[Paper Review] Call Attention to Rumors: Deep Attention Based Recurrent Neural Networks for Early Rumor Detection

Tong Chen, Lin Wu+4 more|arXiv (Cornell University)|Apr 20, 2017

Misinformation and Its ImpactsSocial Sciences33 references20 citations

TL;DR

This paper proposes CallAtRumors, a deep attention-based recurrent neural network that leverages soft-attention mechanisms within LSTMs to learn temporal representations from sequential social media posts for early rumor detection. The model outperforms state-of-the-art methods by effectively focusing on salient, distinguishing features despite high textual duplication, achieving F-measures of 0.8672 on Weibo and 0.9061 on a sample dataset, with detection occurring up to 20.47 hours earlier than human fact-checking on average.

ABSTRACT

The proliferation of social media in communication and information dissemination has made it an ideal platform for spreading rumors. Automatically debunking rumors at their stage of diffusion is known as extit{early rumor detection}, which refers to dealing with sequential posts regarding disputed factual claims with certain variations and highly textual duplication over time. Thus, identifying trending rumors demands an efficient yet flexible model that is able to capture long-range dependencies among postings and produce distinct representations for the accurate early detection. However, it is a challenging task to apply conventional classification algorithms to rumor detection in earliness since they rely on hand-crafted features which require intensive manual efforts in the case of large amount of posts. This paper presents a deep attention model on the basis of recurrent neural networks (RNN) to learn extit{selectively} temporal hidden representations of sequential posts for identifying rumors. The proposed model delves soft-attention into the recurrence to simultaneously pool out distinct features with particular focus and produce hidden representations that capture contextual variations of relevant posts over time. Extensive experiments on real datasets collected from social media websites demonstrate that (1) the deep attention based RNN model outperforms state-of-the-arts that rely on hand-crafted features; (2) the introduction of soft attention mechanism can effectively distill relevant parts to rumors from original posts in advance; (3) the proposed method detects rumors more quickly and accurately than competitors.

Motivation & Objective

• To address the challenge of early rumor detection in social media, where manual feature engineering is impractical due to high volume and duplication.
• To develop a flexible, end-to-end deep learning model that captures long-range temporal dependencies in sequential posts.
• To improve detection accuracy and earliness by selectively attending to the most relevant words and phrases in evolving rumor discussions.
• To reduce reliance on hand-crafted features while maintaining robust performance on large-scale, duplicated social media data.

Proposed method

• The model uses a bidirectional LSTM to encode sequential social media posts into contextual hidden representations.
• A soft-attention mechanism is applied over the LSTM outputs to dynamically assign attention weights to different time steps based on relevance to rumor detection.
• Attention weights are computed using a feed-forward network over the LSTM hidden states, allowing the model to focus on salient phrases and reduce noise from duplicated content.
• The attended representations are pooled and passed through a final classification layer to predict whether a post sequence is a rumor or not.
• The model is trained end-to-end using cross-entropy loss with backpropagation through time.
• Textual features are represented using TF-IDF vectors of posts, preserving semantic content while reducing dimensionality.

Experimental results

Research questions

• RQ1 Can a deep attention-based RNN model detect rumors earlier and more accurately than traditional methods relying on hand-crafted features?
• RQ2 How effective is the soft-attention mechanism in identifying and emphasizing key distinguishing phrases within duplicated social media posts?
• RQ3 To what extent does the model maintain performance under high textual duplication common in viral rumor discussions?
• RQ4 How does the model’s earliness compare to human fact-checking timelines and other SOTA methods?

Key findings

• On the Weibo dataset, CallAtRumors achieved a precision of 87.10%, recall of 86.34%, and F-measure of 0.8672, outperforming state-of-the-art methods.
• On a sample dataset with 1,193 rumors and 6,387 non-rumors, CallAtRumors achieved an F-measure of 0.9061, compared to 0.8801 for CERT.
• The model detected rumors with a mean time lag of 20.47 hours after emergence, significantly earlier than the average human fact-checking time of 54–72 hours.
• At 10–20% of training data, CallAtRumors outperformed ML-GRU by 5% in precision and 4% in recall on both Twitter and Weibo datasets, demonstrating superior early detection capability.
• Attention visualization confirmed that the model effectively highlights distinctive keywords (e.g., 'disqualified', 'Trump') while downplaying redundant or repetitive phrases.
• The attention mechanism improved model robustness to textual duplication, enabling better feature distillation even when 80% of posts in an event were duplicates.