260 likes | 277 Views
Attention Based Neural Architecture for Rumor Detection with Author Context Awareness Sansiri Tarnpradab, Kien A. Hua Department of Computer Science, University of Central Florida. Today’s Agenda. Introduction Methodology Dataset Experiments Results Conclusion. Introduction.
E N D
Attention Based Neural Architecture for Rumor Detection with Author Context AwarenessSansiri Tarnpradab, Kien A. HuaDepartment of Computer Science, University of Central Florida
Today’s Agenda Introduction Methodology Dataset Experiments Results Conclusion
Introduction The motivation
Information sharing on social media is prevalent • The widespread of misinformation • False Rumors. Fake news. • Social stability • Economy • Politics
Determine whether it’s fake or real.. • Time. Every second counts! • Human fact-checking process: Evidence gathered verify conclude • False rumors spread like wildfire! • A large amount of money spent toresolve the problem
Automatic Rumor Detector • Save us time. • Save us cost. Invest in what really matters • Save us from the hassle • Immediately notifying the public that rumor is false • False rumors debunked sooner, less negative consequences triggered
Factors to determine the veracity: • Content What the event is about? • Context Other helpful information? Our research • Twitter • Content = tweets discussing the event • Context = user historical tweets • Present style and characteristics
Methodology Problem Definition Proposed Approach
Event Interval Word Problem Definition
Dataset Source of our dataset Statistics
Collected from two well-known rumor tracking websites • snopes.com • emergent.info
Experiments Model Configurations Baselines
Model Configurations • 80:10:10 Training : Test : Dev • Exploring the best hyperparameters • {Dropout rates, Hidden layer size, Feature maps, Optimizer} • 10% of total events were held out for model tuning User Embedding • 3,200 historical tweets per user were retrieved • usr2vec [1] was employed to initialize each user’s vector
Baselines DTC : Decision Trees Classifier by Castillo et al. [2] RFC-ext : Random Forest Classifier by Kwon et al. [3] SVM-TS : Support Vector Machine based on time-series structures by Ma et al. [4] RNN : Recurrent Neural Network GRU-2 : Depp learning based model with 2 GRU hidden layers by Ma et al. [5] CNN : Convolutional Neural Network
Results Let’s see what we found
Concatenate user embeddings to content vector of RNN, GRU-2, and CNN • Results with context included reveals performance improvement • RNN is improved by ~14% with the context from users
Conclusion The takeaways What’s next?
The attention applied at the word-level facilitates the model Semantic features from content + user embeddings allows the deep neural networks to better discriminate between rumor and non-rumor events Future work Tweet-level rumor detection Non-balanced dataset reflects the real-world scenario
References [1] Amir, Silvio, et al. "Modelling context with user embeddings for sarcasm detection in social media." arXiv preprint arXiv:1607.00976 (2016). [2] Castillo, Carlos, Marcelo Mendoza, and Barbara Poblete. "Information credibility on twitter." Proceedings of the 20th international conference on World wide web. ACM, 2011. [3] Kwon, Sejeong, et al. "Prominent features of rumor propagation in online social media." 2013 IEEE 13th International Conference on Data Mining. IEEE, 2013. [4] Ma, Jing, et al. "Detect rumors using time series of social context information on microblogging websites." Proceedings of the 24th ACM International on Conference on Information and Knowledge Management. ACM, 2015. [5] Ma, Jing, et al. "Detecting Rumors from Microblogs with Recurrent Neural Networks." IJCAI. 2016.