Source-linked AI summary
Call Attention to Rumors: Deep Attention Based Recurrent Neural Networks for Early Rumor Detection
Tong Chen, Lin Wu, Xue Li, Jun Zhang, Hongzhi Yin, Yang Wang
TL;DR
Early rumor detection must handle rapidly evolving social-media posts, duplicated text, and the limitations of hand-crafted features. CallAtRumors addresses this with an attention-based recurrent model that selectively learns temporal representations. It outperforms state-of-the-art competitors and detects rumors earlier on Twitter and Weibo.
Problem
Early rumor detection requires representations that capture long-range dependencies and changing contextual relevance in highly duplicated post sequences without relying on labor-intensive hand-crafted features.
Method
CallAtRumors uses an RNN with deterministic soft attention to weight feature positions over time and learn latent representations from sequential tf-idf post matrices.
Results
CallAtRumors outperforms five state-of-the-art methods and achieves strong precision, recall, and F-measure on both Twitter and Weibo datasets.
Takeaways & Limitations
The model detects rumors more quickly and accurately, including when early posts are sparse and highly duplicated.
Takeaways & Limitations
Future work identifies efficiency improvements using hashing techniques over multi-level feature spaces.
Abstract
from arXiv · showhide
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 \textit{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 \textit{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.
1 INTRODUCTION
Early rumor detection matters because social-media rumors can spread rapidly and cause severe harm. The paper proposes CallAtRumors, a deep attention-based RNN that learns temporal representations while focusing on distinctive features in duplicated, changing post streams.
- 1 INTRODUCTION: Rapidly spreading social-media rumors can trigger severe social panic and major financial losses.A false White House bombing report was followed by a $136.5 billion stock-market loss.
- 1 INTRODUCTION: Early detection is difficult because conventional methods depend on time-consuming, biased, labor-intensive, and data-dependent hand-crafted features.These features may fail to capture contextual variations across posts.
- 1 INTRODUCTION: Event-related posts arrive as time series but contain substantial textual duplication from repeated forwarding, reviewing, and inquiry.The resulting streams require representations that account for both temporal context and repeated phrases.
- 1.1 Challenges and Our Approach: CallAtRumors uses an attention-based RNN to learn hidden representations from streaming post sequences and predict whether an event is a rumor.The model processes feature matrices, captures long-range dependencies, and adapts its focus to contextual variations.
- 1.2 Contributions: The paper’s contributions include automatic early rumor detection, long-range dependency modeling, and deterministic soft attention for extracting distinct features from duplicated posts.Experiments compare the method with state-of-the-art approaches on Twitter and Weibo.
2 RELATED WORK
Prior rumor-detection work uses classification with manually engineered content, sentiment, user, propagation, or temporal features. Attention mechanisms provide a related route for automatically emphasizing source information relevant to prediction.
- 2 RELATED WORK: Rumor detection is commonly formulated as a binary classification problem whose performance depends strongly on discriminative feature extraction.Existing approaches use features such as sentiment, propagation structures, and temporal context.
- 2 RELATED WORK: Many prior methods require substantial manual feature engineering and remain restricted by their data structures.These limitations motivate models that learn representations more automatically from sequential posts.
- 2 RELATED WORK: Early rumor detection targets viral rumors during their formative stages so that action can be taken sooner.Earlier work includes enquiry phrases, manually defined features, and sparse learning for selecting discriminative features.
- 2 RELATED WORK: Attention mechanisms allow models to automatically search for source-sentence parts relevant to predicting a target word.This NLP approach motivates selective feature weighting in the proposed rumor-detection framework.
3 RECURRENT NEURAL NETWORKS
RNNs process sequential inputs one element at a time while maintaining hidden states that summarize prior elements. Because standard RNNs struggle with long-distance dependencies during training, the paper uses LSTM memory cells for streaming post series.
- 3 RECURRENT NEURAL NETWORKS: RNNs process sequential data element by element, updating hidden states that implicitly contain information about earlier inputs.They begin from an initial state and generate an output vector at each step.
- 3 RECURRENT NEURAL NETWORKS: The basic RNN formulation uses input-to-hidden, hidden-to-output, and hidden-to-hidden weight matrices together with bias vectors and a tanh nonlinearity.These parameters govern recurrent state updates and output generation.
- 3 RECURRENT NEURAL NETWORKS: Standard RNNs are difficult to train because vanishing or exploding gradients limit their ability to capture long-distance temporal dependencies.Back-propagation through time is used for gradient computation.
- 3 RECURRENT NEURAL NETWORKS: LSTM memory cells address this training difficulty by storing information over time, making them suitable for variable-length streaming post series.The proposed model employs LSTM units to capture long-term temporal dependencies.
4 CALLATRUMORS: EARLY RUMOR DETECTION WITH DEEP ATTENTION BASED RNN
CallAtRumors converts event-level streams of social posts into variable-length time series, then uses deep LSTMs with deterministic soft attention to classify rumors and non-rumors. Attention dynamically emphasizes relevant textual features while recurrent states model temporal dependencies.
- Problem Statement: The model detects rumors at the event level by batching sequential posts related to the same topic rather than classifying individual posts.Each event consists of relevant posts with timestamps and receives a rumor or non-rumor label.
- Constructing Variable-Length Post Series: Posts are grouped into fixed-size batches, producing variable-length series with at least Min intervals for sequential processing.Events with insufficient posts are distributed across intervals, while the final interval may contain the remainder.
- Constructing Variable-Length Post Series: Each post is encoded with K-dimensional tf-idf features, and each interval forms a K×N matrix, padded with zeros when necessary.The representation uses the most frequent K vocabularies and fixed post amount N per interval.
- Long Short-Term Memory: Stacked LSTM layers process the time-series feature slices, using gated memory to capture long-distance temporal dependencies among posts.LSTM gates regulate information entering, retained in, and emitted from the memory cell.
- Deterministic Soft Attention Mechanism: At each time step, the model predicts attention weights over tf-idf positions and a binary rumor classification from recurrent hidden states.The location softmax determines attention for the next input matrix, while an additional hidden layer produces the rumor/non-rumor prediction.
- Deterministic Soft Attention Mechanism: Deterministic soft attention forms the next context vector as an expectation over feature vectors, enabling end-to-end back-propagation.The attention distribution assigns higher focus to words considered distinctive for rumor or non-rumor representations.
- Loss Function and Model Training: Training combines cross-entropy classification loss with doubly stochastic regularization and weight decay.The attention regularizer encourages coverage of input-matrix elements during training.
5 EXPERIMENTS
Experiments use event-level Twitter and Sina Weibo datasets, evaluate CallAtRumors against established baselines, and test both effectiveness and early detection. CallAtRumors performs strongly on duplicate-content data and identifies rumors before typical reporting times.
- Datasets and evaluation: The experiments evaluate CallAtRumors on event-level Twitter and Sina Weibo datasets, comparing precision, recall, and F-measure with state-of-the-art approaches.Events aggregate posts about the same topic and are labeled as rumor or non-rumor.
- Datasets and evaluation: More than 80% of users repost original news with short comments, creating substantial textual duplication and scarcity of distinctive patterns within events.The datasets are therefore challenging for methods that depend on manually engineered textual features.
- Settings and baselines: A three-layer LSTM with 1,024, 512, and 64 hidden states respectively gives the best result among the tested layer configurations.The number of layers was varied from two to six using a 15% hold-out dataset.
- Effectiveness validation: CallAtRumors achieves 88.63% precision, 85.71% recall, and 0.8694 F-measure on the Twitter dataset.These results are reported as outperforming competing approaches.
- Effectiveness validation: CallAtRumors achieves 87.10% precision, 86.34% recall, and 0.8672 F-measure on the Weibo dataset.The same reported advantage is observed on the second social-media dataset.
- Earliness analysis: With only early posts, CallAtRumors starts at 74.02% and 71.73% precision and 68.75% and 70.34% recall on Twitter and Weibo, respectively.The reported average time lag is 20.47 hours, compared with average reporting times of 54 and 72 hours.
6 CONCLUSION
The paper concludes that CallAtRumors uses soft attention within an RNN to learn latent representations from sequential social posts for early rumor detection. Experiments report sensitivity to distinguishable words and competitive performance under sparse and duplicate textual features, while future work includes richer propagation modeling and efficiency improvements.
- Conclusion: CallAtRumors embeds soft attention in an RNN to learn latent representations from sequential social posts for early rumor detection.The model is intended for time-sensitive detection as rumors spread quickly and broadly.
- Conclusion: Experiments with five state-of-the-art methods report that CallAtRumors remains sensitive to distinguishable words when textual features are sparse at a rumor’s beginning.The paper also reports handling duplicate data through further comparison.
- Future work: Future work proposes modeling rumor propagation patterns as sequential RNN inputs to improve detection accuracy and investigating hashing over multi-level feature spaces for efficiency.These are presented as directions rather than evaluated components of the current method.