Search

Fake news is a threat for the society and can create a lot of confusion to people regarding what is true and what not. Fake news usually contain manipulated content, such as text or images that attract the interest of the readers with the aim to convince them on their truthfulness. In this article, we propose SceneFND (Scene Fake News Detection), a system that combines textual, contextual scene and visual representation to address the problem of multimodal fake news detection. The textual representation is based on word embeddings that are passed into a bidirectional long short-term memory network. Both the contextual scene and the visual representations are based on the images contained in the news post. The place, weather and season scenes are extracted from the image. Our statistical analysis on the scenes showed that there are statistically significant differences regarding their frequency in fake and real news. In addition, our experimental results on two real world datasets show that the integration of the contextual scenes is effective for fake news detection. In particular, SceneFND improved the performance of the textual baseline by 3.48% in PolitiFact and by 3.32% in GossipCop datasets. Finally, we show the suitability of the scene information for the task and present some examples to explain its effectiveness in capturing the relevance between images and text.

Predicting mortality for ICU patients while ensuring fairness across different demographic groups is a multifactorial issue. This study aims to address this challenge by leveraging the Medical Information Mart for Intensive Care (MIMIC-IV) dataset to develop robust machine learning models. The study compares neural network and logistic regression models using both a comprehensive set of predictors and a subset of the most significant predictors. Bias mitigation techniques, including reweighting and threshold modification, were applied to address disparities in model performance. Results indicate that while overall accuracy was high, significant biases were observed, particularly against Asian patients and Medicaid insurance holders. The logistic regression model trained on a balanced dataset and adjusted through threshold modification emerged as the optimal choice, achieving minimal inequalities across subgroups while maintaining high accuracy and F1 scores for mortality prediction. These findings underscore the need for continuous evaluation and advanced bias mitigation strategies to ensure equitable healthcare outcomes.

This research investigates the application of pre-trained transformer-based models, including BERT, XLM- RoBERTa, and DistilBERT, in the context of the EXIST 2023 shared task, which focuses on identifying and categorizing online sexism. The study emphasizes the crucial role of Natural Language Processing (NLP) in detecting harmful content, and it draws on previous competitions that have incorporated tasks to detect hate speech and abusive language. The methodology combines various advanced techniques from the text classification domain, including the use of additional datasets, data preprocessing, and model building. The research also explores data augmentation techniques and label encoding as preprocessing steps. The study’s findings indicate that the developed model performs optimally in English, and it suggests that the use of a voting system and the combination of outputs from multiple models contribute to the overall performance. The research concludes with a call for sustained initiatives to curb the prevalence of harmful content on digital platforms, and it outlines future work directions, including incorporating additional information about annotators, the assessment of annotator reliability, and exploring more sophisticated techniques for handling imbalances.

Text-based personality computing (TPC) has gained many research interests in NLP. In this paper, we describe 15 challenges that we consider deserving the attention of the NLP research community. These challenges are organized by the following topics: personality taxonomies, measurement quality, datasets, performance evaluation, modelling choices, as well as ethics and fairness. When addressing each challenge, not only do we combine perspectives from both NLP and social sciences, but also offer concrete suggestions. We hope to inspire more valid and reliable TPC research.

The rise of social media has offered a fast and easy way for the propagation of conspiracy theories and other types of disinformation. Despite the research attention that has received, fake news detection remains an open problem and users keep sharing articles that contain false statements but which they consider real. In this article, we focus on the role of users in the propagation of conspiracy theories that is a specific type of disinformation. First, we compare profile and psycho-linguistic patterns of online users that tend to propagate posts that support conspiracy theories and of those who propagate posts that refute them. To this end, we perform a comparative analysis over various profile, psychological and linguistic characteristics using social media texts of users that share posts about conspiracy theories. Then, we compare the effectiveness of those characteristics for predicting whether a user is a conspiracy propagator or not. In addition, we propose ConspiDetector, a model that is based on a convolutional neural network (CNN) and which combines word embeddings with psycho-linguistic characteristics extracted from the tweets of users to detect conspiracy propagators. The results show that ConspiDetector can improve the performance in detecting conspiracy propagators by 8.82% compared with the CNN baseline with regard to F1-metric.

Artificial intelligence (AI) is increasingly being used in healthcare, particularly for interpreting medical images. However, there are growing concerns regarding the presence of biases in these AI models, which raise important fairness considerations. This study investigates biases in artificial intelligence (AI) models for chest X-ray diagnosis and explores the role of Explainable AI (XAI) in understanding model decisions. Biases were observed in model performance across different patient groups and diseases. Various XAI techniques were employed to generate explanations for model decisions, and comparisons were made with explanations provided by doctors. We identified an optimized version of occlusion as the most accurate XAI technique in this case, which also provided a consistent accuracy of the explanations across all patient groups. Indeed, the explanations remained equally accurate regardless of variations in model performance for different subgroups, suggesting the absence of model bias amplification. Evaluating the correctness of XAI explanations posed challenges due to the limited availability of ground truth. In order to increase the power of our analysis, we explored alternative evaluation methods, like deletion or insertion curves, but reported them as unsuitable for chest X-ray images. We have therefore established some recommendations for using XAI on chest X-ray images. Given the reported absence of biases in the explanations, our aim is also to instill confidence in clinical stakeholders regarding XAI techniques.

This study investigates the performance of two deep learning models, RoBERTa and Bi-LSTM, in predicting the Big Five personality traits and capturing the correlations among personality traits from text data. The models’ performance was evaluated on two datasets, PAN 2015 and PANDORA, using RMSE and R² values. The study found that RoBERTa outperformed Bi-LSTM in predicting personality traits on both datasets. However, both models demonstrated varying performances across the two datasets, highlighting the influence of data diversity on model performance. The study also examined the correlations among predicted personality traits and found that the models could capture the sign of the correlations present in the original datasets. However, divergences in the direction of correlations were observed in instances of weak correlations. Besides, the correlations of original datasets and predictions may not align with the findings in psychological studies, which address the importance of annotations when researching the correlations of predicted personality traits.