TL;DR
This paper introduces a novel public emotion feature extraction algorithm leveraging Transformer technology to analyze public opinion on social media during natural disasters. By accurately identifying emotional states, the system aims to provide timely insights for disaster response and governance.
What Problem Does This Paper Address?
Natural disasters are inherently unpredictable and can have devastating consequences. Effectively managing public opinion during and after such events is crucial for disaster response, resource allocation, and post-disaster reconstruction. Social media has become a primary channel for public discourse, but the sheer volume and complexity of information, often mixed with emotional content, make it challenging to extract actionable insights. Existing methods struggle to accurately and quickly identify public sentiment vectors in large datasets, potentially leading to a delayed or misinformed response. This paper addresses the need for an intelligent system that can efficiently analyze public emotions expressed on social media during natural disasters to inform effective governance and aid.
Methods
The proposed approach integrates social media data with a public opinion analysis system. The core of the method is a Public Emotion Feature Extraction algorithm built upon the Transformer network.
Here's a breakdown of the key steps:
- Data Collection: Social media data (text, audio, video) related to natural disasters is collected using crawlers.
- Text Preprocessing: Standard NLP techniques like word segmentation, stop-word removal, and part-of-speech tagging are applied.
- Transformer-based Feature Extraction: The Transformer network is employed to process the collected information. This involves:
- Self-Attention Mechanism: Computing query (Q), key (K), and value (V) matrices from the input features (F) using learned weight matrices (WQ, WK, WV). This step aims to increase feature association and universality.
- Multi-Head Self-Attention: Combining outputs from multiple self-attention layers to capture diverse aspects of the emotional features.
- Feedforward Network: Further processing the attention outputs through a feedforward neural network, followed by summation and layer normalization.
- Intelligent Recognition of Public Emotional Information: Multiple convolution kernels are used to extract emotional features from the processed text. This involves:
- Input Layer: Convolution, pooling, and fully connected layers to represent emotional features as vectors.
- Convolutional Operation: Applying kernels of varying sizes (2, 3, 4, 5, 6, 7) to extract features from different windows of the emotion feature.
- Downsampling and Classification: Using maximum downsampling to obtain a final feature vector, which is then fed into fully connected and softmax layers for classification. Dropout is used to prevent overfitting.
The system aims to classify public sentiment into categories like positive, negative, and neutral, and then extract key features such as sentiment words, intensity, and polarity.
Datasets
The study utilizes a Twitter dataset (https://figshare.com/articles/dataset/tweets_csv_gz/3465974/2) for training and evaluation. The authors extracted news and public opinion comments related to natural disasters from this dataset.
Key Insights and Ideas
- Transformer's Efficacy in Emotion Analysis: The paper highlights the effectiveness of the Transformer architecture, particularly its self-attention mechanism, in capturing nuanced emotional features from social media text, outperforming traditional methods.
- Multi-faceted Emotion Extraction: By combining Transformer-based feature extraction with convolutional neural networks (CNNs), the system can capture both global semantic context and local emotional cues, leading to more robust sentiment analysis.
- Real-time Public Opinion Monitoring: The proposed system offers a pathway to near real-time monitoring of public sentiment during natural disasters, enabling faster and more informed decision-making by authorities.
- Bridging the Gap between Social Media and Governance: The research demonstrates a practical application of social media analytics for public governance, translating raw public opinion into actionable insights for disaster management.
Key Results
The proposed method achieved an accuracy of 90.54% in identifying the category of public opinion on natural disasters. This performance surpasses several other advanced technologies, including CNN, LSTM, CNN-LSTM, CNN-FPN, Vit, and Blip. Specifically, the accuracy was improved by:
- 9.22% and 8.31% compared to CNN and CNN-FPN, respectively.
- 11.98% and 7.32% compared to LSTM-based models.
- 4.85% and 2.89% compared to Vit and Blip (Transformer-based methods).
While the model has a larger number of parameters and a longer inference time compared to some simpler models, the authors prioritize accuracy and argue that their approach appropriately balances model structure to prevent feature loss.
Limitations and Future Work
The authors acknowledge that their model has a larger number of parameters and a longer inference time compared to some other methods. They also note that CNN-based models can quickly lose details of emotional features, and LSTM-based models may miss some global features, leading to classification inaccuracies.
Future work could focus on optimizing the model for reduced computational cost and inference time without sacrificing accuracy. Further exploration into more granular emotion detection and the integration of multimodal data (beyond text) could also enhance the system's capabilities. The paper suggests that the system can accurately identify public opinion status and deconstruct the impact of natural disasters on the public, implying potential for more in-depth analysis and prediction.
Summary updated