Abstract
Due to various and serious adverse impacts of spreading fake news, it is often known that only people with malicious intent would propagate fake news. However, it is not necessarily true based on social science studies. Distinguishing the types of fake news spreaders based on their intent is critical because it will effectively guide how to intervene to mitigate the spread of fake news with different approaches. To this end, we propose an intent classification framework that can best identify the correct intent of fake news. We will leverage deep reinforcement learning (DRL) that can optimize the structural representation of each tweet by removing noisy words from the input sequence when appending an actor to the long short-term memory (LSTM) intent classifier. Policy gradient DRL model (e.g., REINFORCE) can lead the actor to a higher delayed reward. We also devise a new uncertainty-aware immediate reward using a subjective opinion that can explicitly deal with multidimensional uncertainty for effective decision-making. Via 600K training episodes from a fake news tweets dataset with an annotated intent class, we evaluate the performance of uncertainty-aware reward in DRL. Evaluation results demonstrate that our proposed framework efficiently reduces the number of selected words to maintain a high 95% multi-class accuracy.
Introduction
The proliferation of misinformation on social media has profound societal consequences. While much research focuses on the veracity of information, understanding the intent behind sharing is equally vital for developing effective intervention strategies. Not all spreaders of fake news are malicious; motivations range from accidental sharing and socialization to emotional venting and political campaigning.
Technically, intent analysis faces two primary challenges: the inherent difficulty of mining intent from short, noisy social media text, and the "delayed reward" problem in reinforcement learning (RL). In typical NLP tasks using RL, the agent only receives feedback at the end of a sequence, making the learning process inefficient. This paper introduces an Uncertainty-Aware Reward-based Deep Reinforcement Learning (DRL) framework to address these issues by denoising input text and providing immediate, uncertainty-based rewards.
Methodology
The proposed framework integrates a Long Short-Term Memory (LSTM) network with a DRL agent. The LSTM serves as the environment and base classifier, while the DRL agent (the Actor) processes tweets word-by-word to decide whether to keep or mask each token. This process aims to extract a "signal" from the "noise," retaining only the words most relevant to intent classification.
A key innovation is the use of Subjective Logic (SL) to quantify multidimensional uncertainty, specifically vacuity (lack of evidence) and dissonance (conflicting evidence). The system provides an immediate reward based on the reduction of these uncertainty metrics as the agent processes the text. This immediate feedback, combined with a delayed reward for final classification accuracy, guides the agent toward more effective decision-making.
Experimental Setup
The study utilizes the LIAR 2015 dataset, a benchmark for fake news detection. To facilitate intent analysis, the researchers manually annotated a subset of 1,500 tweets (835 fake, 665 true). Three human annotators categorized the tweets into five intent classes: socialization, political campaign, information sharing, rumor propagation, and emotion venting. The DRL model was trained over 600,000 episodes to optimize the word-masking policy.
Results and Discussion
The experimental results demonstrate that the DRL framework significantly outperforms standard LSTM models. By selectively masking irrelevant words, the DRL agent achieved a multi-class accuracy of 95-97%.
As shown in the results, the DRL-based models (DRL, DRL-CV, DRL-CD) consistently improved the "gold class" prediction rate and overall accuracy compared to the baseline LSTM. Specifically, the DRL approach improved overall intent classification by approximately 11%. Notably, the framework was slightly more effective at identifying the intent behind fake news (up to 98.3% accuracy) than true news (95.8%).
The "less is more" principle was validated: the model maintained high performance even while processing a reduced sequence of words, proving that denoising the input text enhances the clarity of the underlying intent. The use of uncertainty-aware rewards effectively solved the sparse reward problem, leading to faster and more stable convergence during training.
Conclusion
This research presents a novel DRL framework for intent analysis that leverages uncertainty-aware rewards to improve both accuracy and efficiency. By distinguishing between different types of fake news spreaders, this work provides a foundation for more nuanced social media interventions. Future work will focus on scaling the annotation process through semi-automated techniques and exploring the model's performance on diverse, evolving social media platforms.
Summary updated