Abstract
Twitter exhibits several characteristics, including a limited number of features and noisy text information. Extracting valuable information from Twitter has made hot topic detection a challenging task. In this paper, a novel four-stage framework is proposed to improve the performance of topic detection. Data preprocessing is the first stage. Deep learning is then exploited to enrich short text information via image understanding. Next, improved latent Dirichlet allocation is used to optimize the image effective word pairs, which improves the accuracy of the extracted topic words. Finally, both short text and images are integrated for topic detection, in which the corresponding topics are mined based on fuzzy matching of topic words. A large number of experiments show that the proposed framework significantly improves the performance of topic detection and outperforms the selected baseline methods.
Introduction
Social media platforms like Twitter serve as real-time sensors for global events. However, the "short text" nature of tweets—often limited to 280 characters—creates a significant data sparsity problem for traditional topic modeling. While many tweets are accompanied by images that provide rich context, most existing frameworks either ignore visual data or struggle to integrate it effectively without introducing noise. This research addresses the semantic gap between raw pixels and short text to create a more robust hot topic detection system.
The Proposed Framework
The authors introduce a four-stage pipeline designed to bridge visual and textual information and overcome the sparsity of social media data.
1. Text Preprocessing
The initial stage involves standard cleaning procedures, such as word stemming and the removal of special characters, URLs, and noisy user tags to reduce the dimensionality and noise of the input data.
2. Image Understanding via Deep Learning
To enrich the sparse tweet text, the framework employs a Neural Image Captioning (NIC) architecture. As shown in the figure below, a Convolutional Neural Network (CNN) acts as an encoder to extract visual features, which are then passed to a Recurrent Neural Network (RNN/LSTM) decoder to generate natural language descriptions.
3. Improved Latent Dirichlet Allocation (LDA)
Raw image captions can be verbose or contain irrelevant terms. To mitigate this, the authors developed an "improved LDA" model that identifies and optimizes image effective word pairs. By filtering out low-frequency noise and focusing on recurring semantic pairs, the system ensures that only the most relevant visual information is merged with the original tweet text.
4. Topic Detection & Fuzzy Matching
The final stage clusters the enriched data into topics. To handle the informal nature of social media (e.g., typos and slang), the system utilizes Levenshtein distance for fuzzy matching of topic words, increasing the model's resilience compared to strict keyword matching.
Experimental Evaluation
The framework was tested on a dataset of 34,554 tweets across 10 diverse topics, including "World Cup" and "Traffic Jam." The integration of 22,647 images generated over 240,000 additional descriptive words.
Key results include:
- Performance Gains: The framework achieved F1-scores as high as 90%.
- Recall Improvement: For specific topics like "Giraffe," recall jumped from 0.51 to 0.90 after integrating image data, demonstrating the model's ability to connect tweets that lack shared keywords but share visual context.
- Noise Mitigation: The improved LDA stage successfully filtered out irrelevant words that a naive text+image approach would have included, maintaining high precision.
Conclusion
By treating images as semantic anchors, this framework overcomes the inherent limitations of short-text analysis on social media. While challenges remain regarding computational costs for real-time processing and potential captioning errors, the study proves that visual data is a vital, underutilized resource for understanding the pulse of social media.
Summary updated