisions of its opponent. We model the planning process of such a goal-directed adversary as an Attack Stochastic Game and show how we can infer its intention by inverting this model. We evaluate the potential of our approach for early and accurate intention recognition on an illustrative scenario where a decision maker has to defend a set of valuable assets against an attacker, and we discuss planned extensions.

In recent years homeland security is becoming increasingly sensitive to threats posed by the tactics of subversive groups or individuals with malicious intent. Networked groups and organizations leverage various means of communication, ranging from simple phone calls to more sophisticated forms of collaborations. Such data provide a rich collection of evidence from which to infer relationships of individuals or even the structure of networks and organizations. This paper describes an approach developed to combine information from disparate data sources in order to identify relevant relationships in heterogeneous environments. The objective is to discover entities and the relationships they share by a joint combination of sensor and soft data. By analyzing patterns of sensor-based communications and content of reports, the solution highlights entity associations which are indicative of social relationship. First, the approach identifies a set of entity relations thanks to sensor data. Then, association rules extract relations from texts by using a combination of part of speech features provided by natural language processing tools and the entity type, as labeled by a domain ontology. The overall approach is designed for domains such as intelligence analysis, where the analyst attempts to build a picture of relations holding between individuals by combining sensor streams that can refer to real-world ongoing events and more complex reports provided by human sources. Results are evaluated over a heterogeneous data set and experiments show that the combination of soft and hard data outperforms individual approaches in terms of recall.

Data fusion in heterogeneous environments plays a major role in assisting end users by providing them with an increased situational awareness so that decisions can be made about events in the field. Heterogeneous fusion involves combining different types of soft and hard data such that the situation or the resulting output is more precise, accurate, complete or easy to comprehend by decision makers. If soft data conveys more sophisticated information that is difficult to measure and hard data can be described with specificity, the question of how to take advantage of their complementarities is attracting considerable attention from data fusion community. This paper presents two methods for heterogeneous fusion, differing in procedures used to combine information items. We propose two methods that enrich a situation by adding supplementary attributes to entities, so that entities have a better characterisation. A domain ontology and reasoning capacities support both methods, although they implement different enrichment solutions. First, a picture of entities and relationships is created by using only hard data provided by sensors and then this picture is enriched thanks to soft data, in the form of succinct or more complex observation reports. The enrichment allows the situation to be understood and processed in a meaningful way by end users; however uncertainty arises as various items are matched. The paper also discusses underlying uncertainties induced by both methods along criteria of the current URREF framework.

Adversarial reasoning is of the first importance for defence and security applications since it allows to (1) better anticipate future threats, and (2) be proactive in deploying effective responses. In this paper, we address the two subtasks of adversarial reasoning, namely adversarial plan recognition and strategy formulation, from a generative, game-theoretic perspective. First, a set of possible future situations is computed using a contextual action model. This projected situation serves as a basis for building a set of Markov games mod-eling the planning strategies of both the defender and his adversary. Finally, a library of critical plans for the attacker and a library of best responses for the defender are generated automatically by computing a Nash equilibrium in each game. The adversarial plan recognition task therefore consists of inferring a probability distribution over the set of possible plans of the adversary, while the strategy formulation problem reduces to the selection of the most appropriate response. Initial results on a urban warfare scenario suggest that our framework can be useful to model complex strategic interactions inherent to plan recognition in adversarial situations.

In case of emergency and evacuation, it is often impossible to interpret manually the complex behaviour of a crowd, essentially due to the lack of staff and time needed to understand a situation. In the literature, a monitored system using data fusion methods makes it possible to perform automatic situation awareness. Using Swarm Intelligence domain, the authors propose an approach based on multi-agent system to simulate and detect primitive collective behaviours emerging from a crowd panic. It enables anticipating collective behaviours in real-time as well as their anomalies according to specific scenarios. Detection is the possibility to learn, recognize and anticipate different behaviours by a probabilistic model. The collective behaviour detection of a crowd panic in real-time is based on a learning method on an extended model of Hidden Markov Model. This paper presents experiments of simulation and detection using an implementation of a virtual environment.