Preface

We are living in unprecedented times. Humanity is working together to fight the pandemic. The COVID-19 pandemic has altered our way of life and how we go about our daily routines. For many, kitchen and dining room tables have become the new office, classroom, or social meeting place. Virtual is the new reality, and while it allows many of us to continue to work, learn and socialize safely, it comes with its own unique set of challenges and dangers. As we transition to using more online platforms, scammers, hackers and other cybercriminals are finding new ways to target people. With social distancing as the new norm, socialization has turned even more digital than ever before. Unfortunately, cyber bullying is rampant and has been linked to a variety of negative effects, including anxiety, depression, and substance abuse. In addition, according to Freedom Housefoot_0 , disinformation campaigns have been on the rise since March 2020. Cooordinated and covert attempts by China-linked actors to manipulate information, particularly regarding COVID-19, have been detected in many countries, including the United States, Argentina, Serbia, Italy, and Taiwan, with the relevant content often delivered in local languages. Our workshop tries to address these challenges.

Vijjali et al., for example, develop a two stage automated pipiline for COVID-19 fake news detection: the first model retrieves the most relevant facts concerning user claims about specific COVID-19-related claims; the second model verifies the level of truth in the claim by computing the textual entailment between the claim and the true facts retrieved from a manually curated COVID-19 dataset. The authors demonstrate the effectiveness of pre-trained Transformer-based language models in retrieving and classifying fake news and suggest that the model performance can be further boosted by adding domain-specific knowledge. Shin et al. describe the communication patterns on the coronvirus pandemic from three Asian countries using a time-topic cohesive model, which can detect contextualized events based on topical and temporal information via contrastive learning. It then can be applied to multiple languages, enabling a comparison of risk communication across cultures. They present a case study and discuss future implications of the proposed model.

Information warfare has recently shifted its focus to social media, but traditional media continue to play an important role. In their work with civial society organizations and journalists around the world, Niven & Kau have identified the usefulness of NLP tools to detect and to track media that are aligned with the Chinese Communist Party (CPP). Niven & Kau operationalize alignment in terms of sociological definitions of media bias. They take as a case study the alignment of four Taiwanese media outlets to the CPP state media. They present the results of an initial investigation using the frequency of words in psychologically meaningful categories. Their findings suggest that the chosen word categories correlate with framing choices.

Prakash & Madabushi propose a novel architecture for the task of stance detection, which integrates count-based features into pre-trained models such as B E RT and RoBERTa, with specific regard to the identification of disinformation and propaganda on social media, where handcrafted and count-based features can be particularly helpful.

Lily et al. focus on the task of satire detection, noticing that satire is often misinterpreted by readers as legitimate news. They propose a multi-modal approach based on the state-of-the-art visiolinguistic model Vi L B E RT. They create a new dataset consisting of images and headlines of regular and satirical news, fine-tune Vi L B E RT on the dataset and train a convolutional neural network that uses an image forensics technique. Evaluation on the dataset shows that their proposed multi-modal approach outperforms imageonly, text-only, and simple fusion baselines. Abstract: Billions of people on the internet are exposed to the outputs of downstream natural language processing (NLP) applications on a daily basis. Many of these NLP applications use word embeddings as general purpose language representations. In this talk, Aylin Caliskan will introduce the Word Embedding Association Test (WEAT) to demonstrate that word embeddings trained on language corpora embed the biases and associations documented by the Implicit Association Test in social psychology. In particular, WEAT measures how statistical regularities of language capture biases and stereotypes, such as racism, sexism, and attitudes toward social groups. Word embeddings are trained on text col-lected from the internet, that includes society's organic natural language data in addition to text from information influence operations. The adaptation of WEAT to the information influence domain automatically characterizes overall attitudes and biases associated with emerging information influence operations. Accurate analysis of these emerging topics usually requires laborious, manual analysis by experts to annotate a large set of data points to identify biases in new topics. We validate our practical and non-parametric method using known information operation-related tweets from Twitter's Transparency Report. We perform a case study on the COVID-19 pandemic to evaluate our method's performance on non-labeled Twitter data, demonstrating its usability in emerging domains.

Bio: Aylin Caliskan is an Assistant Professor of Computer Science at George Washington University. Her research interests lie in A I ethics, bias in AI, machine learning, and the implications of machine intelligence on fairness and privacy. She investigates the reasoning behind biased A I representations and decisions by developing explainability methods that uncover and quantify biases of machines. Building these transparency enhancing algorithms involves the heavy use of machine learning, natural language processing, and computer vision in novel ways to interpret A I and gain insights about bias in machines as well as society. In her recent publication in Science, she demonstrated how semantics derived from language corpora contain human-like biases. Prior to that, she developed novel privacy attacks to deanonymize programmers using code stylometry. Her presentations on both de-anonymization and bias in machine learning are the recipients of best talk awards. Her work on semi-automated anonymization of writing style furthermore received the Privacy Enhancing Technologies Symposium Best Paper Award. Aylin holds a PhD in Computer Science from Drexel University and a Master of Science in Robotics from the University of Pennsylvania. Before joining the faculty at George Washington University, she was a Postdoctoral Researcher and a Fellow at Princeton University's Center for Information Technology Policy.

Invited Talk: Diversionary Agenda Setting and Micro-Targeted Persuasion

Stephan Lewandowsky University of Bristol Abstract: We are said to live in a "post-truth" era in which "fake news" has replaced real information, denial has compromised science, and the ontology of knowledge and truth has taken on a relativist element. I argue that to defend evidence-based reasoning and knowledge against those attacks, we must understand the strategies by which the post-truth world is driven forward. I depart from the premise that the post-truth era did not arise spontaneously but is the result of a highly effective political movement that deploys a large number of rhetorical strategies. I focus on two strategies: Diversionary agenda-setting by social media and the use of "micro-targeting" of political messages online. I present evidence for the existence of each strategy and its impact, and how it might be countered.

Bio