The Surprising Performance of Simple Baselines for Misinformation Detection

Abstract

As social media becomes increasingly prominent in our day to day lives, it is increasingly important to detect informative content and prevent the spread of disinformation and unverified rumours. While many sophisticated and successful models have been proposed in the literature, they are often compared with older NLP baselines such as SVMs, CNNs, and LSTMs. In this paper, we examine the performance of a broad set of modern transformer-based language models and show that with basic fine-tuning, these models are competitive with and can even significantly outperform recently proposed state-of-the-art methods. We present our framework as a baseline for creating and evaluating new methods for misinformation detection. We further study a comprehensive set of benchmark datasets, and discuss potential data leakage and the need for careful design of the experiments and understanding of datasets to account for confounding variables. As an extreme case example, we show that classifying only based on the first three digits of tweet ids, which contain information on the date, gives state-ofthe-art performance on a commonly used benchmark dataset for fake news detection -Twitter16. We provide a simple tool to detect this problem and suggest steps to mitigate it in future datasets. CCS CONCEPTS• Computing methodologies → Artificial intelligence; Natural language processing;

Towards Fine-Grained Reasoning for Fake News Detection

“…Content-based methods mainly utilize the textual or visual content from the news article and related posts for news verification (Yang et al 2012;Afroz, Brennan, and Greenstadt 2012;Kwon et al 2013;Przybyla 2020;Ma et al 2016;Zellers et al 2019;Qi et al 2019;Gupta et al 2013;Jin et al 2016b;Kaliyar, Goswami, and Narang 2021). These methods enable the detection of fake news at an early stage (Wei et al 2021;Pelrine, Danovitch, and Rabbany 2021). However, their performance is limited as they ignore auxiliary knowledge for news verification.…”

Folk Models of Misinformation on Social Media

“…Aside from 'myth busters' around COVID-19 and election ballot processing, online users could seek little guidance from the institutions in discerning misinformation. Discerning misinformation is, according to current evidence [23,36,63], not something users are endowed with, and experts for now only provide strategies for inoculation against misinformation [32,38] in addition to the automated detection [31,50,79].…”

“…The distribution of participants per their self-reported gender identity was 102 (43.4%) were female, 117 (49.78%) male, and 16 (6.82%) preferred not to say. Age-wise, 32 (12.76%) were in the [18][19][20][21][22][23][24][25][26][27][28][29][30] bracket, 100 (42.55%) in the [31][32][33][34][35][36][37][38][39][40] bracket, 60 (25.53%) in the [41][42][43][44][45][46][47][48][49][50] bracket, 28 (11.91%) in the [51][52][53][54][55][56][57][58][59][60] bracket, and 15 (6.38%) were 61+ old. The distribution of the political leanings within the sample was: apolitical 10 (4.25%), left-leaning 115 (48.93%), moderate 61 (25.95%), and 49 right-leaning (20.85%).…”

Probing Spurious Correlations in Popular Event-Based Rumor Detection Benchmarks

“…For instance, [4,9] respectively employ Recurrent Neural Networks and Convolutional Neural Networks to model the variations of text and user representations over time. Hierarchical attention networks [10] and pre-trained language models [11] have also proven effective. Another line of work leverages propagation-based information diffusion patterns to encode information flow along user interaction edges.…”

“…Investigation of Spurious Correlations Despite the promising performance of deep learning models, reliance on dataset-related cues has been observed in a wide range of tasks including text classification [17], natural language inference [32] and visual question answering [30]. In fact-checking scenarios, language models can capture underlying identities of news sites [19], and rumor instances can possess time-sensitive characteristics [11]. Spurious artifacts lead to model failure on out-of-domain test instances, as empirically observed by [23,29,41].…”