SciLander: Mapping the Scientific News Landscape

TL;DR

SciLander introduces a novel unsupervised method to learn representations of scientific news sources by extracting heterogeneous indicators, improving misinformation detection and revealing source reliability and bias during the COVID-19 pandemic.

Contribution

The paper presents SciLander, a new approach for modeling scientific news sources using multiple indicators, addressing challenges in misinformation related to complex scientific topics.

Findings

Outperforms state-of-the-art baselines in news veracity classification

Learned representations encode source reliability and bias

Effective on a large COVID-19 news dataset

Abstract

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. SciLander extracts four heterogeneous indicators for the news 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 (i.e., the semantic shift of terms), 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.