ManiTweet: A New Benchmark for Identifying Manipulation of News on Social Media

TL;DR

This work introduces ManiTweet, a benchmark task and dataset for identifying manipulation of news on social media, addressing the gap where posts distill or distort information from linked articles. It defines three subtasks—tweet manipulation detection, manipulating span localization, and pristine span localization—and builds ManiTweet from $3,636$ tweets across $2,688$ articles using a two-stage annotation process. The study finds that state-of-the-art LLM prompting under zero-/few-shot settings underperforms on this challenging task, while a fine-tuned LongFormer encoder-decoder (LED-FT) model outperforms LLMs, and a combined LLM+LED-FT pipeline yields the best results. Large-scale analyses reveal higher manipulation rates for low-trust or political articles and show that manipulated sentences tend to carry main or consequential discourse, offering actionable insights for misinformation research and future model development, including the potential of opinion-focused features to improve detection.

Abstract

Considerable advancements have been made to tackle the misrepresentation of information derived from reference articles in the domains of fact-checking and faithful summarization. However, an unaddressed aspect remains - the identification of social media posts that manipulate information within associated news articles. This task presents a significant challenge, primarily due to the prevalence of personal opinions in such posts. We present a novel task, identifying manipulation of news on social media, which aims to detect manipulation in social media posts and identify manipulated or inserted information. To study this task, we have proposed a data collection schema and curated a dataset called ManiTweet, consisting of 3.6K pairs of tweets and corresponding articles. Our analysis demonstrates that this task is highly challenging, with large language models (LLMs) yielding unsatisfactory performance. Additionally, we have developed a simple yet effective basic model that outperforms LLMs significantly on the ManiTweet dataset. Finally, we have conducted an exploratory analysis of human-written tweets, unveiling intriguing connections between manipulation and the domain and factuality of news articles, as well as revealing that manipulated sentences are more likely to encapsulate the main story or consequences of a news outlet.

Figures (1)

• Figure 1: Two illustrative examples that highlight the challenge of identifying manipulation of news on social media. The first example expresses a personal opinion about watching a well-reviewed movie without distorting any facts from the associated article. Conversely, in the second example, the tweet falsely asserts that the movie is directed by John Smith instead of Jane Doe, thereby misrepresenting the information contained in the reference article. Hence, the second tweet misrepresents the information contained in the reference article.

Theorems & Definitions (1)

• Definition 1